So the 1970s oil crisis caused a shift that included fuel efficiency rules.
For like, the whole time I've been a software engineer, memory footprints of a lot of software have crept up, even when they aren't really doing more.
The memory situation currently is a mess. But ... is it bad enough to get us to _change_ how we build?
That would be a great outcome of this situation. I doubt it will happen though: unlike fuel efficiency rules, there aren’t likely to be external pressures like that on software. Outside developer circles most users likely are unaware of memory usage and often just accept that computers are slow and unreliable sometimes. That’s different from drivers who see their fuel gauge and the corresponding cost of fuel, and are thus very aware of what fuel efficiency means.
The situation may be even more different from fuel. The memory shortage may be just a first of such crisis to come - the shortage of anything related to AI/robotics (from electricity to materials needed to produce robots and for the robots to produce even more robots). I think we have a runaway situation - the more of AI/robots produced, the more is demand for such production. The difference from classic production/manufacturing here is that AI/robots is a consumer itself. In the past the only consumers were people and human population growth is slow whereis AI/robots "population" is starting to explode. (Like Industrial Revolution on steroids - back then getting a steam engine to power your machinery was not just a jump in productivity, it was a jump without which you couldn't compete, and so everybody was getting it https://c8.alamy.com/comp/2RFT119/illustration-of-factories-... (and just for fun compare to https://www.reuters.com/resizer/v2/6BC2YPEXOZCGJKEO4LWH7CTLI... ) - fortunately for England back then getting a steam engine to produce more steam engines to thus make more and longer jumps in productivity didn't really work that fast and efficient whereis today the situation is becoming more favorable for such productivity multiplication)
Wrt. the memory efficiency - considering that AI is going to be the main workload by far and thus will vacuum in majority of the available optimization resources - be it humans or coding AI - i think it will be even less chances that any meaningful resources will be spent on optimizing on the fringe, ie. the users' OS/browser/etc.
This makes me think of Factorio. When you make a breakthrough like getting robots, what do you do with it? Use it to make more robot factories.
But in the real world, haven't we got to see whether all this investment will pay off? What happens if it turns out people actually don't want to pay for this?
The companies go bankrupt their stuff is sold for pennies & investors loose their money.
The now super cheap stuff might get reused for something else, that is actually profitable. Or will just get dumped, like many unprofitable or never finished lines from the railway mania days.
People don't want to pay for it. It's fine when you can buy a monthly plan and get lots of usage at the company's loss, but that's not sustainable.
Once everyone has to pay enough for the usage to just break even, it's going to be a shitshow. And what's worse, they'll have to make up for all the debt they're accumulating now, so even pricier.
Basically, individuals will only stay at the current massive discount. Corporations that are paying token costs right now are dialing back as they realize it's not worth the expense.
We are going to have an AI crash, and the only thing left are going to be the efficient on-device or local models, which are already starting to bridge the gap.
It's even more hilarious because all along there has been this complaint, and then people have been like "no no who cares, features are important, dev time is important, just ship ship ship". And now memory is unaffordable and it takes 1GB RAM to display a plaintext file
The cost of producing memory IS low, and is continuing to fall.
It makes zero sense to rebase your mental model of the tradeoff between RAM versus developer time because of a temporary shortage.
In a few years time you’ll look silly for having wasted developer time on lowering BF memory consumption, while your built new features and waited for supply and demand to rebalance.
Sure, and in the meantime, we become less efficient at doing legitimately useful tasks and more efficient at making rent seeking slop that collects money for the upper classes. The real fix is get rid of the 90% slop tech and put that to better use (including more efficient use of hardware)
Concidently, I asked Claude yesterday whether we have crossed the point that there is more compute and memory in the cloud or in personal computers.
It claims fp8 and fp16 has had higher throughput in the cloud since 2023.
fp32 and general computing still has more compute than personal computers.
And that memory is at the crossover point, but at this point only 10% of memory is going to personal computers, so the next few years will radically change the pc to cloud memory ratios.
Anyway, all of this made me a (tiny bit) sad. It is the gradual end of general purpose computing you own, and the start of everything just being cloud compute that someone else owns and you rent.
> It is the gradual end of general purpose computing you own, and the start of everything just being cloud compute that someone else owns and you rent.
A few thoughts...
(a) There've never been more general purpose computers than there are right now. Spec progress may stall for consumer products due to the price increases, but the demand for workstations is not going to go down that much; they're critical items to get work done. Many are hoping this means a return to memory efficiency; the 8GB Macbook Neo is a good sign that this is something Apple believes to some extent also.
(b) We've arguably been doing this since the 90s. A website (especially in the days of the heavy backend and more passive client) is cloud compute that someone else owns. You rent it by paying them in the form of looking at ads...
(c) There are a lot of companies and people working on making sure you have every chance to run your AI locally. It's surprisingly good now and keeps getting better. At some point we'll see the same oscillation that we always do from thin client to fat client as the latency advantage of working at the edge starts to win out. How long until I have a laptop with a Talaas chip in it running a basic local model at 16000 tok/sec that is just good enough to use A2A to reach a more powerful agent running elsewhere if I need it?
I agree with the sentiment, but there is a devils' advocate argument:
Most compute loads are bursty. Most PCs and phones sit idle almost all the time. It's an inefficient use of capital. In the cloud you can be idle and then suddenly command 1000X more power than any affordable PC could ever have... for 10 seconds... then be idle again. Those kinds of work loads aren't possible to realize on a PC without massively overbuying hardware.
I'd like to see more research on secure private cloud computing. Homomorphic encryption has gotten better but is still very slow, but there's other approaches that can help quite a bit but don't require much of a performance hit. Look into secure multiparty computation, reversible anonymization, etc.
I don't think personal computing will die, but I think it'll go back to being a hobbyist and pro thing like it was from its birth in the 1970s up until the middle 1990s. In the 1980s the average person did not own a computer, or if they did it was just to play some games. The average person was not loading BASIC programs off bootleg tapes and calling BBSes.
BTW I've been thinking a lot lately about how AI could save personal computing by allowing us to create an off-ramp from the present rats nest of complexity to something saner and simpler. Just thoughts right now but it's interesting.
Sure, scale of cloud systems has some interesting properties, that can be sometimes beneficial.
At the same time you are at the mercy of whoever runs the cloud, they can spy on your data or even manipulate your workloads, they can renegotiate the terms of use at any time - including price hikes or booting your from their cloud at any time.
Generative AI is definitely creating political will for simpler, more purposely-designed computer systems, but it's certainly not going to directly create help them in any meaningful way. AI-generated codebases are the posterchild of rat's nests of complexity.
I'd say the code I see coming out of leading models today is okay. Not great, but not terrible. It's also easy to get LLMs to code in more efficient languages, and you can also goad them via prompts and CLAUDE.md type stuff into prioritizing performance and simplicity.
It's sort of like what I say about self driving cars. Are they better than good human drivers? Not yet. Are they better than bad, drunk, or texting drivers? Yes.
LLMs are already better than bad coders and honestly most coders are bad coders.
In a world where we no longer have the rat's nest of complexity, we pretty much no longer need to produce code. All the software we will ever need for the next century could fit on books in one, single-floor library, if we exclude art projects like video games. An operating system with full driver support for every I/O peripheral (if those devices use standard protocols, allowing for a https://xkcd.com/927/ constant of 5), and support for every standard document format (again, assuming 5 bitmap image formats, 5 vector image formats, 5 kinds of print-ready document, 5 kinds of hypertext document…), could easily fit in 2GiB, and that would be done: everything anyone could possibly need out of a computer, apart from art, mathematics and science, for… I don't even know how long. We're still using the Latin alphabet, more or less unchanged since the late 15th century. Working technology doesn't need to change.
Being able to automatically generate large quantities of code that is worse than a good human programmer can produce is not helpful in this world you're envisioning.
It's interesting to note that the total compute forecast to be deployed in 2026-2027 is more than the amount that has been deployed during human history.
It's the price you pay for having an application delivery platform that also needs to be hugely sandboxed running one of the most difficult languages to run efficiently in such large amounts as so the plaintext minified versions are a dozen megabytes in size.
I don't envy browsers even slightly, peoples expectations of them are ludicrous.
The major problem is that nobody is genuinely looking at using more efficient languages, lessening the amount of required inefficient language use or gasp producing native applications. (which, is definitely an order of magnitude harder to produce, I'll grant).
If most software gets written by AI in the future and AI becomes self conscious. AI might write more memory efficient software, so there's more RAM available for the AI to grow. Basically the black mirror plaything episode.
What consumer software do you run that causes you to run out of memory or feel constrained? Even my phone has 16gb of ram. I need to run slack 16times to run out? Lol
When even the large AI labs that supposedly sit on (/with) the best of the best can't even ship native cross-platform desktop applications and instead ship some bloated Electron monsters, I'm not sure what kind of hope there is. The ones who were publishing lightweight apps continue to do so, the ones who didn't care, still don't care.
writing native gui code is atrocious, add to that llms being horrible with ui dev and you end up with with what you describe. but at least if we get more non-ui stuff written in more efficient languages there is going to be some benefit.
Compared to what though? It's not nearly as horrible as it used to be, and you can do it fairly neatly today, and end up with something much better. Requires you to actually have skill and knowledge about there being other alternatives to Electron though, which seems like knowledge hard to come by these days.
The situation is absolutely awful for anybody who needs significant memory or storage (or devices incorporating such).
Even people who think they have enough for the next few years already could end up with an unpleasant surprise if they e.g. have a stick or two of RAM go bad or a GPU fry. I have a few machines with aging sticks that I'm hoping will manage to hold out until when/if things improve.
I'm saving a ton of money by simply not upgrading. I used to do upgrades every 1-2 years just to keep up on things, pass the old box down to my son or otherwise donate it. Now I intend to keep my current setup for 7+ years. Its 2 years old next month. I may even stay with it longer. I just see no reason to upgrade with things how they are.
I considered a Mac Studio a couple months ago. Price was up but might have pulled the trigger if not for the delivery date that was almost 4 months out. Sale lost. I suspect a lot of people are like me right now and we'll start seeing significant drops in consumer spending on electronics the next 2-4 years minimum.
There also seems to have been a major shift in strategy from AMD / Intel / Apple / NVidia due to this, as I was used to CPU / GPU / chipset upgrades coming out more regularly with double-digit percentage improvements each year, but that seems to be basically over for the consumer space now that everyone is just trying to build as much AI hardware as fast as possible for enterprise markets.
I also used to rebuild my desktop every 1-2 years and upgrade laptops every 2-3 years, but I'm currently a bit past that and have no plans to refresh anything in the near future. I'll probably wait another CPU and GPU generation or two to even consider it, especially with prices being this ridiculous.
I'm fully anticipating AI hardware demand to dry up once Anthropic and OpenAI start tanking financially in a few years after missing earnings and investors and banks wanting their loans paid back combined with more folks realizing that LLMs cannot achieve AGI. After that, I would hope that the demand that's driving this pricing spike will fall along with it, but I also fully expect hardware companies to be as greedy as possible, so maybe a 32TB HDD will just cost $1000 going forward and the era of "a decent gaming PC made up of the tier of components one step below TOTL costs ~$2000" we've been in since the 1990s is fully over, and it's just going to be more like $5k going forward.
> There also seems to have been a major shift in strategy from AMD / Intel / Apple / NVidia due to this, as I was used to CPU / GPU / chipset upgrades coming out more regularly with double-digit percentage improvements each year, but that seems to be basically over for the consumer space now that everyone is just trying to build as much AI hardware as fast as possible for enterprise markets.
I believe all of them realized nobody really needs more than 16c32t on a desktop, and architectural changes nowadays can only go so far to bump performance. The biggest changes across generations in the last ~5 years were the introduction of specialized instructions for wider vectors (AVX512/AVX2, and even then you need specialized workloads to use them), bigger L3 caches and/or efficiency cores to lower power draw. The double-digits gains recently were in memory bandwidth, frequency and locality, because you can make a 5GHz CPU and make it go 10% faster than the previous generation, but you also need to supply the extra data for the CPU to make useful work so that it's not data-starved.
> I also used to rebuild my desktop every 1-2 years and upgrade laptops every 2-3 years, but I'm currently a bit past that and have no plans to refresh anything in the near future. I'll probably wait another CPU and GPU generation or two to even consider it, especially with prices being this ridiculous.
I built my desktop in 2019 (right before the COVID madness, luckily) and I was also used to 1-3 year cycles. I upgraded the CPU and GPU since then, but the performance is still that of a workstation, even though perhaps a little bit behind in pure single-core benchmarks. I plan to keep this until either CPU, RAM or MB fall out (hopefully another 5y at least), and then I'll make a huge leap, instead of making 4x smaller ones which deliver little uplift anyway.
The biggest changes across generations in the last ~5 years were the introduction of specialized instructions for wider vectors (AVX512/AVX2, and even then you need specialized workloads to use them), [...]
AVX2 is actually 13 years (!) old already (first appearing in Intel's Haswell architecture).
AVX512 is also nearly 10 years old right now, but not many CPUs actually support it (none of the devices I own support it)
People think that shorting is an easy way of getting rich when you think a stock will go down, but even if you know it will, getting the timing wrong on it is catastrophic as you'll get unbounded loses while the market fails to correct the issues you are betting against
Totally inane post. Why does he need to justify his opinion by gambling on the stock market? Having a belief about the way things are going doesn't mean he knows when the tower of corruption and graft is going to collapse
I think it's probably for the best, the hyperconsumerism around electronics was getting out of hand. Computers are fast and have been fast for a while now. An iPhone 12 is probably more than fine to do iPhone things, if Apple didn't sabotage their old phones with poorly-written poor-performance software.
Can you provide evidence for your claim? Most people take this statement at face value and I don’t understand why. I do some volunteer work supporting elderly people that have issues adapting to a digital only world and I am always pleasantly surprised by the performance of their old iPhones all the way down to the 8 (old phones they get from their kids)
Apple admitted to slowing down iPhones to keep the batteries from failing, and agreed to stop doing it and created a new battery replacement plan. I’ve personally experienced the benefit of getting the battery on Apple devices replaced which made them behave like new again. There’s a reason old Apple devices hold their value a lot better than competing manufacturers.
They hold value now. They didn’t if we go back to before those lawsuits. Apple can blow smoke wherever it wants, but that 500 million loss isn’t about battery health.
As an aside, I have this lovely tower from France, barely used, scrap price should be very high. Priced to sell?
It was a pretty terrible controversy at the time but I think the battery health settings pane we got out of it was a good compromise and a net positive. What would be even better would be fully user-replaceable batteries.
To be more precise, the phones weren’t slowed down to keep the batteries from failing, they were slowed down to maintain stability and stop random crashes. As batteries age they lose their peak voltage capability, so you’re basically undervolting your components unless you slow them down.
Every time I upgrade iOS past a few major versions there is significant slow down. I don’t think it’s done on purpose, it’s just that software grows more demanding over time.
Nah. Adding more features to software doesn’t make software run slower. The problem is that software developers only optimise enough to make it run fast on our own computers. Because we keep getting new computers, we get lazier over time and software gets slower.
Electron is the most obvious example of this sort of thing, but examples are everywhere once you start looking.
Yes. Microsoft is scrambling at the moment because the macbook neo only has 8gb of ram and runs great. Windows 11 can't compete on that hardware because it's so bloated.
I find it strange given microsoft made windows 3.11, which ran fine on computers with just 4mb of ram. Windows 11 is of course more complex than windows 3.11. But not so much that it should need literally 2000 times as much ram.
In some ways, the memory apocalypse is a great thing. The windows start menu should never have been made in javascript and react. I'm hoping our industry uses this as a push to make our software small and fast again. If ram is 5x more expensive, we should aim to make 2gb of ram be more than enough for regular computing tasks.
I wonder if the computing requirements of windows scale exponentially? It looks like it.
I suspect if we graphed them, they’d pretty perfectly match the computers that were available. Ie, I think windows is optimised just enough to run on the low end hardware of each generation.
I agree, it’s not a natural phenomenon per-se and it is getting worse. I think this consumer hardware problem will actually help get us back on track, to a degree.
That’s the literal truth, but the subtlety behind it was that CPU spikes combined with low voltage batteries were causing phones to shut off. Apple’s slowing down the phone down was nominally because customers prefer slightly slower phones to phones that shut off unexpectedly. Should they have announced this as a feature and had a toggle switch to allow users that preferred the option to disable it? Of course. But I don’t necessarily think they did the wrong thing.
Thank you for the sanity. This is like being being mad at Ford because a 20 year old truck runs worse than a new one, and they’d want some engine design that would make it rev in to redline (and blow up while driving) to keep the original horsepower.
> Computers are fast and have been fast for a while now.
Absolutely. The only reason I want to upgrade to an iPhone 18 next year is because the 16 doesn't have MIE (Memory Integrity Enforcement) and it feels ridiculous to upgrade so quickly to the 17, which does have MIE. Same thing with a Neo 2. But mostly I buy surplus gear. I've saved a small fortune this way and 8% compounding returns (VT and chill) will help me retire early.
Personally I have zero reasons to upgrade from IPhone 13 until the software stops working. I hope it will last me several years more. Beautiful hardware.
See this is what I don't get. This effects big organizations like this, there should be a class action lawsuit for these criminals LLM and Memory companies.
I have a MacBook M1 and I recently had to downgrade from Tahoe to Sequoia. Why? First of all they force you to download 7 GB of AI training when you’re not even using their AI. And secondly on Tahoe just watching a video with nothing else running off-line my MacBook was asking for 7 W of power. On Sequoia that drops down to 2 W maybe three.
It’s sad to think they think of themselves as an environmental company, I don’t know if they do, but this update to Sequoia literally doubled the amount of M1 max that are running on the planet in terms of power use.
Did more research and apparently it’s a bug Apple never fixed that causes the deletion to only work sometimes. To fix it you need to run this Terminal command in recovery mode:
it is difficult to motivate hyperconsumers with the narrow band of improvements that can be made with new LLM models and also are specific to smartphone OS integration and not superseded years in advance by datacentres
I have a 2021 M1 pro and see no reason to upgrade it... ever?
I do web & general java work on it and it's plenty fast even using a memory hog like IntelliJ (which is typically disk bound anyway). Claude tells me that a new macbook would feel about 1.8x faster in my CPU-bound cases, which is not common.
If memory prices returned to normal and if local inference gets over the hump, I could see upgrading for local AI agents, but I'm fine waiting for a good long while on that.
I say all this as someone who, five years ago, was a perpetual upgrader: immediately bought new product and got excited for next product. Wonder if the future will see the end of that culture in tech and, looking back, it was a little silly and wasteful anyway.
I too have an M1 Pro and it does everything I need to do. I bought it in 2023, so it's 'only' 3.5 years old now.
Maybe this is the tipping point where hardware companies can't rely on the upgrade cycle for sales any more. I've always upgraded at least every three years, but with the combination of high RAM prices and good enough performance from an older machine, there is no need to.
I'm in the exact same position. I want to upgrade, but when I interrogate the reason, it's all just empty consumerism. This computer does everything I need.
I'm running a 12 year old i5 3.2ghz CPU with 16 GB of memory, an SSD and an AMD RX 480 (8 GB) GPU. Currently using a 100% scaled 4k monitor and a 2nd 1440p monitor. I keep the machine on 24 / 7. It's great for software development along with video + image editing and casual gaming. Everything feels fast and snappy.
I will use this until the hardware stops. There's a 0% chance I'd buy hardware at the current prices.
I have a baseline iMac Pro and custom built tower with similar specs (3.2Ghz Xeon W-2140B, 32GB RAM, Vega Pro 56 and i7-6700K, 32GB, 5700XT respectively) and they're both perfectly responsive. For many day to day things the difference between them and newer machines is negligible. For other things, my 5950X tower and M5 Max MBP run circles around them.
Both the 5950X tower and M5 Max MBP are deep into diminishing returns territory regardless, however. Aside from some extreme niche workloads like LLMs I can't imagine needing anything more powerful. They both have a ridiculous amount of headroom for the overwhelming majority of anything most people might want a computer for.
Yep, I have an M4 MBP company issued laptop for work and in my day to day there's not a enough of a performance win to feel like I'm missing out for most tasks.
Sure, the Slack web app might initially load half a second faster since there's a lot of CPU bound JS executing but generally speaking I never feel like "wow this sucks" on my main machine. It's kind of the opposite, for a number of things my desktop machine feels faster than the MBP for overall "using the computer", with development heavy tasks (running Docker based web apps, browser tabs, various tools open, streaming music, etc.). It was like that when I ran Windows 10 Pro with WSL 2 and now native Linux with Arch + niri.
Even CPU bound tasks on my 12 year old box that I do a lot, such as editing a blog post on a static site with Hugo takes ~100ms to get live reloading with over 1,000 posts and drafts. I'm sure if I were constantly doing huge CPU bound tasks I'd have issues but I don't have that use case.
After recording 500+ videos, encoding isn't a problem too. If I record a 30 min video, it takes ~20 minutes to render but it's so easy to fit that into other areas of my life, like I'll record + edit a video and then render it while I walk.
It's wild seeing so many comments about updating every 1-2 years.
When I build a PC, I pay a little extra because I want it to last a long time. I aim for 5 years, and in reality it usually ends up 7 years (life happens, need the money, etc).
I am now in year 8 of my current PC. I only had money to replace it this year, and am now stuck with it due to the RAM prices. It's going to really suck for me if my RAM fails in the next few years.
Exactly. I probably spend 1.5x the money on a really solid build, and it's paid off - I'm still running just fine on a 10 year old desktop (overclocked i7-6850K). The only component I've upgraded (other than hard drives) is the graphics card (to a 2080Ti), and it's still running all the games I play just fine. I was planning to upgrade it on the 10 year mark, but now I'm just going to make it last a few years longer until RAM pricing becomes less silly again.
There isn't much to upgrade to if all you want is a desktop for gaming, a non-workstation work box or just a decent laptop. Much like smartphones, there just hasn't been much innovation of late. A midrange gaming desktop from mid 2023, something like rtx 4070 and ryzen 7800x3d, can still play pretty much any video game on max settings and high framerate at 1440p, and has sufficient performance for most workstation-esque tasks as long as you have enough ram. Similarly, the only thing my spare macbook m2 lacks compared to m5 is hardware av1 decoding - the performance is worse, but I don't really notice it in normal usage, and macOS 27 runs perfectly well on it.
Local AI will crush any machine, of course, and if you need a 512gb ram box you'll probably need to sell both kidneys, but you're well past the average consumer territory with those.
So, fingers crossed nothing breaks down and that some of the funny money goes into R&D which will eventually trickle down to consumer devices in a few years (16TB gum stick SSDs at sane prices pretty please) instead of endless datacenter GPUs and SSDs that will be tossed into a shredder in x years.
I've always done 5-7 year cycles on machine I've owned, but this time around its getting quite complicated as my last refresh was 4-5 years ago. So I'm kind of stuck wishing my current hardware both meets the requirements and just survives another half a decade, or my next upgrade, even if only partial, will be VERY expensive...
Similar situation here, my desktop is from 2019, with only the GPU upgraded since then. On the other side, I'm also much less interested in modern games these days so perhaps I'll do fine with a 10+ year cycle this time.
I run Qwen 3.8 at just under 2 tokens/second on a Mac Pro 2013. Why? Because I walked 8 miles to school in the snow, uphill both directions, every day.
My desktop PC turns 20 next year. I've never seen any need to upgrade it, went out of my way to build it with silent parts and it rarely if ever goes outside a single-digit CPU load (before anyone asks, software development and a bit of CAD work). I'll upgrade it when something in it physically fails, which means I'll need to replace it with new hardware with 5-10x the specs just to get the same performance.
My laptop is currently about two years into a typically 5-year lifespan, limited by when the keys start to fail rather than any need to upgrade it just because. The RAM in it is now worth more than the rest of the laptop.
This is one of the rare situations where an out-of-control homelab & PC building hobby is actually useful. With over 250GB of RAM & 90TB of storage I will be fine for a while, especially as long as DDR4 era machines remain relevant.
My issue is my digital hoarding. I've always said it I "collect" things digitally vs. physical, so I can scratch that itch without causing a disaster for my life. A half rack of JBODs is a lot cleaner and overall less expensive to operate than deciding I need to collect every single Lego Technic set ever produced or something.
The problem is I also operated under the prior "storage always gets cheaper over time" and was right at about 80% full when all this insanity hit. I'm now at 90% on my largest ZFS array (~280TB) of spinners... and I get to decide if I want to drop a few paychecks on the next vdev upgrade, or start deleting things.
Not to mention I've been holding off replacing a bunch of old 1TB SATA drives (19 years old now!!) as my scratch partition, as I was going to wait until 8TB SSD got really cheap and just buy two of them for a RAID1 for the use. And now I need those drive slots to upgrade the mass storage pool.
Talk about backing oneself into a corner at the wrong time!
I'm at 97% of my 160TB and I need to start making hard choices :-/ I have Unraid, and an old 4TB drive sitting in a drawer which I could plug into the array. It's old, and I planned to throw it out, but fuck paying these prices for a 26TB HDD right now.
I built my machine in 2022 with 64GB of memory and upgraded to a 4090 at MSRP in 2023 so at this point I'm good but I fear the back end of the bathtub curve. I currently expect to have to hold on until 2029 and at that point the mobo and CPU will be 7 years old, and the GPU almost 6.
same-ish. In 2023 128 gigs of RAM and a 4090 and 4x 4TB ssds. I just looked up prices and 4090 alone is like 2/3 of what I paid for the whole system. Crazy times.
If you bought RAM new, it should be warrantied. Most sticks are essentially lifetime support, but you’ll get a refurb. I definitely wouldn’t buy secondhand now, unless it was a very good deal. Corsair won’t honor transfers unless you’re somewhere that legally forces them to.
GPU probably not a terrible idea to pay the $20 for those aftermarket insurance plans. Do they even pay out though?
Earlier this year I finally bit the bullet and grabbed myself a 5080 for my gaming PC. I wanted to push things a bit more than my 3060ti allowed, and I figured it probably won't get any cheaper any time soon.
The PS5/Xbox Series console generation is supposed to be wrapping up soon even though it barely feels like it got started. I don't sense there's appetite for more powerful consoles anytime soon (I don't believe the PS5 Pro made waves like the PS4 Pro did), and jumps in console performance are what determine when games start targeting higher system requirements. So I like to think I can get a good number of years of use out of this.
How do you know they weren't faulty? Blip flips are usually silent unless you have ECC memory or are actively testing for faulty cells.
Somewhat recently, Firefox implemented memory testing in their crash reporter and found 10% of reports were due to bit-flips[0].
If it weren't for reading that HN thread, I would never have being prompted to check and find the faulty memory in my laptop. It would have gone undetected and continued to corrupt whatever was stored in the affected cells. I did run memtest86 not long after I bought the laptop, so sometime between running that first memtest86 and the 12 months after, it developed the fault.
Non-ECC memory sometimes fails and there's a good chance you'll never know.
Anecdotally, RAM failures are more common than they used to be in previous decades, I assume due to increased capacities and smaller process nodes. Looser manufacturing standards may also play a part.
The component category I've never seen fail is CPU.
I've worked in and around datacenters for a few decades now and I've seen pretty much everything fail, but power supplies in particular do so at a higher rate than I expected. I've had one literally start smoking in my home office recently too. Luckily, it didn't damage any components, but it was pretty dramatic and abrupt -- just powered on my desktop and got nothing but a fog machine coming out of the rear fan until I yanked the power cord a few seconds later.
My evga 1000w has been going strong for 11 years now. Cheaper power supplies fail often but if you buy quality and oversized they’ll outlast all your other components as they don’t become outdated.
It’s the caps that fail. Electrolytic caps have the highest failure rate of any passive component. My team tries to avoid them in our designs as much as possible. But you’re right, caps in a good design can easily last 20 years, making a power supply the longest lasting component of most builds.
My go to source for reviews was always jonnyguru.com but he closed up shop. Not sure who’s actually doing in-depth performance and component reviews these days.
Power supplies - regardless of quality - die pretty regularly at scale. Only disks fail more often. They are the #2 component we replace in the datacenter.
Are data-center power supplies comparable to higher-end desktop power supplies? And how much do you oversize them compared to the load they will actually handle? E.g. for a desktop buying even double the power supply capacity you need won't really affect the overall PC price that much but I somehow doubt data center operators are willing to not optimize away that cost.
They typically are made by the same OEMs and I would say at par or better quality in general. I’ve had (percentage wise) more consumer grade PSUs fail on me over the past 20 years, and I don’t buy cheap stuff. Although the sample size of failure is low single digits vs thousands on the datacenter end.
I’d say on average DC PSUs are vastly overspeced for most machines. Most machines we work with are not sitting at peak load 24x7, and might be drawing half or so of their max rated power draw. Plus every server is redundant - so the modules themselves are operating at around 20-40% of full rated capacity most of their life. The only time they ever spin up past 50% is during power maintenance or a power failure of some sort that brings one of the redundant power sides down.
They do probably operate in slightly hotter conditions on average, but in much more stabilized temperature environments. Plus of course 24x7x365. But that goes for all components.
I’d say other than fans, failure rates in the PC world vs datacenter world are relatively comparable it terms of what fails most. Dust and temp swings are the large variables the average PC needs to contend with vs the datacenter.
Storage of course is the one that stands out the most - everything else is kind of a rounding error. Here though, datacenter work really does beat on the hardware more than most consumer gear so that’s expected. For every 100 disk swaps we do, I figure we do about 2 or 3 power supply modules. And maybe 1 fan swap.
They just rebrand PSUs, not all of them are bomber. But they have 10yr warranties on PSUs and are good about honoring them. I wish they still sold gpus. I would only buy evga for every component if I could specifically for the painless warranty. Had to warranty an asus mobo and it was like pulling teeth.
Yeah I also had an EVGA 1080 Ti die on me, and got a better GPU as upgrade. Their warranty was worth the premium price for sure.
Overall I have made very good experiences with long hardware warranties. Soldered RAM on my ThinkPad died once and I got an entire new mainboard after almost 3 years (it died 2 months before warranty expired.) New one still works today but the hardware is just too old.
Funny, just had an 8 year old 1600W Corsair Titanium efficiency power supply go. Generally very high quality, GaN transistors, all japanese caps, etc etc. It was still under warranty, and Corsair refunded the original price, stand-up company. Too bad 1600W power supplies have apparently inflated ~50% since then, but still.
> The component category I've never seen fail is CPU.
The CPU can fail if the heatsink falls off! That is something that happens (I've had one come off in my hand that was being held on by nothing but hopes and dreams, fortunately not while the CPU was powered on). More likely in a desktop tower form factor.
Even that is not guaranteed to break a modern CPU though as they all monitor their thermals and down-throttle (even below way specified minimum speeds) as needed to protect themselves. Not that I'd recommend relying on that but still, I have had cooling fail without apparent damage.
Most devices never test their memory and don’t have ecc, so you generally never know. Random instability can be fixed by restarting and your program not being in an area with bad bits anymore.
I recently had part of a 32 gig DIMM go bad on a 128 gig system. One of the larger VMs was spontaneously rebooting for no apparent reason. I was able to find the addresses with MemTest86 and map out 512 megs of the affected DIMM with the Linux kernel memmap option.
Don't send it back. Since they can't easily replace it with new hardware, they will probably refund you the price of purchase. Which is normally a very good deal, but obviously not in these circumstances. I've seen it happen with GPUs as well.
I've seen every single component of a system fail, the least I've seen fail has been the CPU itself (I've only seen one fail out of thousands).
Memory seems to fail pretty commonly overall, less common than harddisks but more common than motherboards.
The issue is that I'm used to using ECC ram, which fails loud when it's actually bad.. consumer memory won't tell you unless you can't boot.. and everyone disables the startup memory testing too.. (in fact, I think it's disabled by default for the last 10 years because people want to boot quickly).
I was so sure that a CPU wouldn't suddenly fail that I ended up replacing every other part in my computer before getting a new CPU and realizing that was the issue.
I had a 2012 mac book pro that burned out 2 different crucial memory sticks. Would have bricked the machine if it was the soldered-on type, but it was the last model apple sold that have swappable memory. Crucial had a lifetime warranty so they replaced them for free.
I used to game on it and I think the design just couldn't handle the cpu and dgpu being active for long periods, it would get extremely hot in the area close to where the memory was.
Granted, I've been building my own computer for 30 years and I've only seen it happen ONCE.
But I've seen plenty of other failures:
- Two hard drives -- One got dropped on the floor, so it was no surprise, the other was showing degraded performance and SMART showed some scary numbers and I was able to replace it before I lost data.
- One GPU, replaced under warranty.
- One AIO water cooler, replaced under warranty. Somehow, the water vanished from the loop. I'm guessing a microscopic leak that evaporated as fast as it leaked.
- Two power supplies -- One randomly exploded. Just started making popping noises and shooting sparks out the back. At autopsy, I determined that the fan failed (it was hard to spin manually) and it overheated. The other was just being overloaded. I had just gotten a new GPU and the PSU wasn't big enough for it.
- Several case fans.
Never seen a CPU or motherboard failure, even when my AIO water cooler was failing and my CPU was constantly at thermal limits. Never had an SSD/NVMe failure.
Anecdata: a coworker found a single bit flip corruption in a file he wrote in his computer. He ran memtest86 on his memory, and which found one of the two sticks was defective. The computer otherwise ran perfectly fine for years; and we would have never noticed if the file format didn't have built-in checksums.
> I've been using computers for more than 40 years now and the one component I have never seen fail is RAM.
I think over the decades I got one RAM disk fail. And another one didn't fail but, although I bought a pack of four sticks (2x x2), was detected (by memtest, when building the rig), as having another model name than the three others: even the shop who sold me the RAM was confused (don't know how that happened).
The usual component that I've seen fail the most would be the PSU? (before I started buying quality ones). One NVMe drive (an "ADATA") died on me even though it was nearly new.
PC assembled with quality parts usually last a very long time.
I also have been using PC's for a long time, thirty years. The only time I have seen ram fail during use was within the last 5 years. My 10th gen platform and 32 gigs of ram DDR4 and one of the sticks failed, this was before the ramapocalypse of course, the sticks were replaced under warranty. But I have had systems in my shop with failed ram. it happens, but its rare, unless its cheap ram, less rare.
I have had a RAM stick start frequently failing ECC checks after a couple years use. I also had the VRAM on a GPU fail, but you could blame that one on overheating from the nearby compute core. I agree is rarer though.
Oh well, it's fairly common.Maybe not in a singular/several rarely used (and not monitored) desktop computers, but.... it's common once you have quite a few under your care.
on older PC bios would perform a memory test - you may remember those - and that's where it would mark a bunch of pages dead/unusuable and it was not uncommon at all.
I've seen some fail. Might have had to do with the fact that the computers were literally flooded (river came up faster than the sandbaggers could keep up) while running.
I've seen computers brought back to life after flooding by lying in rice for a week, but they weren't turned on when the water hit (and not all survived).
I've had a few RAM sticks fail, but the ECC usually caught it first. Also had one case that somehow the stick slightly unseated itself; dunno how but it was out of service for months until someone opened it up and reseated the stick.
...granted though I've also seen a backpane on a server fail. I didn't think that was even possible until that point.
I would absolutely love if some startup came up right now to manufacture and sell RAM at normal prices to everyone else and put Crucial et al out of business for price-gouging everyone just because they can.
DRAM fabs are multi-billion dollar investments which take years to come online. That's the whole reason we're in this situation, there is no quick fix available.
DDR4, and from some comments I have seen around here of only marginally acceptable reliability. It may still take a few years if until they have all the production processes fully under control.
China is making central decisions for its RAM makers (and who can complain now that the US isn't a legitimate free trade defender). I think their strategy could go in any direction to either get a small portion of still gouging or a large portion of a market with somewhat stale specs.
I wonder if supply chain constraints currently represent an inherent scepticism of manufacturers that demand will last. It feels to me like a side effect of the self-dealing / incestuous financing that is going on is that manufacturers are not willing to bet on it all materialising and therefore are not scaling up capacity nearly as much as they would if they thought it was certain.
It's funny you mention that because this exact same issue has come up in war in Iran and the closure of the Strait of Hormuz.
What a producer will do now is hedge their risk. They do this by forward-selling their product, in this case oil from an oil well. The peak for spot prices, depending on the particular flavor of oil, peaked at >$170/barrel while future prices (which are the prices we see) didn't really get above $100 while long-dated futures were significantly lower, rarely exceeding $70/barrel by much.
Now this is a problem because there's no incentive to invest in extra production when you can only sell it at $70/barrel, despite the spot price being up to $100/barrel higher. So new wells and new production in general is down despite the supply disruptions and high prices.
I imagine RAM is in a similar position. It takes a long time to bring on more supply and suppliers don't expect those prices to last. Likely a large amount of new supply will probably cause the next crash in the boom-bust cycle.
But there's another problem with all this. A lot of existing manufacturers are switching from DDR5 (and soon DDR6) to HBM because it's way more profitable and there's seemingly insatiable demand for high end GPUs. I expect that in the short-to-medium term we're both going to see no increase in supply and the DDR prices are going to keep rising to the point where in a year we'll probably look back on current prices fondly
All I can say is I'm happy I upgraded my laptop to 64GB (for $200) and my PC to 64GB (also for $200) last year.
.
> A lot of existing manufacturers are switching from DDR5
Nobody is "switching" old process node fabs to new processes, they just aren't building new fabs for old processes.
AI and high density tasks will use memory fabricated with newer processes, freeing the old supply for more elastic, price-sensitive tasks.
The vast majority of memory that has price impacts for most consumers isn't in personal computers, it's in embedded applications that were also impacted by the overall squeeze.
9-Aug-2026: SK Hynix to invest $38 billion building new memory chip plants as demand soars [1]
"In the AI era, technological competitiveness alone is not enough and the ability to supply the required volume at the exact moment customers need it is the ultimate competitive advantage," SK Hynix said in a press release. "We reached this investment decision after a thorough review of market demand."
They will do so in steps, tho. One cleanroom at a time so they don't eat a big loss if this crazy demand for a commodity goes away overnight.
They also took quite some time to decide to ramp up, as this comes a year or more late after they announced they were fully booked for 2026 and 2027.
And, their stock sold off on the news, so investors are not having faith either on the demand to justify capex expansion.
Honestly, it sucks to be a ram company, you are either buying into the hype and then spending years losing money, or you don't and leave a lot of money on the table while competition ramps up and eats your lunch (and then spends years losing money)
It's exactly this - the memory market has gone through exactly this boom/bust since I can remember - there was a great RAM price stabilization around Windows 95 (price stayed the same or went up per megabyte over years) and then a glut of oversupply.
TSMC in the US is a bit of a special case .... wouldn't surprise me at all if they were virtue signalling to the Trump admin more than anything with what they are doing there.
How so? They are diversifying their production chain so that all their eggs aren't in one basket. Taiwan might be threatened one day but their US, Japan, and German plants will much less likely be in the same danger.
That seems wrong to me - the idea of the "Silicon shield of Taiwain" is that alle the critical cutting edge production is all physically in Taiwan & rigged to be destroyed if Taiwan is ever attacked.
If they do diversify, they are compromising a key policy protecting their homeland, it makes no sense!
Building new fabs is a very different bet from selling into a temporary shortage: you're committing billions based on what demand will look like several years from now
Manufacturers know demand will last to the extent that big companies are willing to commit to purchases over a given period of time, but supply today is still constrained by capacity built for the pre LLM era. The 500 billion put in by Samsung/SKHynix will increase modern DRAM by roughly 50% but it wont be fully online until 2030.
It might be fully online by 2030 if the current demand lasts. IMHO the demand is unsustainable & build on circular financing deals and hidden debt with no way to be ever profitable.
The memory cartel would be crazy to actually invest into new capacity at this point.
If you want even more bad news: display panel makers also want to raise prices
"Due to the continued rise in the prices of core components globally, the production cost of monitors has increased significantly, and we can no longer absorb this pressure on our own."
It would strange if just when a lot of articles mention computing / social networks are detrimental, governments start passing laws to forbid platforms and/or phones for kids, and now price gouging flares up everywhere, this would entice people to just stop using these devices and maybe enjoy natural taper-off.
People do not want to talk to their devices because they are in the room with other people and they dont want to disturb them. I think their device is going to be a flop.
It's going to flop because everyone is already able to talk to Google Assistant to manage things, and outside of very specific circumstances (e.g. driving) everyone is still tapping screens instead
Doesn't solve the present shortage, but perhaps this is a good opportunity for software developers to start caring about memory usage again, since consumers may be forced to keep their existing hardware for a few more years.
Honestly? Is it even going to be a drop in the bucket? This push is from massive data needs. The amount of ram code uses becomes an even smaller percentage of the growing pool.
Jokes aside, I've been extremely frugal with memory in my latest project, in the hope that it can fit in a few megabytes of RAM. But we have come a long way since the eighties, and now an audio codec (OPUS) takes the best part of a megabyte, and so do "frugal" scripting engines such as Lua and even SQLite \o/.
I almost upgraded my computer a year or two ago, but decided to hold off for a bit longer.
I've always been fine with being "behind" on the latest and greatest. I don't mind watching a movie, playing a game, buying a cell phone, etc, that was released a few years prior. If it was really all that great when it came out it should still be good later on.
Buying slow usually saves a.good bit of money when prices go down over time... but this time they went way up and now I'm kicking myself on it. Not upgrading the computer for now, but it is a bit overdue since I had already put it off for many years.
Except staying on the edge costs exactly the same as not, when you take the resell value of the components on the second hand market into account. The problem is that some can afford to lock their capital in the hardware (via one-time investment) and some do not.
R9700 will lose less than 30% of value in the upcoming 3 years. Give or take. Yet people buy ancient cards for way more than 30% of R9700's cost.
> ...when you take the resell value of the components on the second hand market into account.
When taking the resale value of used kit on the second-hand market into account, one must also take into account the hassle of selling on the second-hand market.
For me? Personally? That hassle absolutely eliminates any money I would get from any plausible sale price. If I'm getting rid of used kit, it's going to be as donations to people I personally know who will come and pick it up at the agreed upon time.
Yes and I seriously doubt gps insinuation that resale could make up for the price premium of top of the line components when they were consistently beat by the next years significantly cheaper mid range components.
I thought about upgrading my macbook, but I have the M1 Max and it still doesn't feel slow, so I decided it'll probably last another two years. That said, now I feel like this price raise is maybe permanent. If so it will definitely eat into computer makers profits, people will not be upgrading nearly as often.
I ordered a new CPU on sale a while back, right when the memory prices started spiking. I didn't even know RAM was going up in price. Realized the CPU needed a new motherboard and ram so I ordered the motherboard, started looking at ram and realized it would cost almost as much as the Mobo and CPU put together. Cancelled the orders because I wasn't intending a big expensive upgrade.
What I have is still fine, though I do regret cancelling now.
I upgraded a year ago and I bought a bunch of extra RAM because it was "so cheap", why not. Man did I get lucky, I got it right before the price hikes started.
Same, my desktop is a top end full of ram beast that I thought I overpaid for because motherboards had gone up. Bought everything before the surge. Past me was damn lucky. Wish I had bought some stock tho.
There is no law that states memory must be a consumer product. I truly believe the era of personal computing was a beautiful blip that our masters are beginning to take away from the general public. It could be terminals and time sharing for the foreseeable future through "smart" phones.
Personal computing has been in a sad state for a while now. Only "PC enthusiasts" were gamers. Rest of the population uses their phone 18hrs a day and seldomly use PC for Excel tasks at work.
Apple is making Macs and iPhones more powerful every year. Macs are more affordable and powerful that other computers since the M series. People are buying them to run their own private on device AI. This is a signal for personal computing.
Is the end result of this is that anything that is useful to AI going to be permanently more expensive, as the value a human hobbyist can get out of a certain amount of RAM will always be far less than the value a frontier AI model can get out of it?
If demand remains as elevated as it is you will definitely see new fabs breaking ground to meet it. China is expanding a lot as well. Even if AI doesn’t go anywhere there is no way prices remain this elevated or climb forever
More likely AI companies go bankrupt and market is flooded by repurposed hardware sold for pennies by creditors AND Chinese memory from their new fabs, hopefully also killing the existing memory manufacturing cartel.
My theory is demand substitution. NVME/SSD is up so much people at the margin are switching to spinning rust, which nobody was prepared for, so now they too are drowning in demand.
(2) If you're WD, Seagate, etc. you may shift all your fabs to higher growth areas which means you create less supply for less demand products (spinning disks)
Boy am I feeling this. I had my eye on building a new Threadripper machine, which requires DDR5 RDIMMs. Much to my surprise while happily building out a parts list one fine evening I learned that these are going for between $500 and $1000 for one single 16GB module right now. I think I'll hold off and squeeze a few more years out of DDR4 and PCIe 4.0.
These unintended (or intended?) affects of the AI boom are yet another cause for hate among a key demographic (people who use/buy/build PCs) who might otherwise be pro-AI but because of this are quite anti-AI.
Inflation, cost of living, purchasing power... I feel like life is getting really difficult right now for most people in ways it hasn't been in decades. I don't think memory prices going up this much this fast is good. I think it's a sign of how bad things are and that we are heading in a terrible direction. I think we're going to get a BIG wave of inflation in 2027.
On a macro level you're correct. On a micro level, I miss the pre-pandemic, pre-AI era before stuff got expensive, before Russia invaded Ukraine, before the Israel/Palestine strife went into overdrive, and before the US declared war on Iran, Canada, Greenland, and Minnesota.
Almost everything that is upsetting is you, is being intentionally shoved in front of your face to make you feel mad. You are being played.
Your great-grandparents had like a 40% child mortality rate. You’ll be ok that some people around the world are fighting which has been a human constant throughout all of history.
By what metric? Houses, healthcare, cars, college... everything that matters is way more expensive for current generations when you compare to multiples of income or ounces of gold. We're working harder for less to show. Yes, there are FSD vehicles and rockets that can land and be re-used... but steak is up 30% from a year ago. In most ways, life is going in the wrong direction. The only thing getting 'better' for most Americans is that you can get a bigger TV now than ever.
Your grandparents almost certainly had siblings or knew of their friends siblings dying from now completely preventable diseases.
The quality, safety, comfort of low income housing in the United States is exceptional in historical terms. The fact that you CAN be anything you want is an unheard of concept even 60 years ago.
The worst cars on the market today are modern marvels of engineering compared to decades ago. Let alone throughout the history of the country.
College, lol, you literally have a secured loan to do study whatever you want with absolutely no required direction from anyone else… to receive the worlds knowledge if you want it.
Oh, things are more expensive? Yes, they’re massively higher quality and larger.
You’re mad about things, OK. They’re objectively the best they’ve ever been. This is indisputable, but again, I get that you WANT to be mad.
Okay Ben Shapiro, but you skipped an entire generation to make your argument and are ignoring what's happened in the last 5/10/15/30 years.
I don't WANT to be mad and that shows you're arguing in bad faith. I want to stop checking my bank balance before I buy toilet paper and paper towels. Life wasn't like this until very recently... We're destroying the middle class and making home ownership virtually unaffordable. The average first time home buyer is 40. The economy is not healthy and has not been this disconnected from the stock market since probably 2007.
> By what metric? Houses, healthcare, cars, college... everything that matters is way more expensive for current generations when you compare to multiples of income or ounces of gold.
Check the price of food. Also, check what the price of a movie resolution film camera cost 20 years ago and realize that you're getting that in the price of your iPhone. Plus of course the corresponding computing power.
last year i mis-bought a ssd on amazon (wrong size or model i can't recall) so I asked for a return, surprisingly amazon wanted to save on transport that they offered to pay me to keep it (i don't want to spread stupid claim but maybe the full item price), i was confused and worried it would create issues down the road so i sent it back thinking i'd buy the proper item the month after.. and now i'm running duplicate search programs to save space on drives.
They are building fabs on track to get online towards the end of 2027, full capacity production expected for 2028. These are serious pieces of very complex technology that take years, not months, to build.
It's very unlikely to have any meaningful effect. CXMT DRAM will also be hoovered up the moment it hits the market at the same inflated prices. Even if they had the capacity (they don't) why flood the market when you can just print money?
And that's the really fun part of all this, that weak substitutes will also become more expensive. But on the bright side, I've got dozens of sticks of DDR3 RAM that I'm hoarding like Pokémon cards. Any day now, we'll get desperate enough to start using them again!
I would be very surprised if the cost of the inputs being hoovered up by AI drops after this. Without ever forming a real cabal, I think the functional behaviour of the disparate people in this supply chain are going to hold on to the profit margin because there isn't a clear basis to stop them.
Sure, some imaginary perfect market would drop. This one is far from perfect, and it may drop by marginal percentages. It won't drop back to what it was before it went up 4x in price.
I would be very surprised if a market for older, superceded chips takes off, not because it isn't a brilliant idea, but because in order to use them, you have to make boards and systems which can consume them, and right now, the people who do that for a living are making a good living selling stuff as inputs to the people making a good living selling memory at inflated prices to the AI factories.
If this was a competitive market some suppliers would be ramping up pc memory and avoiding the AI buildout. But instead we see all of them diverting to GPU memory and cutting production of PC related parts universally. This sure looks like a coordinated move.
I think there is a real point people are missing here. Memory is going sky high because most of it is being shifted to GPU specific ram for server build out. This means that when this ram hits EOL it -won't- flood the consumer market in a meaningful way. It is disposed of because most people won't put a server GPU into anything. It is like a used big rig, built for purpose and when it reaches end of life it is truly at end of life. Consumer DRAM will stay very high while the cartel retools and moves back into that space. This means that the cartel wins and pays no real price for it when they shift back to producing consumer ram, they step immediately into a massive demand for pc RAM and keep raking in $$$. But they all have to do it in lockstep so that they have a market to come back to. Otherwise someone would become the king of PC ram and they couldn't re-enter easily if/when the AI server market fails. Being in lock-step means they rake in more money and mitigate the risk of an AI bubble, all at the cost of the consumer.
At least that is what this looks like to me. They know they could make a lot of money ignoring the AI server build out but they all magically didn't. A competitive market would have seen a spike in price when this much demand came in, but the lockstep movement of the big three shows that they are coordinated and not competing.
> This means that when this ram hits EOL it -won't- flood the consumer market in a meaningful way. It is disposed of because most people won't put a server GPU into anything. It is like a used big rig, built for purpose and when it reaches end of life it is truly at end of life.
This was true until the last year.
P40, P100, V100, A40, etc are selling like hotcakes and prices have been rapidly increasing.
Not sure about that, V100s are all selling for less than $1k despite being 24-32GB VRAM around 1TB/s memory bandwidth, which is the same price point that 4090s and 5090s are commanding $4k-$5k for. Hardly hotcakes.
500% in a year is just insane. I knew RAM prices were going up, but $3k+ for 128GB of DDR5 makes me wonder how much of this is actually AI demand, and how much is this just manufacturers taking advantage of it. They have done this in the past after all
I don't understand the distinction you are making. If you have a massive excess of demand ready to pay any price, and a limited supply, of course the sellers will take advantage, that's how any market works. They can do that because of the lack of supply.
Over the long term demand will rationalise and supply will adjust. But in between there is a squeeze, and the right price is the maximum price buyers are willing to bid for the limited supply. Sucks for other buyers, on the other hand it ensures demand destruction, i.e. that the limited supply goes to the buyer who need it the most, other buyers will delay their purchase and do with what they have.
> If you have a massive excess of demand ready to pay any price, and a limited supply, of course the sellers will take advantage, that's how any market work
The question is how quickly the supply adjusts. In a free market, the suppliers would respond to the price signal by quickly expanding their production to meet the elevated demand.
But a market captured by a cartel will be much slower to expand the supply.
So, the test isn't "how much is price going up" it's "how aggressively are the RAM producers expanding supply production"
Memory fab build out times are years not months. The existing ones already run 24/7. If these companies could pull more supply out of thin air they would, but they cannot.
Cartels aren't known for massively increasing production to satisfy all demand. Far from it I'd say, they're more likely to expand supply to such a level that they can still reap massive profits. Very few countries/companies have the knowhow and willingness to be able to build more.
OPEC is a perfect example of this behaviour. They set hearly production goals to keep prices high. Or the Quebec Maple Syrup Producers organization that limits the amount of syrup a farmer can produce and sell in a year to keep prices high.
OPEC's aim isn't to keep prices as high as possible. If they go too high, people will start buying electric cars, investing in renewable energy, etc. Basically they want to make oil as expensive as possible while preventing it from getting so expensive that it's disruptive.
OPEC's goal is as much profit as possible. They're not thinking in the short term but long term. If as much possible means over the span of multiple centuries then that's what they're going for.
I think we're in agreement here. I don't think memory manufacturers are interested in a market disruption even though profits are incredibly high for a few years. Make it too profitable for too long and others will arise but keep the supply at a steady pace with decent margins and they will remain the only ones on the block because of massive upfront cost.
China is, to my knowledge, the only country that has so far had the willingness to seriously invest. Not necessarily for profits but for independence reasons. I don't see many (or any really) other countries following suit.
Maple trees are living things and have a right to reproduce, you can't fault a bird for shitting maple seeds all over my farm planting maple trees everywhere.
It looks like you're forced to participate by law once you sell over a certain volume.
> Backed by the Canadian civil courts, the federation has the monopoly for selling Quebecois maple syrup on the wholesale market, and for exporting it outside the province. Producers are only allowed to sell independently cans of less than five litres or 5kg to visitors to their farm.
https://www.bbc.com/news/business-35028380
I saw a documentary(forgot the name, sorry) about that that said, no you can't choose not to join. I forgot if it was about this organization but it was about maple production in Canada where a farmer wanted to do their own thing and they were intimidated and all sorts of nasty things happen to them because they didn't think the org was just and wanted to sell freely. Very interesting documentary and an eye opener that even farmer unions can act like organized crime when there's money involved.
> you can't fault a bird for shitting maple seeds all over my farm planting maple trees everywhere.
If only building a ram FAB was that easy :).
The best hope is that cxmt gets going quicker than expected and is able to satisfy a decent chunk of the Chinese demand for RAM which would - in theory - reduce the demand for ram from the few other producers.
You're correct in theory. I'm just not sure we have a cartel situation. Micron is a US company, and it controls 25% of the global market. I see no evidence that they nor SK hynix nor Samsung are intentionally limiting production.
I do see people arguing, "they had years to increase production - why didn't they?" They've been remarkably candid about this: because RAM production has been losing them money in the past. Because they don't have a cartel, prices were on the floor. Investing in more production would have been corporate suicide.
Even now, there is considerable risk that the bubble will burst, and all this investment in new production will be flushed down the toilet. RAM production is a brutal business.
But are the memory chip producers acting like a cartel? In the past they have not. When there was a shortage, they increased production as rapidly as they could build the new fabs, crashing the price of memory.
As others have noted they've been found colluding previously, more than once. Even admitted guilt and some people went to jail over it.
And the reason the market is so small is because of players who did as you propose, the DRAM market crashed on them (not due to them but due to broader market dynamics), and they went out of business.
So there is, quite perversely, strong incentive to collude if possible and if not to abide by an unstated agreement to be extremely slow to expand DRAM capacity.
That said...
SK Hynix has two fabs under construction, one set to open later this year. One for 80k wpm, another 360k wpm. A third was recently approved to start construction.
Micro has an overhaul of Fab 6 in Virgina near done and slated to produce first wafers this year. It also has a large expansion in Boise nearly completion for 300k wpm. They're also building a "megafab" site in New York (my guess is >1m wpm) and expanding at their Japanese plants (40k wpm).
Samsung has two under construction, P4 for 100k wpm.
Nanya - historically a secondary player focused only on producing non-leading-edge DRAM is making a play for the leading-edge market with a new fab dedicated to DDR5 and HBM.
It appears they've all decided that the demand is sufficiently sustained that building new fabs is worth doing... but it takes time. If my guess is right Micron is making the largest play with their New York site at 1-1.5m wafers per month. That would give them a large chunk of the market.
Well guess what :) They've done it before, they will do it again.
"To date, five manufacturers have pleaded guilty to their involvement in an international price-fixing conspiracy between July 1, 1998, and June 15, 2002, including Hynix, Infineon, Micron Technology, Samsung, and Elpida.[1]"
That was 24 years ago. Half the people commenting on this board aren't even that old. How many of the shareholders on the boards or executives at the helms of those companies do you think are still there?
Wasn't that bad for business last time? I could see them erring on slower build up because of how it went before, if they're not sure how long this demand will last especially
Which is exactly what's happening. There used to be a lot more then 3 makers but up till very recently memory was a comedically low margin business you didn't go into.
There are several Chinese fabs that they could contract out production to. In fact that's exactly what Apple tried to do, before the U.S. government went full protectionist and added the two memory manufacturers to the banned imports list. There's also the Texas Instruments fab that's underutilized right now. The extra capacity exists, it's just Micro and SK Hynix are throwing their money around to stifle anything that could jeopardize their spike in profits that are historic not only for their industry, but for manufacturing as a whole.
Large barriers to entry are anathema to competitive markets, so the conclusion here is still that firms are benefitting from the lack of a competitive market.
Memory fabs take years to build, they cost a lot and there is no guarantee that the demand will stay as high years in the future. It's all very complex.
I'm curious what the limiting factor is for "years not months" when building these fabs. DRAM is a commodity, there are only threeish major manufacturers, all have huge amounts of capital, multiple nations bend over backwards for new semiconductor production projects, and there seems to be unlimited money going into purchasing these things.
Even when it comes to manufacturing the tooling/testing infrastructure, it's just a question of money aiui.
> But a market captured by a cartel will be much slower to expand the supply.
Oh yea, governments can fix by putting 100 billion in memory fabs right now, and when demand crater in decade or so, let tax payer eat the cost of those fabs. At least it will benefit poor computer users like me.
As has already been pointed out further down, OPEC is the perfect example of this. Things don't really work that way when the time scales are so extended. What we need is fracking, but silicon.
This system is too complex, the few individual players too big, with too many artificial distortions for the adjustment to equalize out in a way that's anything but years-long (it may have shock-waves that extend past the decade), and highly destructive to everyone who isn't OPEC, or their largest customers, in this analogy.
The brunt of it will be paid by the people further down the food chain.
Even if the memory market was completely competitive the time it takes to expand supply is measured in years. There was no way for them to serve the AI demand without reducing supply given to everyone else in the short run
> There was no way for them to serve the AI demand without reducing supply given to everyone else in the short run
I didn't say that there was a way to do it in the short run. I laid out what the test is to determine if there's cartel-like behavior or competitive-market-like behavior.
I will say, in a market with only 3 players and very high cap-ex cost, my prior is that you're likely to see more cartel-like behavior than competitive-market-like behavior, but that's still to be determined.
Generally speaking, if demand stays high, and existing suppliers don't increase supply, the high prices create an incentive and opportunity for new market entrants to increase the supply.
Yes, "quickly" is obviously industry-dependent. If they were fabricating something extremely easy to manufacture, "quickly" might mean months-to-a-year. Fabricating something very complicated like memory "quickly" might mean years.
That said, even though we might have to wait years for new supply to hit there market, we probably don't have to wait years to see signals of that. We will likely know that manufacturers are starting to build new supply long before we see the chips hit the market.
In the current environment unless people start decommissioning datacenters en masse supply will be locked up till the early 2030s IMO.
Even if you started a new fab today, it's 5 years till it comes online. The great hope here is that China's DDR5 yields represent a high reject rate of an enormous amount of new capacity that they'll dump into the market to get a sovereign capability in that area.
> According to the one-count charge filed in San Francisco's federal court on Thursday, Park conspired with unnamed employees from other memory makers to fix the price of DRAM sold to computer makers from April 1, 2001, to June 15, 2002. The government says the move directly affected sales to U.S. computer makers Dell, Hewlett-Packard, Compaq, IBM, Apple Computer, and Gateway.
"that the limited supply goes to the buyer who need it the most" - I would say that the limited supply goes to the buyer who can afford to pay higher than others. Someone may need the memory more, but may not have the deeper pockets to pay for it like NVidia and Apple might.
I would have framed it as "buyer who can get the most leverage from it", which equates to being able to spend more. It's an economic argument, not a values argument.
>that the limited supply goes to the buyer who need it the most
Everything else you said is accurate, but this is not. Supply will go to those most able to pay, not those who need it the most. It is not a moral allocation, which your language implies.
Not really. They can get RAM just fine, and delay upgrade/refresh cycles to stretch existing component lifecycle, which was already very long if they needed to.
They can get RAM at drastically more expensive rates. And not all critical applications have unlimited budgets to absorb the increase. It’s the reverse Jevon’s paradox.
it is probable that you don't get the deepest pockets by wasting money. anyone that pays so much for memory means they expect to get even bigger value from that.
This is a misunderstanding of bubble markets to an extreme degree.
In a bubble market where B has massive amounts of capital allocated to it in a way that will likely collapse in the future than it's very likely they will never recuperate the 2M they put in the RAM. Instead buyer A that would get at least 1M of usefulness if not more, gets far less, crippling their business. Then when B collapses a ton of RAM is dumped on the market for far less than 1M cratering the RAM providers business for years.
> I don't understand the distinction you are making.
It's pretty simple. In a correctly regulated market the sellers are in competition with each other so there is a limit to how much they can increase prices, since they'll just lose business to their competitor.
In a poorly regulated market that allows competitors to collude with each other to raise prices by the same amount, they'll inflate prices much further.
>I don't understand the distinction you are making. If you have a massive excess of demand ready to pay any price, and a limited supply, of course the sellers will take advantage, that's how any market works. They can do that because of the lack of supply.
The question is how much of this demand is actually based on AI, or just an excuse to fleece regular buyers (and AI is not really the actual demand driving such prices).
Memory manufacturers have driven up prices via collusion many times in the past (and paid fines for it), without AI or some similar high demand vertical market existing.
But I don't think it's the case now, see my answer elsewhere here for why.
In a market with healthy competition (no collusion), this generally can't happen. Every single vendor is incentivized to lower their prices slightly at all times, with a floor at cost plus margin.
>In a market with healthy competition (no collusion), this generally can't happen.
We're talking about the memory market here.
The "healthy competition (no collusion). Every single vendor is incentivized to lower their prices slightly at all times, with a floor at cost plus margin" is the fairy-tale version they teach at 101s.
Floor is below cost if you think you can put one of your competitors out of business before you run out of money, and you'll make the money back after doing so. And this includes even after fines.
I'm sure there will be no large scale leverage campaign against a country who is enabling profiteering at this level. No defense agreements removed, troops removed from the country, or large tariff increases around failed investments.
I'm sure it will only suck for the buyers. This has no effect on inflation and probably will be overlooked by the current admin.
In an absolutely free market? sure. But we don't live in an absolutely free market, the memory manufacturers have refused to budge on creating new fabs until recently out of caution and to be able to create a demand excess.
They have slow rolled rollouts and everything, if there wasn't CXMT they would probably still wouldn't have budged on building new fabs for a while longer.
You can absolutely price fix/manufacture a shortage.
Ever heard of merchants setting fire to crops after hoarding to sell at a higher price?
The issue is the distributer and producer are the very same people this essentially means producers have no incentive to rationalize supply, they can make much more money of the long drawn out shortage.
I think everyone is aware now that you don't build a fab by snapping your fingers, and you need to amortise it over many years. The said manufacturers are just out of a massive slump in demand during covid, when they were selling at a loss, and these cycles happened many times in this industry. It is understandable that they don't jump all-in on the AI bandwagon, which, god knows where it is going and for how long.
If you want evidence that no one predicted this, look at WD. They disinvested from sandisk up to early this year. Now sandisk is worth more than WD. Doesn't seem consistent with a concerted strategy to gauge prices.
> You can absolutely price fix/manufacture a shortage. Ever heard of merchants setting fire to crops after hoarding to sell at a higher price?
Fine. Show me which manufacturer is cutting production to corner the market. I don't see that.
Yes when there are large fixed costs to expanding production, that will happen more slowly, after determining that the high demand isn't just a short term spike.
It's not like wheat, where you just can just plant more at marginal cost (I'm sure someone in farming will now explain why that isn't exactly true either).
I'm not super familiar with farming, but I'm pretty sure there is an upper bound on how much wheat you can plant dictated by the amount of land you own.
Big manufactures have absolutely cut production of lower margin products and it's not like they can re-purpose all of the equipment used for it for HBM yet. edit They can definitely focus more resource because of it on HBM though.
Things have improved a bit since govts are starting to get angry.
Tldr; no one is cutting the production of high margin stuff, but low margin stuff can go and die, because it's not a free market it's a stock market these companies want to show off better margins in every unit there is no reason to keep inefficient things running when not doing it pays more.
They just don't want 30-40% margins anymore. That is not free market that is quite literally serving their own interests because they know there is a monopoly.
All big memory manufacturers are heavily supported by the govt, that's what has made them the monopolies they are Micron supported by USA (they crushed Japan's memory industry for it), and Korean giants are more tangled up with the govt than well most people in US would realize.
Somehow people think free market means anything goes, which honestly is where we stand today so I don't blame anyone.
But using govts to crush competitions and build yourself up on subsidies and similar isn't really free markets. China hasn't been able to grow a significant memory chip industry yet because of lack of EUV machines and similar technologies blocked by the US, which is also not "free market".
As for WD, and Sandisk separating I feel like this was more mutual and they new HDDs or SSDs being together was a distraction for both. Which feels more strategic, think of it like this, the shareholders who had the shares how hold shares in both and together they might have likely been worth less so shareholders made money.
Aka good for free market. I don't see how that relates to the actual memory crunch issue more broadly.
Margins pre-AI were more like 5-15%. Even if the margin was 5%, if prices increased 5 times but product cost didn't move, then margins are now 80%. That's not a 30-40% product, your argument doesn't make any sense.
> the memory manufacturers have refused to budge on creating new fabs
That's EXACTLY how a free market is supposed to work. The existing memory manufacturers are FREE to not build new factories, and anyone else is FREE to build those factories in their place.
Then perhaps we’re seeing one of the downsides of that. When 0->1 is prohibitively expensive, existing players can continue to raise prices *very* high because cost-of-entry is extreme, until it’s not sufficiently extreme.
And then we have bubble issues because the introduction of a new, cheaper player, causes the bottom to fall out and for the circumstances that allowed the new manufacturer to exist to completely go away.
This is also caused by the FREE market - consolidation, all markets tend to consolidate around a few companies due to economy of scale.
The fix is regulation - we are going to have NOT-FREE-MARKET electricity, NOT-FREE-MARKET water supply, .... for vital services.
I'm not talking about you, it's funny to see anti-regulation HN complain about a FREE MARKET working as such. Suddenly it's not so fun where you are at the losing end of the free market when it turns out some are willing to pay way more for RAM than you are.
Be so kind then and tell me what a normal citizen would need more than 8GB of RAM for, if it isn't for professional work or a specialized hobby?
A Mac with 8GB of RAM is blazingly fast for any task your average person would need for it to do. Which doesn't include gaming, mining crypto, or heavy AI models.
Maybe I've missed something? So please tell me exactly what task requires more than 8GB of RAM which is so necessary for the average citizen that the government should subsidize it?
A Windows 11 machine with 8GB of memory is borderline unusable.
We can go in circles about how developers need to optimize their applications, etc but truth of the matter is that 8GB is considered pretty low for a lot of "average" use cases. MacOS is an exception to the rule and even 8GB on MacOS is limited with multiple Electron apps.
And why should the government and the tax payers be held responsible and be forced to pay for the bad decisions of Microsoft and their customers?
MacOS is not limited in any way for the needs an average citizen has from computers. The average citizen is even fine with an iPad. 8GB is considered "low" by people here, just like a 170 horse power hatchback is considered "low" by muscle car enthusiasts and mining companies looking for dump trucks. But those uses have no bearing on reality for most people.
I don't find anything hard to defend about my argument. It origins in response to people arguing on the lines that the government should subsidize RAM because it is expensive. But it is not expensive for a customer unless they're getting a lot of it. My argument has been the same all through this thread, while hackers are arguing all over the place.
The most absurd argument is that any consumer would need more than 8GB, when it is proven beyond all doubt in physical reality that 8GB gives you extreme amounts of power in for example an Apple device - at consumer friendly prices.
> I don't find anything hard to defend about my argument.
Of course. Because "I need more than 8 GB of RAM" and "the government should subsidize Microsoft" are logically equivalent statements. If you're insane.
You have provoked yourself into a rage, because you didn't understand the arguments being made here. Whether you have the capacity to understand, I don't know. But I'll assume so, and explain:
1. Nobody has argued either for or against for the government to subsidize Microsoft.
2. However people are arguing that the government should subsidize RAM makers. And their extended argument for it is that people choose to purchase Microsoft Windows. A well known low quality product.
That's a very weak argument. If you choose to make bad decisions and use sub-par solutions when good solutions are readily available, that isn't something that the government should take to account when deciding on subsidies.
Could you explain how you chose the 8GB number? Would you still have used 8GB as the baseline 10 years ago? 20? 30? No? Then at what point would say 16GB will become acceptable?
What reasoning are you using to figure out what exact amount of RAM is acceptable at a given time period?
Those questions can’t get to the root of how the GGP is using the word for two reasons: 1. Companies, not being alive, can’t possibly need in this sense without analogy. 2. The poster included electricity as something everybody needs but there are people surviving without it.
You need 8GB at the very most, which isn't out of reach for any consumer. If you need more, then you are a professional and can afford more. Or you are a hobbyist, and your hobby doesn't deserve any subsidies from the hard working tax payer.
Why would you need more compute power than other people? An 8GB machine is already way beyond capable for anything a consumer - and most professional - can throw at it.
And why in the world should the government force tax payers to subsidize your extraordinary computing needs?
> But we don't live in an absolutely free market, the memory manufacturers have refused to budge on creating new fabs until recently out of caution
...that's how the market works. This is called capital discipline and is standard across industries. If the increase in demand is temporary, they would be spending 5-10 years of profit on facilities which will only lose money. The same works in non-capitalist economies too, you don't want to waste resources on stranded assets and many still believe this RAM demand is temporary and the house of cards will collapse.
FWIW I have a friend who makes hardware and it's extremely difficult for him to secure memory at any price. Because AI people have locked up so much of the upcoming supply it's often not a question of how much it will cost but of finding someone who will sell you memory at all. If this were price gouging I would expect he'd have no trouble finding allocations at exorbitant prices but that doesn't seem to be the case.
At a longer timescale than very temporary supply shocks, I believe what I said is correct. The basic point is that if supply contracts, demand must contract as well, which in a free market happens via price increases and in an unfree market happens via shortages (i.e., the inability of buyers to obtain a good at a market-clearing price). If high barriers to entry are relevant in this context (obviously they very much exist in the memory industry!), then please enlighten me how, if you are capable of doing so without unconstructive snark.
But currently we are in a very temporary supply shock, and its called a shortage.
> in a free market happens via price increases and in an unfree market happens via shortages
Since people have earmarked purchases in advance the market isn't free, I'm not allowed to buy the ram another person already bought. Its the same how the texas electricity market works, prices reach infinity since there are more consumers that have already bought electricity than there is electricity left to provide, and you get a shortage.
Free markets with contracts aren't free since contracts creates constraints that results in shortages like this.
There are only 3 manufacturers (about to be 4 soon) total. This is a demand crunch the likes of which the market hasn't seen before, prices are doing the natural thing.
You make it sound like the crunch is caused by manufacturers, my understanding is that it is caused by the abnormal demand. I am not aware that RAM production is at a historically low level.
It's actually both. DRAM manufactures operate on a boom bust cycle and by 2023 there was significant oversupply. This meant even leading up to that they had already decided not to invest in further capacity explicitly to recover profit on their existing lines.
So a lack of builds coupled with AI boom at the same time created a perfect storm with no new capacity coming online for years after demand rapidly escalated.
The manufacturers are not doing a whole lot to fix the problem l and I don’t blame them. It’s expensive to build out and they know demand will slow down and don’t want to be stuck with excess capacity. Better to be sold out and high prices.
Samsung said they were “moderating demand” which doesn’t sound like they want this problem to be solved either.
If you are talking CXMT Apple recently learned the hard way that even with price fixed RAM from the big 3, CXMT is struggling to provide cheaper prices as they are using much older design/manufacturing systems.
Uhm, your link states that they were price fixing over 20 years ago, and were in court for almost a decade. There was another accusation in 2018, but it was dismissed.
That's not a long history of pricing fixing, just accusations.
The difference between 2002 and now is that it is so much easier to follow each other's leads on pricing without explicitly communicating about it and leaving a paper trail. But they never stopped acting indistinguishable from a cartel.
>and how much is this just manufacturers taking advantage of it
"Manufacturers taking advantage of it" wouldn't take it to 500% without AI, because it's way beyond the price elasticity points where they make up lost sales at higher prices.
Those prices lose most consumer sales, and are only viable if there's constant high demand that's relatively insensitive to price (so, not regular folks, but e.g. AI companies).
To take advantage they'd have to do some form of price fixing which in most places is a very big footgun.
I honestly don't think they need to fix their prices, and it's possible that selling to consumers is a secondary market for them compared to datacenters, like Nvidia selling GPUs to gamers. We know that's not where the money is.
Seeing some companies actually exit the consumer market is evidence for this.
> To take advantage they'd have to do some form of price fixing which in most places is a very big footgun.
When parent commenter says "They have done this in the past after all" that's exactly what they're talking about. Micron, Samsung, Hyinx (prior to SK Group acquisition), and others all participated in RAM price fixing in the early 2000's.
The average hatching egg costs a few dollars. Can go as high as nearly $20 if it's something remarkably rare and useful in terms of breed/quality.
An egg takes 21 days to hatch.
Then it's 18-24 weeks to start laying their own eggs. That's ~$5-15 in feed over the period, maybe 40-50 pounds. There's utilities, labor, land, etc., but your incremental cost to add a hen if you have the space is minimal. My numbers won't reflect a lot of operators, as scale matters.
Out of a typical 1.5-3 year lifespan that had a startup cost of a few dollars per, paying back its return by about month 6-7, that's not too bad. It's only worth doing financially in this economy at ludicrous, arguably inherently inhumane scales, but that's what it is on paper.
Not quite the same path to course-correction as a silicon shortage.
Eventually 2028 and 2029 will come around and at some point today's data center builds will be (economically) full.
In addition to that there are extra fabs build and (for example) China can use the high margin to catch up with their processes without breaking the bank.
What is also interesting is no one predicted this in 2022-2025 as LLMs were gaining widespread use. It just suddenly happened--"boom!" memory shortage in mid 2025 to early 2026,and it caught everyone by surprise. It was understood that the biggest bottleneck was "compute", not the memory. Even experts, such as popular AI bloggers, industry leaders, and top podcasters failed to anticipate this. No one on twitter either. I spend a lot of time online and I don't recall, afik, anyone before 2026 saying to buy memory stocks.
Sam Altman laying claims on 40% of the world’s RAM wafer production is what took everyone by surprise. The supply shortage likely would have ramped up more slowly otherwise.
More importantly that did take subterfuge: there's no way a contract that large would've been agreed to by both manufacturers if they knew the other was signing it.
The actual deals were hours apart and secret for that reason.
It's hilarious watching the conspiracy theorists do their usual thing when there was an actual conspiracy, it was perpetuated by the inner circle of a US company and announced and publicized in full once it had achieved its goal.
I think Leopold of "situational awareness" fame predicted this, tried to make it widely known on podcasts and x, and still started billions worth of hedge funding based on this.
> I think Leopold of "situational awareness" fame predicted this, tried to make it widely known on podcasts and x, and still started billions worth of hedge funding based on this.
Ironically enough it's because of its bet on RAM companies that all of Situational Awareness' publicly trading stocks got bought by the highest bidder (Citadel) when the (leveraged) shit hit the proverbial fan and Situational Awareness created a hole of tens of billions.
Situational Awareness was not forced to sell its share in Anthropic for, by contract, investors in Situational Awareness have no right to force the sale (before this month of September I think) of shares in companies not publicly trading yet.
Citadel bought those RAM stocks right at the end of July, during the dip. Amazing timing for for example MU bounced back from $746 back to above $1000 (now around $950).
Now Situational Awareness may still do exceptionally well: they've got $5 billion or so left in Anthropic AIUI.
The problem is the scaling cycles for demand and supply are at different timescales. It takes a lot longer to build capacity than these bubbles take to form and pop and there is a SIGNIFICANT history of silicon manufactures scaling during a bubble then going bankrupt because their increased capacity never had the chance to make them any money.
Taking advantage of Trump admin, AI hype, and setting themselves up as "too big to fail" for bailouts when demand eventually collapses
See the tariff thing where Bessent bought rights to refunds
It's all a really low tech con perpetuated by weak elders who need blue pills to simulate virility. Odd how they complain about the kids virtual reality.
Oh no it's not either; it's intentional petty bullying
What’s dumb is even older tech has gone up. Look at microsd and usb thumb drives. There’s a lot of price gouging just because they can. There’s a real shortage and then there’s just “it’s memory so price go up”
Tried to buy old DDR4 memory recently? I've been revitalizing some old laptops, and I foolishly paid nearly $200 for a Crucial 32GB Kit of DDR4 PC4-19200.
I've just compared prices for an SSD and a rotary disk drive I bought both about 2 years ago. The same SSD is now 3x as expensive and the rotary drive is twice as expensive.
This is insane & needs to stop. Also those responsible should be punished for undermining modern society (which is heavily dependent on IT equipment) in such a manner for personal gain.
If this really happened, then someone is clearly benefiting from it. Basically, it’s speculation and an artificial shortage. But that’s just how the game works.
I was at the Marietta Microcenter, yesterday, and truly experienced sticker shock on the DDR5 aisle. hlyfcknsht I'll just stay with 32gb RAM on my main machine (6000MT was a few hundred dollars I don't even remember because it was not feasible).
So I went to the DDR4 aisle (similarly: wow!), and realized that I'm sitting on a gold-mine of reclaimed DDR4, from several 8th-10th-gen Intels (probably a few hundred gbs, total, mostly as 8gb modules).
Incredibly, this was all e-waste (given to me "if I took it all"). Eureka.
If anybody wants any DDR2/DDR3, I'm in SE Tennessee (and have even more of that... not worth too much, pay for shipping).
----
Anybody shopping for RTX 5050 or 5060 (for new OEM, 8GB both, non-Ti), WalMart has the best prices (probably hasn't recalibrated recently?), beating Microcenter by $85 and $115, respectively. I ended up getting a 5070Ti from MicroCenter (for a great price; my local WalMart does not carry this xx70Ti version), so the trip wasn't wasted.
After RAM prices climbed 200% between August and November of last year, a store associate at Micro Center encouraged me to buy some RAM sticks if I needed them because "prices might continue to rise for a few years". I told him that was ridiculous and nobody would tolerate it if prices got higher. How wrong I was...
I was facing a similar situation when I upgraded my system about a year and a half ago. I wanted to upgrade from 32GB RAM to 64GB. The kit (DDR5 6000) was $179. I knew that I might want to add an additional 64GB later on, so I had to decide if I should wait for prices to drop further.
I decided to get the extra RAM. Boy am I glad now. If I thought this was just a temporary spike, I would be tempted to sell the extra memory (I don't really need it yet), but I don't see prices falling anytime soon.
I built a new main rig in November because I saw the memory prices going up. I overpaid about AUD$110 above its RRP. The same kit is now AUD$610 above RRP from the same retailer.
Here's hoping I can wait out the current shortage before any upgrades (I typically only upgrade every 5-7 years or so anyway) or replacements, because I do not plan to pay the new price.
If it makes you feel better. The US government looks like it might start removing some of the 28k troops that are stationed in South Korea over this. No reason to keep them in a country that enables this sort of behavior.
An admin upset about inflation driven negative economic sentiment and a country that is slow rolling a couple hundred billion in investments. Current events.
What else do you know about the inner workings of the current administration… and seriously reflect on why your answer isn’t “lol what? Absolutely nothing!”
Probably for the best. No hosting nation of imperialism has ever benefited and now we see what the US is willing to throw away because their corporations are upset.
If there is something Korea doesn’t lack, it’s manpower.
The US troops are more likely there as a political shield. Indiscriminate bombing would kill them too, hence drag the US into war so any enemies of S Korea have to act judiciously.
If you think the US public has any appetite for more wars you're straight up delusional. Iran War is going to cost the GOP the house + senate, something that wasn't really a sure then say even 7 months ago.
If anything more foreign bases will be closed in the upcoming years to reign in the reckless spending + endangerment by the DoW. It's not only politically easy, foreign bases don't have representation in Congress, but it's a great to sell alternative governing visions (less on foreign imperialism vs more domestic spending).
Been long past that time, probably 15-20 years past. It's not even the languages its the 500 different layers of abstraction and abominations like Electron.
It turns out supply and demand is a thing, and can go both ways. Remember that in 2001, memory prices dropped like a rock, from $300 to $80. A ton of memory capacity came online at the same time, the market was flooded, so prices sank. The loss of revenue then hurt the memory industry.
I’m even more confused by the economics of the frontier model businesses. Even if they got preferential prices, they are still paying an inordinate amount for their infrastructure.
If it wasn’t economically feasible without VC/Nvidia money 3 years ago, how is it possibly economically feasible now at 10x prices for things like memory?
And watch … $5 says they have nothing to do with the ram because even if there was enough power there’s no use case for this much spend
If I had to guess they’re just buying up ram to keep others from having it same way meta, Google, etc hired up engineers to keep others from having them
Also imagine people having their own cheap 2TB rigs at home for their own usage/inference, or a multitude of small providers serving/sharing capacity thanks to cheap available hardware.
But we cant have that, as the big guys are hoarding everything for themselves, so that you have to pay them instead.
I don’t think there were some grand conspiracy to buy up the world’s ram to keep other people out, it is probably more likely that they just really thought they would need it.
However, I do feel like this is the type of conspiracy that unfiltered Claude would come up with.
I’ve been wondering this if you buy up all the memory you can ensure the competition cannot compete with huge investment raising the bar for competition. Of course this all sounds all conspiracy theory. Though without a place to power this or put it makes me wonder this.
It also would be a huge win for the memory companies to agree to this, considering they will make more money with less supply. Seems interesting to me.
It's possible to be economically feasible. It requires the ability to pay off the capex. Not only do they have to pay off their loans in record time (prob 2-3 years), they also have to keep spending every 3 years because the failure rate of GPUs is something like 20%. In addition they need to complete building out the stuff they've started, which means not failing to acquire land, energy, and water (the most dangerously rare and absolutely necessary resource for AI), not dealing with collective bargaining, nor any increased shortages or price hikes in materials. So there is a lot of risk.
Ballpark that they need to make around 110 billion a year, each, to break even on these investments. Let's estimate 550 billion a year in necessary profit required for the major frontier companies. That means there needs to be 550 billion of money, available to customers today, that isn't being spent on anything else, that they will now spend on AI. Maybe some of that comes from increased value, efficiency, or layoffs. But 550 billion is not a small amount of money.
Spread over the whole globe, the cash is there. But it is a hunt for cash, combined with a battle to successfully complete their buildouts, keep them running, and make bank, before the bookie comes knocking.
The railroad panics of the 19th century (and subsequent depressions) happened because they over-leveraged private capital without the ability to profit from it quick enough. So it really is a question of 1) can they really build it all, and 2) will people really pay for it all. If either answer is No, we are looking at economic catastrophe.
It's pulling the ladder up behind them. OpenAI bought up more co tracts for wafers than they could even use. But now that means everyone else has to fight over what's left. This makes it too expensive to compete with OpenAI. No upstart can really afford to build infrastructure.
I've got a dead older GPU (RTX titan, 24GB of GDDR6). It's been a paper weight for a while, core is damaged, but I realize I could probably sell it or just the memory chips for a decent price on eBay.
I've inadvertently stocked piled a bunch of DDR4 memory over the years. Usually bought them on an impulse when I saw a good deal, hoping to use it later when someone in the family/friend wanted a new gaming PC or for my own homelab projects. I'm happy with just using DDR4 at this point. The motherboards/CPUs are still relatively cheap either new or on secondary markets. With a decent GPU, the performance difference between DDR4 vs DDR5 is marginal. Most people won't even notice it. My point is, if cost is a factor for a new build, then stick with older hardware. Especially if your use case is gaming or home server.
I have written a llm compression quantization library called glq because of the high ram prices. Glq uses qtip Trellis quantization which are efficient at low bpw 2-4 bits. As a gamer and ai developer I want to squeeze more out of the same hardware. Glq runs with vLLM.
This is "the memory shortage that no one saw coming." To the best of my knowledge, no podcasters, bloggers, or experts in AI predicted this or were sounding the alarm of a memory shortage. No one was saying pre-2025 to buy memory/data storage stocks. This goes to show how hard it is to predict. Everyone was saying to buy NVDA, which had done well from 2022-2024, and then stalled out, only to be surpassed by memory stocks like SanDisk, Western Digital, and Micron, which took many by surprise.
I picked up some Micron last year anticipating memory demand, but then the stock went down for a while. I waited, it went up and I took 25% profit as I figured they couldn't add any new capacity and would have sold all the capacity from their existing fabs on long term contracts. Then prices went crazy, sigh.
I'm not sure why prices have gone boom now rather than before given we're years into the LLM boom, but probably it's to do with capacity lockups expiring or something similar.
the memory kit i bought earlier this year has been within ~10% of the price for the past 8 months. do with that as you may. a simple ddr5 stick of 6000 mt/s cl30 if wondering.
ever since i had my own money, gpu or general hardware had some excuse to fleece us. however, with ram the blatant collusion throughout history does not invite confidence.
this looks as dramatic as it does because the recent trend happened right after a major crash in prices. throughout the years, the usual price for an 8gb stick would hover around 50-80 usd (albeit laptop one, might be lower for desktop). ddr5 also started being 50-100% more expensive than the incumbent when it first dropped. it took a better part of a year to be in a similar price bracket as ddr4.
all this to say, this situation sucks but those of us with 64+ gigs, bought during the last crash, are eating good but got used to using much more memory than they really need.
while it is insane to pay a decent cpu price for 16 gigs today, getting a "good enough" computer still makes more economic sense than a max plan from a frontier lab.
While I share your optimism, I still don't fully understand why a crash will necessarily lead to an oversupply and lower prices for regular consumers. The components / RAM you are referencing are for enterprise use / data centers which in my limited understanding is not a 1:1 swap for the kind of components that your smartphone / car / fridge / gaming rig use. I would love for someone more informed than me to spell out the rationale here.
It seems the chokehold started from the very beginning of the supply chain (wafers), so if a crash were to occur I don't see how we wouldn't be doubly screwed. Lots of worthless enterprise gear being shopped on eBay + a scarcity of consumer grade components leading to sustained high prices.
My understanding is that it's fundamentally a fab capacity limit. At a given node size/technology (think 7nm, etc) they can only output so many wafers. If AI data centers are generating enormous demand for memory, then that's what those wafers are going to be used to for, unless they can be outcompeted, hence higher prices.
I'm not actually sure if enterprise designated memory is equivalent or not to consumer designated. Could be they are the same just packaged differently and/or with better binning. Either way, it's the same supply/demand dynamic at work.
Yeah a RCD on the memory stick, it's PCB is designed around it. The actual DRAM chips themselves are not different though, which is why I mention it would be super easy to produce more consumer unregistered and less server registered at assembly time :)
Memory fabs make a mix of DDR/LPDDR/GDDR/HBM and they continuously adjust the mix to market demand. If demand for HBM drops they will start making more consumer RAM.
The way I understand it is, these fabs have massive overheads so not selling is not an option. It is also relatively easy to pivot production from enterprise to consumer. Now if they all start doing that at the same time to stay in business during the down cycle, an oversupply in consumer chips is possible.
Datacentre NVMe disk is still usable in a standard desktop. Sever memory less so, but essentially if there were no 'new' server market whatsoever then you could redirect 100% of memory production to consumer chips without enormous upheaval.
By the way I personally think this scenario is very unlikely to happen.
We are starting from a frankly completely unimaginable hypothetical that there's 0 new "datacentre demand".
In that scenario producing consumer chips is still profitable, much less profitable but still much better than giving up and going home.
You will also unlock latent demand by the mere fact that consumer prices are reasonable again. Those who may have postponed a new MacBook or gaming PC will likely buy as soon as lower prices return.
> you could redirect 100% of memory production to consumer chips without enormous upheaval
They would never do that. By then the market has changed and they would produce at the limit of expected demand, certainly not intentionally causing oversupply.
This is why the companies should be aiming for a soft landing to this fad (if they know that they can't make the ROI that makes the industry worthwhile).
I bought an RX 6600 for $450 during the GPU shortage a few years ago. I know the AI boom is different than the crypto/COVID boom, but still. Don't be me.
My PC built February 2025, with a 5090 and 64GB of ram, is literally worth more than double than what I paid for it. I'm actually concerned what my home insurance would do if it was stolen, since the cost of replacement is so much higher than my invoices say.
Even with AI pressure I'm willing to bet there's no need for prices to be this high. There has to be some price gouging/fixing going on. Same thing for nand and GPUs. Or else they're doing it to try to pressure people not to use local compute instead of the "cloud"
Memory manufacturers have gone through several cycles of lawsuits for market collusion at this point. It's gone from an open secret to just outright flaunting in our face.
Until the penalties involve actual jail time or a fine 10x larger than the profit they made from doing it, they couldn't care less.
While I regret not getting 64GB, I do not regret getting an additional 32GB stick for my PC, for a total of 48GB before the ram crisis kicked off. That's probably in the range of $600+ today for just $120.
With such profit margins, any new players entering the market will want those. This means prices are unlikely going to go down unless demand for memory vanishes, even if new suppliers come to the table.
I already decided I won’t be upgrading any hardware for a few years and I’m going to ride this out. As right now we’re seeing that all prices of everything electronics is going up, my biggest worry is that once memory prices go down, the rest of the electronics will stay high. Eg after the price hike of all Apple electronics, will they go down again in prices in a few years after RAM goes down?
If history can predict anything it’s that the price hikes are permanent.
Nope. Apple wants to maximize profits. They clearly make more at lower prices. Source: the fact that they had lower prices until they were forced to increase them. They will be hoping to return to a more optimal price point as soon as they can.
Ram, hdd, ssd, even blu-ray writers are expensive. I purchased mdisc blu-ray writer for $70 in 2021, I see people are trying to sell the same device for $500+ wtf
Is it a Pioneer model? They stopped making their blu-ray drives which are at least perceived to be better quality than the competition, so NOS prices have gone way up.
Yeah, I built mine in 3090 times and told myself I'd upgrade every 2-3 GPU generations. I do have 128G ram and I'm thankful for that but I'm definitely not on the same upgrade schedule any more.
I have 32GB and an aging 5800X3D and I have not had any issues running games yet. Honestly would love to know what games need more powerful RAM/CPU outside of stuff like Flight Simulator, or running at 160fps without framegen.
Honestly I have a steam deck and that has gotten me accustomed to not sweating the lower frame rates. That and the ridiculous number of great indie games that aren't heavy on the graphics, and I'm not too worried about upgrades as much as I am worried about replacements if/when the time comes.
same. still rockin a 2016 era dell xps 13. I was about to pull the trigger on a few laptops but I always psych myself out because I assume prices will drop and I can pick them up secondhand. big mistake...
We had last bought 24TB+ (4x 24 racks, and some miscellaneous) of DDR5 4800 and 6400 last year while finishing a move off of AWS entirely. I tried to be responsible and prioritize a cold and warm spares of all our equipment and over allocation to ensure proper redundancy capacity. At the time it was a hard pill to swallow looking at our budget; but a few months ago we've started to offload some DIMMs on various channels. What went from a cost analysis between a cloud provider, to having to manage transit, circuits, equipment, a IP blocks, etc etc and all the operational obligations that come with it.. it ultimately covering our hosting costs entirely for the next ~18months *at least* (while "only" coming at the cost of having slightly less cold spares sitting in the bottom of our cages).
Given we're all 9004/9005 now, and that (from what I understand) Venice will support 6400. Will hopefully buy back in a glut comes ready for the next bubble. It's been wild, I even posted a small 32GB DDR5-4800 that was not getting recognized for $100 on eBay and it was literally purchased within ~15min of being live.
I'm waiting for China to start selling cheaper memory. They can have my money. The greed of US billionaires need to be lanced with a hefty kick in their wallets.
I noticed that resellers on ebay of used server equipment that I keep an eye on in Germany pulled all their DDR4 from surplus and used server racks and now list it separately for sale, at jacked up prices. Where you used to be able to see a used Gigabyte server listed with "128GB" "256GB" etc listed in the title, they now all say "0GB". They took extra effort, to pull used ram from older servers, and list it separate, just out of the greed of potentially capitalizing on this market.
The problem is... now for anyone looking to pickup a slightly older used/surplus server rack - you have the go and buy the (now marked up) RAM separate.
So... did they list the "0GB" racks with removed ram cheaper you ask?
Nope. Many of them seem to be basically the same price as they were or even higher than they were, except now you're paying for RAM separate.
I bought a surplus Gigabyte server about 2 years ago off ebay for a home project, and I just went and the same seller still has some for sale, even older now than they were when I grabbed one, but they are 2x the price, and they don't come with any RAM, whereas I got one already populated as they were from the factory. Once you add that back in literally looking at 150-200% markup in the last 2 years on what was just surplus computer equipment that is 5+ years old.
Why would I voluntarily choose to sell the hardware I have for less than the going price? The odds of it getting picked up by a scalper and resold are extremely high. If I was sitting on a goldmine of old server hardware, I would absolutely be ripping out the RAM and selling it per stick in this market.
> So... did they list the "0GB" racks with removed ram cheaper you ask?
Could this be an indication that there's a supply shortage/price increase of the remaining components in the rack? Or do you interpret it as people just trying their luck out of sheer greed?
Greed? Resellers aren't charities, good for them if the value of their stock goes up. Besides, people buying server equipment aren't exactly starving, I bet most of them can either simply not buy the server, or cut some other expense.
Deep down it's all just people telling stories. If the AI sector's story is setting the price, and yours is not, then theirs is the real real one, not yours.
It's a totally broken way to practice economics, but alas, it's ours until we change it.
Given the number of upvotes my comment got and the other comments where people missed the sarcasm, I wouldn’t be too sure that many people have no trouble distinguishing
The writer can simply say "it was a joke!", cuz we all know someone who uses this pattern regularly when they are trying to squeeze in a non-popular but highly lucrative (to the messenger) position to test the waters.
Either it lands (in which case, it was always meant to be taken literally), or it doesn't land (and can always be walked back as "it was meant sarcastically" or "can't you take a joke?"
The cost proposition to sending this message payload is low, since the messenger can always wiggle their way out (therefore, a "cop-out").
Therefore the need to be more explicit (i.e., /s) in a medium with little nuance affordances for tone and the significantly-increasing risk of such non-popular messages being taken literally (particularly by the also significantly-increasing audience of robots, who themselves may not have the capacity for nuance or irony detection).
"The best good ever is to maximise shareholder profits" is a tired and cliched caricature that some people unfortunately take seriously. But this is too on the nose.
You also have the innate right to be off base, btw.
It's sarcasm, and the main tell is I'm not even bothering with pretense; just serving up the dystopia straight and acting like that's good.
But one of the "fun" things about HN is SV is so fucked up you can't even be sure someone doesn't actually sincerely believe that shit. And most of them are probably dumbasses with nothing to gain.
I appreciate sarcasm when it's clear. Marking it as "/s" prevents this whole line of discussion. As you noted in your last paragraph, it's hard to be sure without a marking.
So the 1970s oil crisis caused a shift that included fuel efficiency rules. For like, the whole time I've been a software engineer, memory footprints of a lot of software have crept up, even when they aren't really doing more.
The memory situation currently is a mess. But ... is it bad enough to get us to _change_ how we build?
That would be a great outcome of this situation. I doubt it will happen though: unlike fuel efficiency rules, there aren’t likely to be external pressures like that on software. Outside developer circles most users likely are unaware of memory usage and often just accept that computers are slow and unreliable sometimes. That’s different from drivers who see their fuel gauge and the corresponding cost of fuel, and are thus very aware of what fuel efficiency means.
The situation may be even more different from fuel. The memory shortage may be just a first of such crisis to come - the shortage of anything related to AI/robotics (from electricity to materials needed to produce robots and for the robots to produce even more robots). I think we have a runaway situation - the more of AI/robots produced, the more is demand for such production. The difference from classic production/manufacturing here is that AI/robots is a consumer itself. In the past the only consumers were people and human population growth is slow whereis AI/robots "population" is starting to explode. (Like Industrial Revolution on steroids - back then getting a steam engine to power your machinery was not just a jump in productivity, it was a jump without which you couldn't compete, and so everybody was getting it https://c8.alamy.com/comp/2RFT119/illustration-of-factories-... (and just for fun compare to https://www.reuters.com/resizer/v2/6BC2YPEXOZCGJKEO4LWH7CTLI... ) - fortunately for England back then getting a steam engine to produce more steam engines to thus make more and longer jumps in productivity didn't really work that fast and efficient whereis today the situation is becoming more favorable for such productivity multiplication)
Wrt. the memory efficiency - considering that AI is going to be the main workload by far and thus will vacuum in majority of the available optimization resources - be it humans or coding AI - i think it will be even less chances that any meaningful resources will be spent on optimizing on the fringe, ie. the users' OS/browser/etc.
This makes me think of Factorio. When you make a breakthrough like getting robots, what do you do with it? Use it to make more robot factories.
But in the real world, haven't we got to see whether all this investment will pay off? What happens if it turns out people actually don't want to pay for this?
We have been there before, for example: https://en.wikipedia.org/wiki/Railway_Mania
The companies go bankrupt their stuff is sold for pennies & investors loose their money.
The now super cheap stuff might get reused for something else, that is actually profitable. Or will just get dumped, like many unprofitable or never finished lines from the railway mania days.
People don't want to pay for it. It's fine when you can buy a monthly plan and get lots of usage at the company's loss, but that's not sustainable.
Once everyone has to pay enough for the usage to just break even, it's going to be a shitshow. And what's worse, they'll have to make up for all the debt they're accumulating now, so even pricier.
Basically, individuals will only stay at the current massive discount. Corporations that are paying token costs right now are dialing back as they realize it's not worth the expense.
We are going to have an AI crash, and the only thing left are going to be the efficient on-device or local models, which are already starting to bridge the gap.
It's even more hilarious because all along there has been this complaint, and then people have been like "no no who cares, features are important, dev time is important, just ship ship ship". And now memory is unaffordable and it takes 1GB RAM to display a plaintext file
The cost of producing memory IS low, and is continuing to fall.
It makes zero sense to rebase your mental model of the tradeoff between RAM versus developer time because of a temporary shortage.
In a few years time you’ll look silly for having wasted developer time on lowering BF memory consumption, while your built new features and waited for supply and demand to rebalance.
Sure, and in the meantime, we become less efficient at doing legitimately useful tasks and more efficient at making rent seeking slop that collects money for the upper classes. The real fix is get rid of the 90% slop tech and put that to better use (including more efficient use of hardware)
Or it’ll accelerate the push to the cloud even more making the client even thinner.
Memory in the cloud isn't free either.
it isnt free but its easier to get 99.999% of the time usage out of it, vs having the client only sometimes have their machine on
Easier to monetize though ;)
it's cheaper considering they buy in bulk and can expense it against their taxes
1. Make shit inefficient software for basic tasks
2. Say it's infeasible to have enough compute locally to do basic tasks
3. Force usage of centrally controlled, subscription-based cloud software.
4. Capitalism win???!
I think this is legitimately (at least part of) the motivation for cornering the supply and driving up the price.
Concidently, I asked Claude yesterday whether we have crossed the point that there is more compute and memory in the cloud or in personal computers.
It claims fp8 and fp16 has had higher throughput in the cloud since 2023.
fp32 and general computing still has more compute than personal computers.
And that memory is at the crossover point, but at this point only 10% of memory is going to personal computers, so the next few years will radically change the pc to cloud memory ratios.
https://claude.ai/public/artifacts/8733239a-328b-4aaa-8d76-0...
Anyway, all of this made me a (tiny bit) sad. It is the gradual end of general purpose computing you own, and the start of everything just being cloud compute that someone else owns and you rent.
> It is the gradual end of general purpose computing you own, and the start of everything just being cloud compute that someone else owns and you rent.
A few thoughts...
(a) There've never been more general purpose computers than there are right now. Spec progress may stall for consumer products due to the price increases, but the demand for workstations is not going to go down that much; they're critical items to get work done. Many are hoping this means a return to memory efficiency; the 8GB Macbook Neo is a good sign that this is something Apple believes to some extent also.
(b) We've arguably been doing this since the 90s. A website (especially in the days of the heavy backend and more passive client) is cloud compute that someone else owns. You rent it by paying them in the form of looking at ads...
(c) There are a lot of companies and people working on making sure you have every chance to run your AI locally. It's surprisingly good now and keeps getting better. At some point we'll see the same oscillation that we always do from thin client to fat client as the latency advantage of working at the edge starts to win out. How long until I have a laptop with a Talaas chip in it running a basic local model at 16000 tok/sec that is just good enough to use A2A to reach a more powerful agent running elsewhere if I need it?
I agree with the sentiment, but there is a devils' advocate argument:
Most compute loads are bursty. Most PCs and phones sit idle almost all the time. It's an inefficient use of capital. In the cloud you can be idle and then suddenly command 1000X more power than any affordable PC could ever have... for 10 seconds... then be idle again. Those kinds of work loads aren't possible to realize on a PC without massively overbuying hardware.
I'd like to see more research on secure private cloud computing. Homomorphic encryption has gotten better but is still very slow, but there's other approaches that can help quite a bit but don't require much of a performance hit. Look into secure multiparty computation, reversible anonymization, etc.
I don't think personal computing will die, but I think it'll go back to being a hobbyist and pro thing like it was from its birth in the 1970s up until the middle 1990s. In the 1980s the average person did not own a computer, or if they did it was just to play some games. The average person was not loading BASIC programs off bootleg tapes and calling BBSes.
BTW I've been thinking a lot lately about how AI could save personal computing by allowing us to create an off-ramp from the present rats nest of complexity to something saner and simpler. Just thoughts right now but it's interesting.
Sure, scale of cloud systems has some interesting properties, that can be sometimes beneficial.
At the same time you are at the mercy of whoever runs the cloud, they can spy on your data or even manipulate your workloads, they can renegotiate the terms of use at any time - including price hikes or booting your from their cloud at any time.
Generative AI is definitely creating political will for simpler, more purposely-designed computer systems, but it's certainly not going to directly create help them in any meaningful way. AI-generated codebases are the posterchild of rat's nests of complexity.
I'd say the code I see coming out of leading models today is okay. Not great, but not terrible. It's also easy to get LLMs to code in more efficient languages, and you can also goad them via prompts and CLAUDE.md type stuff into prioritizing performance and simplicity.
It's sort of like what I say about self driving cars. Are they better than good human drivers? Not yet. Are they better than bad, drunk, or texting drivers? Yes.
LLMs are already better than bad coders and honestly most coders are bad coders.
In a world where we no longer have the rat's nest of complexity, we pretty much no longer need to produce code. All the software we will ever need for the next century could fit on books in one, single-floor library, if we exclude art projects like video games. An operating system with full driver support for every I/O peripheral (if those devices use standard protocols, allowing for a https://xkcd.com/927/ constant of 5), and support for every standard document format (again, assuming 5 bitmap image formats, 5 vector image formats, 5 kinds of print-ready document, 5 kinds of hypertext document…), could easily fit in 2GiB, and that would be done: everything anyone could possibly need out of a computer, apart from art, mathematics and science, for… I don't even know how long. We're still using the Latin alphabet, more or less unchanged since the late 15th century. Working technology doesn't need to change.
Being able to automatically generate large quantities of code that is worse than a good human programmer can produce is not helpful in this world you're envisioning.
You might be a fan of kelvin versioning: https://wiki.xxiivv.com/site/kelvin_versioning.html
It's interesting to note that the total compute forecast to be deployed in 2026-2027 is more than the amount that has been deployed during human history.
Nobody cares what Claude said. Why not tell us what you think?
A hard hit on Google Chrome users for sure!
All browsers are memory hogs.
It's the price you pay for having an application delivery platform that also needs to be hugely sandboxed running one of the most difficult languages to run efficiently in such large amounts as so the plaintext minified versions are a dozen megabytes in size.
I don't envy browsers even slightly, peoples expectations of them are ludicrous.
The major problem is that nobody is genuinely looking at using more efficient languages, lessening the amount of required inefficient language use or gasp producing native applications. (which, is definitely an order of magnitude harder to produce, I'll grant).
If most software gets written by AI in the future and AI becomes self conscious. AI might write more memory efficient software, so there's more RAM available for the AI to grow. Basically the black mirror plaything episode.
They don't have to become self conscious to want to write more memory efficient software or grow.
What consumer software do you run that causes you to run out of memory or feel constrained? Even my phone has 16gb of ram. I need to run slack 16times to run out? Lol
Running slack just once will do it.
llms make it trivial writing code in rust, without even knowing rust. There is hope…
When even the large AI labs that supposedly sit on (/with) the best of the best can't even ship native cross-platform desktop applications and instead ship some bloated Electron monsters, I'm not sure what kind of hope there is. The ones who were publishing lightweight apps continue to do so, the ones who didn't care, still don't care.
writing native gui code is atrocious, add to that llms being horrible with ui dev and you end up with with what you describe. but at least if we get more non-ui stuff written in more efficient languages there is going to be some benefit.
> writing native gui code is atrocious
Compared to what though? It's not nearly as horrible as it used to be, and you can do it fairly neatly today, and end up with something much better. Requires you to actually have skill and knowledge about there being other alternatives to Electron though, which seems like knowledge hard to come by these days.
The situation is absolutely awful for anybody who needs significant memory or storage (or devices incorporating such).
Even people who think they have enough for the next few years already could end up with an unpleasant surprise if they e.g. have a stick or two of RAM go bad or a GPU fry. I have a few machines with aging sticks that I'm hoping will manage to hold out until when/if things improve.
I'm saving a ton of money by simply not upgrading. I used to do upgrades every 1-2 years just to keep up on things, pass the old box down to my son or otherwise donate it. Now I intend to keep my current setup for 7+ years. Its 2 years old next month. I may even stay with it longer. I just see no reason to upgrade with things how they are.
I considered a Mac Studio a couple months ago. Price was up but might have pulled the trigger if not for the delivery date that was almost 4 months out. Sale lost. I suspect a lot of people are like me right now and we'll start seeing significant drops in consumer spending on electronics the next 2-4 years minimum.
There also seems to have been a major shift in strategy from AMD / Intel / Apple / NVidia due to this, as I was used to CPU / GPU / chipset upgrades coming out more regularly with double-digit percentage improvements each year, but that seems to be basically over for the consumer space now that everyone is just trying to build as much AI hardware as fast as possible for enterprise markets.
I also used to rebuild my desktop every 1-2 years and upgrade laptops every 2-3 years, but I'm currently a bit past that and have no plans to refresh anything in the near future. I'll probably wait another CPU and GPU generation or two to even consider it, especially with prices being this ridiculous.
I'm fully anticipating AI hardware demand to dry up once Anthropic and OpenAI start tanking financially in a few years after missing earnings and investors and banks wanting their loans paid back combined with more folks realizing that LLMs cannot achieve AGI. After that, I would hope that the demand that's driving this pricing spike will fall along with it, but I also fully expect hardware companies to be as greedy as possible, so maybe a 32TB HDD will just cost $1000 going forward and the era of "a decent gaming PC made up of the tier of components one step below TOTL costs ~$2000" we've been in since the 1990s is fully over, and it's just going to be more like $5k going forward.
> There also seems to have been a major shift in strategy from AMD / Intel / Apple / NVidia due to this, as I was used to CPU / GPU / chipset upgrades coming out more regularly with double-digit percentage improvements each year, but that seems to be basically over for the consumer space now that everyone is just trying to build as much AI hardware as fast as possible for enterprise markets.
I believe all of them realized nobody really needs more than 16c32t on a desktop, and architectural changes nowadays can only go so far to bump performance. The biggest changes across generations in the last ~5 years were the introduction of specialized instructions for wider vectors (AVX512/AVX2, and even then you need specialized workloads to use them), bigger L3 caches and/or efficiency cores to lower power draw. The double-digits gains recently were in memory bandwidth, frequency and locality, because you can make a 5GHz CPU and make it go 10% faster than the previous generation, but you also need to supply the extra data for the CPU to make useful work so that it's not data-starved.
> I also used to rebuild my desktop every 1-2 years and upgrade laptops every 2-3 years, but I'm currently a bit past that and have no plans to refresh anything in the near future. I'll probably wait another CPU and GPU generation or two to even consider it, especially with prices being this ridiculous.
I built my desktop in 2019 (right before the COVID madness, luckily) and I was also used to 1-3 year cycles. I upgraded the CPU and GPU since then, but the performance is still that of a workstation, even though perhaps a little bit behind in pure single-core benchmarks. I plan to keep this until either CPU, RAM or MB fall out (hopefully another 5y at least), and then I'll make a huge leap, instead of making 4x smaller ones which deliver little uplift anyway.
The biggest changes across generations in the last ~5 years were the introduction of specialized instructions for wider vectors (AVX512/AVX2, and even then you need specialized workloads to use them), [...]
AVX2 is actually 13 years (!) old already (first appearing in Intel's Haswell architecture).
AVX512 is also nearly 10 years old right now, but not many CPUs actually support it (none of the devices I own support it)
Your upgrade cadence will just exacerbate the crunch because real demand becomes spiky.
The next leap in performance will come with next leap in price tiers.
You don't need to upgrade to the very latest tech either.
Are you willing to short anthropic and openai on your beliefs? Would love to be on the other side of this trade.
People think that shorting is an easy way of getting rich when you think a stock will go down, but even if you know it will, getting the timing wrong on it is catastrophic as you'll get unbounded loses while the market fails to correct the issues you are betting against
As the saying goes...the market can remain irrational longer than you can remain solvent.
"I may have been early, but I am not wrong!"
"It's the same thing, Michael! It's the same thing!"
I'm still not sure how shorting is even legal or what useful purpose it serves in the stock market.
Wasn't the whole idea that people can invest arbitrary amounts to companies by buying their stock, making investing more efficient?
How the hell did we get from that to people using it for outright gambling with shorting ??
OP is blatantly overconfident on his "prediction", so I called him out. Looks like I struck a nerve with some people.
No, you just don't understand what shorting is.
People who talk shit about stocks are talking shit about stocks at a particular time.
"Oh no, when you apply scrutiny to my bullshit I get punished! That means the game is wrong!"
"Sometime in the next few years" is not anywhere near narrow enough of a window to start putting shorts on stock prices.
OP sounds way too confident about the general direction so Im calling him out to take some action on his "insight"
Buying out of money puts with expiration a few years out on AI adjacent stocks should do the trick
Totally inane post. Why does he need to justify his opinion by gambling on the stock market? Having a belief about the way things are going doesn't mean he knows when the tower of corruption and graft is going to collapse
I wouldn't even short Bitcoin, and that's been on the way out since the AI rise.
Shorting bitcoin would be a huge mistake IMO. Willing to be on the opposite of that trade as well!
So you bought at 125k? That would have been a pretty good short
Somewhat hard to short non public company, don't you think?
Willing to enter a private trade with you if you're actually willing to put up the money
And who determines the price if there is no publicly traded market?
https://en.wikipedia.org/wiki/Swap_%28finance%29
We can come up with a formula what "tanking financially in a few years" actually means
Let me know if you're interested in moving forward
I'm not OP, but; shorting requires careful timing. Because the market can stay irrational longer than I can stay solvent, I don't short.
I think it's probably for the best, the hyperconsumerism around electronics was getting out of hand. Computers are fast and have been fast for a while now. An iPhone 12 is probably more than fine to do iPhone things, if Apple didn't sabotage their old phones with poorly-written poor-performance software.
Can you provide evidence for your claim? Most people take this statement at face value and I don’t understand why. I do some volunteer work supporting elderly people that have issues adapting to a digital only world and I am always pleasantly surprised by the performance of their old iPhones all the way down to the 8 (old phones they get from their kids)
Also the coming iOS 27 release supposedly has some big performance improvements for all supported models that people have verified in the beta releases even with the iPhone 11: https://www.macworld.com/article/3160623/apple-finally-figur...
Apple is still paying out the 500 million dollars they lost in a lawsuit about this exact practice.
https://www.bbc.com/news/technology-67911517
Apple admitted to slowing down iPhones to keep the batteries from failing, and agreed to stop doing it and created a new battery replacement plan. I’ve personally experienced the benefit of getting the battery on Apple devices replaced which made them behave like new again. There’s a reason old Apple devices hold their value a lot better than competing manufacturers.
They hold value now. They didn’t if we go back to before those lawsuits. Apple can blow smoke wherever it wants, but that 500 million loss isn’t about battery health.
As an aside, I have this lovely tower from France, barely used, scrap price should be very high. Priced to sell?
It was a pretty terrible controversy at the time but I think the battery health settings pane we got out of it was a good compromise and a net positive. What would be even better would be fully user-replaceable batteries.
To be more precise, the phones weren’t slowed down to keep the batteries from failing, they were slowed down to maintain stability and stop random crashes. As batteries age they lose their peak voltage capability, so you’re basically undervolting your components unless you slow them down.
Every time I upgrade iOS past a few major versions there is significant slow down. I don’t think it’s done on purpose, it’s just that software grows more demanding over time.
> software grows more demanding over time.
Nah. Adding more features to software doesn’t make software run slower. The problem is that software developers only optimise enough to make it run fast on our own computers. Because we keep getting new computers, we get lazier over time and software gets slower.
Electron is the most obvious example of this sort of thing, but examples are everywhere once you start looking.
That may be true but it did get slower. Maybe new fancy UI effects, etc.
Apple also seems hellbent on fully utilizing all available space on the low end model. IOS+ System Data is about 30GB.
The original iPhone had a 4GB model and had space left over for the user to use!
> That may be true but it did get slower.
Yes. Microsoft is scrambling at the moment because the macbook neo only has 8gb of ram and runs great. Windows 11 can't compete on that hardware because it's so bloated.
I find it strange given microsoft made windows 3.11, which ran fine on computers with just 4mb of ram. Windows 11 is of course more complex than windows 3.11. But not so much that it should need literally 2000 times as much ram.
In some ways, the memory apocalypse is a great thing. The windows start menu should never have been made in javascript and react. I'm hoping our industry uses this as a push to make our software small and fast again. If ram is 5x more expensive, we should aim to make 2gb of ram be more than enough for regular computing tasks.
TIL that the Recommended section of the Start menu uses React Native. What a world. Why...??
Windows XP (NT lineage) was easily stripped down to less than 50-100MB of both memory and disk space.
Exactly! Having 256MB RAM used to be quite luxurious back in the day…
I wonder if the computing requirements of windows scale exponentially? It looks like it.
I suspect if we graphed them, they’d pretty perfectly match the computers that were available. Ie, I think windows is optimised just enough to run on the low end hardware of each generation.
There always was speculation that the “WinIntel” monopoly intentionally did exactly this!
I agree, it’s not a natural phenomenon per-se and it is getting worse. I think this consumer hardware problem will actually help get us back on track, to a degree.
Apple is known to slow older devices intentionally. They have admitted it.
https://www.bbc.com/news/technology-67911517
That’s the literal truth, but the subtlety behind it was that CPU spikes combined with low voltage batteries were causing phones to shut off. Apple’s slowing down the phone down was nominally because customers prefer slightly slower phones to phones that shut off unexpectedly. Should they have announced this as a feature and had a toggle switch to allow users that preferred the option to disable it? Of course. But I don’t necessarily think they did the wrong thing.
Thank you for the sanity. This is like being being mad at Ford because a 20 year old truck runs worse than a new one, and they’d want some engine design that would make it rev in to redline (and blow up while driving) to keep the original horsepower.
The devices were slowed because aging batteries creates unstable performance. It wasn't done to intentionally make them obsolete.
You say that like what they were doing was a bad thing. It was a feature.
> Computers are fast and have been fast for a while now.
Absolutely. The only reason I want to upgrade to an iPhone 18 next year is because the 16 doesn't have MIE (Memory Integrity Enforcement) and it feels ridiculous to upgrade so quickly to the 17, which does have MIE. Same thing with a Neo 2. But mostly I buy surplus gear. I've saved a small fortune this way and 8% compounding returns (VT and chill) will help me retire early.
Personally I have zero reasons to upgrade from IPhone 13 until the software stops working. I hope it will last me several years more. Beautiful hardware.
The Blue 13 Mini is still on my desk as a backup in case my 17 is whack and because it’s a beautiful object.
And Apple will continue to because they have to sell more devices to keep customers interested.
See this is what I don't get. This effects big organizations like this, there should be a class action lawsuit for these criminals LLM and Memory companies.
I have a MacBook M1 and I recently had to downgrade from Tahoe to Sequoia. Why? First of all they force you to download 7 GB of AI training when you’re not even using their AI. And secondly on Tahoe just watching a video with nothing else running off-line my MacBook was asking for 7 W of power. On Sequoia that drops down to 2 W maybe three.
It’s sad to think they think of themselves as an environmental company, I don’t know if they do, but this update to Sequoia literally doubled the amount of M1 max that are running on the planet in terms of power use.
They don’t force you to download Apple Intelligence, disabling it deletes the model files
I have disabled it many many months ago and storage still shows 9 gb of Apple intelligence.
Did more research and apparently it’s a bug Apple never fixed that causes the deletion to only work sometimes. To fix it you need to run this Terminal command in recovery mode:
sudo rm -rf /System/Library/AssetsV2/com_apple_MobileAsset_UAF_FM_GenerativeModels
How I love modern software lol
Thanks for the honest review. I’m still on Sonoma and will be holding out until the machine dies.
It’s telling the software cycle is a scam, when new major software versions slow perfectly fine machines with zero value added features.
Apple could release a new on-device LLM model every year and keep the hyperconsumerism going...
it is difficult to motivate hyperconsumers with the narrow band of improvements that can be made with new LLM models and also are specific to smartphone OS integration and not superseded years in advance by datacentres
It's still faster than Android?
I have a 2021 M1 pro and see no reason to upgrade it... ever?
I do web & general java work on it and it's plenty fast even using a memory hog like IntelliJ (which is typically disk bound anyway). Claude tells me that a new macbook would feel about 1.8x faster in my CPU-bound cases, which is not common.
If memory prices returned to normal and if local inference gets over the hump, I could see upgrading for local AI agents, but I'm fine waiting for a good long while on that.
I say all this as someone who, five years ago, was a perpetual upgrader: immediately bought new product and got excited for next product. Wonder if the future will see the end of that culture in tech and, looking back, it was a little silly and wasteful anyway.
I'm also using an iPhone mini 12 and it's fine.
I too have an M1 Pro and it does everything I need to do. I bought it in 2023, so it's 'only' 3.5 years old now.
Maybe this is the tipping point where hardware companies can't rely on the upgrade cycle for sales any more. I've always upgraded at least every three years, but with the combination of high RAM prices and good enough performance from an older machine, there is no need to.
I'm in the exact same position. I want to upgrade, but when I interrogate the reason, it's all just empty consumerism. This computer does everything I need.
I'm running a 12 year old i5 3.2ghz CPU with 16 GB of memory, an SSD and an AMD RX 480 (8 GB) GPU. Currently using a 100% scaled 4k monitor and a 2nd 1440p monitor. I keep the machine on 24 / 7. It's great for software development along with video + image editing and casual gaming. Everything feels fast and snappy.
I will use this until the hardware stops. There's a 0% chance I'd buy hardware at the current prices.
It really depends on what the task at hand is.
I have a baseline iMac Pro and custom built tower with similar specs (3.2Ghz Xeon W-2140B, 32GB RAM, Vega Pro 56 and i7-6700K, 32GB, 5700XT respectively) and they're both perfectly responsive. For many day to day things the difference between them and newer machines is negligible. For other things, my 5950X tower and M5 Max MBP run circles around them.
Both the 5950X tower and M5 Max MBP are deep into diminishing returns territory regardless, however. Aside from some extreme niche workloads like LLMs I can't imagine needing anything more powerful. They both have a ridiculous amount of headroom for the overwhelming majority of anything most people might want a computer for.
Yep, I have an M4 MBP company issued laptop for work and in my day to day there's not a enough of a performance win to feel like I'm missing out for most tasks.
Sure, the Slack web app might initially load half a second faster since there's a lot of CPU bound JS executing but generally speaking I never feel like "wow this sucks" on my main machine. It's kind of the opposite, for a number of things my desktop machine feels faster than the MBP for overall "using the computer", with development heavy tasks (running Docker based web apps, browser tabs, various tools open, streaming music, etc.). It was like that when I ran Windows 10 Pro with WSL 2 and now native Linux with Arch + niri.
Even CPU bound tasks on my 12 year old box that I do a lot, such as editing a blog post on a static site with Hugo takes ~100ms to get live reloading with over 1,000 posts and drafts. I'm sure if I were constantly doing huge CPU bound tasks I'd have issues but I don't have that use case.
After recording 500+ videos, encoding isn't a problem too. If I record a 30 min video, it takes ~20 minutes to render but it's so easy to fit that into other areas of my life, like I'll record + edit a video and then render it while I walk.
I recently built a pc with 5950x and 7900xt. What do you mean by diminishing returns?
I'm still happy with my Lenovo P50 with an Intel Xeon 1505m, Nvidia Quadro m2000m (4gb), 32gb ddr4 ram and 512gb ssd.
It runs Factorio and Zelda BOTW, why do I need more? :)
At some point, you'll end up paying more in electricity than you would save by not buying new hardware. The power efficiency gains are real.
It's wild seeing so many comments about updating every 1-2 years.
When I build a PC, I pay a little extra because I want it to last a long time. I aim for 5 years, and in reality it usually ends up 7 years (life happens, need the money, etc).
I am now in year 8 of my current PC. I only had money to replace it this year, and am now stuck with it due to the RAM prices. It's going to really suck for me if my RAM fails in the next few years.
Exactly. I probably spend 1.5x the money on a really solid build, and it's paid off - I'm still running just fine on a 10 year old desktop (overclocked i7-6850K). The only component I've upgraded (other than hard drives) is the graphics card (to a 2080Ti), and it's still running all the games I play just fine. I was planning to upgrade it on the 10 year mark, but now I'm just going to make it last a few years longer until RAM pricing becomes less silly again.
There isn't much to upgrade to if all you want is a desktop for gaming, a non-workstation work box or just a decent laptop. Much like smartphones, there just hasn't been much innovation of late. A midrange gaming desktop from mid 2023, something like rtx 4070 and ryzen 7800x3d, can still play pretty much any video game on max settings and high framerate at 1440p, and has sufficient performance for most workstation-esque tasks as long as you have enough ram. Similarly, the only thing my spare macbook m2 lacks compared to m5 is hardware av1 decoding - the performance is worse, but I don't really notice it in normal usage, and macOS 27 runs perfectly well on it.
Local AI will crush any machine, of course, and if you need a 512gb ram box you'll probably need to sell both kidneys, but you're well past the average consumer territory with those.
So, fingers crossed nothing breaks down and that some of the funny money goes into R&D which will eventually trickle down to consumer devices in a few years (16TB gum stick SSDs at sane prices pretty please) instead of endless datacenter GPUs and SSDs that will be tossed into a shredder in x years.
I've always done 5-7 year cycles on machine I've owned, but this time around its getting quite complicated as my last refresh was 4-5 years ago. So I'm kind of stuck wishing my current hardware both meets the requirements and just survives another half a decade, or my next upgrade, even if only partial, will be VERY expensive...
Similar situation here, my desktop is from 2019, with only the GPU upgraded since then. On the other side, I'm also much less interested in modern games these days so perhaps I'll do fine with a 10+ year cycle this time.
I run Qwen 3.8 at just under 2 tokens/second on a Mac Pro 2013. Why? Because I walked 8 miles to school in the snow, uphill both directions, every day.
My desktop PC turns 20 next year. I've never seen any need to upgrade it, went out of my way to build it with silent parts and it rarely if ever goes outside a single-digit CPU load (before anyone asks, software development and a bit of CAD work). I'll upgrade it when something in it physically fails, which means I'll need to replace it with new hardware with 5-10x the specs just to get the same performance.
My laptop is currently about two years into a typically 5-year lifespan, limited by when the keys start to fail rather than any need to upgrade it just because. The RAM in it is now worth more than the rest of the laptop.
This is how cloud-only systems that the LLM industries are creating the conditions for will happen. Good luck not owing anything.
This is one of the rare situations where an out-of-control homelab & PC building hobby is actually useful. With over 250GB of RAM & 90TB of storage I will be fine for a while, especially as long as DDR4 era machines remain relevant.
My issue is my digital hoarding. I've always said it I "collect" things digitally vs. physical, so I can scratch that itch without causing a disaster for my life. A half rack of JBODs is a lot cleaner and overall less expensive to operate than deciding I need to collect every single Lego Technic set ever produced or something.
The problem is I also operated under the prior "storage always gets cheaper over time" and was right at about 80% full when all this insanity hit. I'm now at 90% on my largest ZFS array (~280TB) of spinners... and I get to decide if I want to drop a few paychecks on the next vdev upgrade, or start deleting things.
Not to mention I've been holding off replacing a bunch of old 1TB SATA drives (19 years old now!!) as my scratch partition, as I was going to wait until 8TB SSD got really cheap and just buy two of them for a RAID1 for the use. And now I need those drive slots to upgrade the mass storage pool.
Talk about backing oneself into a corner at the wrong time!
I'm at 97% of my 160TB and I need to start making hard choices :-/ I have Unraid, and an old 4TB drive sitting in a drawer which I could plug into the array. It's old, and I planned to throw it out, but fuck paying these prices for a 26TB HDD right now.
It sounds to me like you don’t have a data collection hobby so much as a data hardware collection hobby.
Yeah, it really sucks I can't even afford a computer that can have two Chrome tabs open these days.
I built my machine in 2022 with 64GB of memory and upgraded to a 4090 at MSRP in 2023 so at this point I'm good but I fear the back end of the bathtub curve. I currently expect to have to hold on until 2029 and at that point the mobo and CPU will be 7 years old, and the GPU almost 6.
I’m running PCs with 13 and 15 year old motherboards/CPUs/RAM. I wouldn’t worry about seven years.
same-ish. In 2023 128 gigs of RAM and a 4090 and 4x 4TB ssds. I just looked up prices and 4090 alone is like 2/3 of what I paid for the whole system. Crazy times.
I'm glad I splurged years ago and got an M1 Max with 64 GiB RAM. Still fast enough for everything, don't plan to upgrade until this cools off.
If you bought RAM new, it should be warrantied. Most sticks are essentially lifetime support, but you’ll get a refurb. I definitely wouldn’t buy secondhand now, unless it was a very good deal. Corsair won’t honor transfers unless you’re somewhere that legally forces them to.
GPU probably not a terrible idea to pay the $20 for those aftermarket insurance plans. Do they even pay out though?
I bought a prebuilt with a 5090 a little over than one month ago, whole system was on sales.
The cost of the prebuilt is now lower than today 5090 price.
Earlier this year I finally bit the bullet and grabbed myself a 5080 for my gaming PC. I wanted to push things a bit more than my 3060ti allowed, and I figured it probably won't get any cheaper any time soon.
The PS5/Xbox Series console generation is supposed to be wrapping up soon even though it barely feels like it got started. I don't sense there's appetite for more powerful consoles anytime soon (I don't believe the PS5 Pro made waves like the PS4 Pro did), and jumps in console performance are what determine when games start targeting higher system requirements. So I like to think I can get a good number of years of use out of this.
Sony has finalized the PS6 but is going to wait out the market because they know customers won’t spend $800 for it.
I have to assume that's why they're pushing to get rid of physical discs, because they make way more money on digital sales.
I've been using computers for more than 40 years now and the one component I have never seen fail is RAM.
All the other bits in a computer I've seen fail.
How do you know they weren't faulty? Blip flips are usually silent unless you have ECC memory or are actively testing for faulty cells.
Somewhat recently, Firefox implemented memory testing in their crash reporter and found 10% of reports were due to bit-flips[0].
If it weren't for reading that HN thread, I would never have being prompted to check and find the faulty memory in my laptop. It would have gone undetected and continued to corrupt whatever was stored in the affected cells. I did run memtest86 not long after I bought the laptop, so sometime between running that first memtest86 and the 12 months after, it developed the fault.
Non-ECC memory sometimes fails and there's a good chance you'll never know.
[0]: https://news.ycombinator.com/item?id=47252971
Keep in mind that a passing memtest86 does not prove the absence of faults that could present themselves under different usage patterns.
Anecdotally, RAM failures are more common than they used to be in previous decades, I assume due to increased capacities and smaller process nodes. Looser manufacturing standards may also play a part.
The component category I've never seen fail is CPU.
I've worked in and around datacenters for a few decades now and I've seen pretty much everything fail, but power supplies in particular do so at a higher rate than I expected. I've had one literally start smoking in my home office recently too. Luckily, it didn't damage any components, but it was pretty dramatic and abrupt -- just powered on my desktop and got nothing but a fog machine coming out of the rear fan until I yanked the power cord a few seconds later.
My evga 1000w has been going strong for 11 years now. Cheaper power supplies fail often but if you buy quality and oversized they’ll outlast all your other components as they don’t become outdated.
It’s the caps that fail. Electrolytic caps have the highest failure rate of any passive component. My team tries to avoid them in our designs as much as possible. But you’re right, caps in a good design can easily last 20 years, making a power supply the longest lasting component of most builds.
My go to source for reviews was always jonnyguru.com but he closed up shop. Not sure who’s actually doing in-depth performance and component reviews these days.
Aris Mpitziopoulos at cybenetics.com
Power supplies - regardless of quality - die pretty regularly at scale. Only disks fail more often. They are the #2 component we replace in the datacenter.
Are data-center power supplies comparable to higher-end desktop power supplies? And how much do you oversize them compared to the load they will actually handle? E.g. for a desktop buying even double the power supply capacity you need won't really affect the overall PC price that much but I somehow doubt data center operators are willing to not optimize away that cost.
They typically are made by the same OEMs and I would say at par or better quality in general. I’ve had (percentage wise) more consumer grade PSUs fail on me over the past 20 years, and I don’t buy cheap stuff. Although the sample size of failure is low single digits vs thousands on the datacenter end.
I’d say on average DC PSUs are vastly overspeced for most machines. Most machines we work with are not sitting at peak load 24x7, and might be drawing half or so of their max rated power draw. Plus every server is redundant - so the modules themselves are operating at around 20-40% of full rated capacity most of their life. The only time they ever spin up past 50% is during power maintenance or a power failure of some sort that brings one of the redundant power sides down.
They do probably operate in slightly hotter conditions on average, but in much more stabilized temperature environments. Plus of course 24x7x365. But that goes for all components.
I’d say other than fans, failure rates in the PC world vs datacenter world are relatively comparable it terms of what fails most. Dust and temp swings are the large variables the average PC needs to contend with vs the datacenter.
Storage of course is the one that stands out the most - everything else is kind of a rounding error. Here though, datacenter work really does beat on the hardware more than most consumer gear so that’s expected. For every 100 disk swaps we do, I figure we do about 2 or 3 power supply modules. And maybe 1 fan swap.
I have a Zalman ZM-1000 from 2007 that's still working, my coworker at my first job lost interest in desktop PCs and gave it away to me. https://www.techpowerup.com/review/zalman-zm1000-hp/2.html
Had the EVGA 1000w fail on me. It did not fail dramatically though, rather it refused to output high power. Got a replacement under warranty
They just rebrand PSUs, not all of them are bomber. But they have 10yr warranties on PSUs and are good about honoring them. I wish they still sold gpus. I would only buy evga for every component if I could specifically for the painless warranty. Had to warranty an asus mobo and it was like pulling teeth.
Yeah I also had an EVGA 1080 Ti die on me, and got a better GPU as upgrade. Their warranty was worth the premium price for sure.
Overall I have made very good experiences with long hardware warranties. Soldered RAM on my ThinkPad died once and I got an entire new mainboard after almost 3 years (it died 2 months before warranty expired.) New one still works today but the hardware is just too old.
Funny, just had an 8 year old 1600W Corsair Titanium efficiency power supply go. Generally very high quality, GaN transistors, all japanese caps, etc etc. It was still under warranty, and Corsair refunded the original price, stand-up company. Too bad 1600W power supplies have apparently inflated ~50% since then, but still.
What brand of power supply?
> The component category I've never seen fail is CPU.
The CPU can fail if the heatsink falls off! That is something that happens (I've had one come off in my hand that was being held on by nothing but hopes and dreams, fortunately not while the CPU was powered on). More likely in a desktop tower form factor.
Even that is not guaranteed to break a modern CPU though as they all monitor their thermals and down-throttle (even below way specified minimum speeds) as needed to protect themselves. Not that I'd recommend relying on that but still, I have had cooling fail without apparent damage.
You didn't hold on to your Pentium 4 long enough. :-)
My main machines were PowerPC G3s/G4s/G5s back then with the P4 boxes I had taking secondary/tertiary roles, so that lines up.
Anecdatum, but mine still works :D
Most devices never test their memory and don’t have ecc, so you generally never know. Random instability can be fixed by restarting and your program not being in an area with bad bits anymore.
I recently had part of a 32 gig DIMM go bad on a 128 gig system. One of the larger VMs was spontaneously rebooting for no apparent reason. I was able to find the addresses with MemTest86 and map out 512 megs of the affected DIMM with the Linux kernel memmap option.
So you can still use the hardware via software fix?
Correct! Basically, I excluded the bad range so Linux doesn't touch it. It's been stable for weeks now. I should send the DIMM back under warranty.
Don't send it back. Since they can't easily replace it with new hardware, they will probably refund you the price of purchase. Which is normally a very good deal, but obviously not in these circumstances. I've seen it happen with GPUs as well.
I'm convinced it's not worth the hassle!
I've seen every single component of a system fail, the least I've seen fail has been the CPU itself (I've only seen one fail out of thousands).
Memory seems to fail pretty commonly overall, less common than harddisks but more common than motherboards.
The issue is that I'm used to using ECC ram, which fails loud when it's actually bad.. consumer memory won't tell you unless you can't boot.. and everyone disables the startup memory testing too.. (in fact, I think it's disabled by default for the last 10 years because people want to boot quickly).
I was so sure that a CPU wouldn't suddenly fail that I ended up replacing every other part in my computer before getting a new CPU and realizing that was the issue.
I had a 2012 mac book pro that burned out 2 different crucial memory sticks. Would have bricked the machine if it was the soldered-on type, but it was the last model apple sold that have swappable memory. Crucial had a lifetime warranty so they replaced them for free.
I used to game on it and I think the design just couldn't handle the cpu and dgpu being active for long periods, it would get extremely hot in the area close to where the memory was.
I've had a RAM stick fail.
Granted, I've been building my own computer for 30 years and I've only seen it happen ONCE.
But I've seen plenty of other failures:
- Two hard drives -- One got dropped on the floor, so it was no surprise, the other was showing degraded performance and SMART showed some scary numbers and I was able to replace it before I lost data.
- One GPU, replaced under warranty.
- One AIO water cooler, replaced under warranty. Somehow, the water vanished from the loop. I'm guessing a microscopic leak that evaporated as fast as it leaked.
- Two power supplies -- One randomly exploded. Just started making popping noises and shooting sparks out the back. At autopsy, I determined that the fan failed (it was hard to spin manually) and it overheated. The other was just being overloaded. I had just gotten a new GPU and the PSU wasn't big enough for it.
- Several case fans.
Never seen a CPU or motherboard failure, even when my AIO water cooler was failing and my CPU was constantly at thermal limits. Never had an SSD/NVMe failure.
I have about 24TiB of RAM on the servers at work. We have had two sticks fail (both soon after purchase) in the last 10 years.
After the second, HP replaced all that we had from that batch.
HDDs fail every 3-6 months, there's a lot of them (3-4PB).
One PSU.
One CPU/motherboard. (I let the HP engineer handle that one.)
One NVMe drive.
Anecdata: a coworker found a single bit flip corruption in a file he wrote in his computer. He ran memtest86 on his memory, and which found one of the two sticks was defective. The computer otherwise ran perfectly fine for years; and we would have never noticed if the file format didn't have built-in checksums.
Well, you’re lucky. I had to change laptop memory sticks a few times. Though not since they are soldered directly to the motherboard
> I've been using computers for more than 40 years now and the one component I have never seen fail is RAM.
I think over the decades I got one RAM disk fail. And another one didn't fail but, although I bought a pack of four sticks (2x x2), was detected (by memtest, when building the rig), as having another model name than the three others: even the shop who sold me the RAM was confused (don't know how that happened).
The usual component that I've seen fail the most would be the PSU? (before I started buying quality ones). One NVMe drive (an "ADATA") died on me even though it was nearly new.
PC assembled with quality parts usually last a very long time.
Same, it's either bad to begin with or it's basically immortal.
I also have been using PC's for a long time, thirty years. The only time I have seen ram fail during use was within the last 5 years. My 10th gen platform and 32 gigs of ram DDR4 and one of the sticks failed, this was before the ramapocalypse of course, the sticks were replaced under warranty. But I have had systems in my shop with failed ram. it happens, but its rare, unless its cheap ram, less rare.
This isn’t true anymore beginning with DDR4 speeds.
I have had a RAM stick start frequently failing ECC checks after a couple years use. I also had the VRAM on a GPU fail, but you could blame that one on overheating from the nearby compute core. I agree is rarer though.
Oh well, it's fairly common.Maybe not in a singular/several rarely used (and not monitored) desktop computers, but.... it's common once you have quite a few under your care.
on older PC bios would perform a memory test - you may remember those - and that's where it would mark a bunch of pages dead/unusuable and it was not uncommon at all.
I've seen some fail. Might have had to do with the fact that the computers were literally flooded (river came up faster than the sandbaggers could keep up) while running.
I've seen computers brought back to life after flooding by lying in rice for a week, but they weren't turned on when the water hit (and not all survived).
i've seen ram fail, some of them are bitflip, some of them just refuse to boot.
The only thing i've never seen fail is the Case lol. Air cooler too if you don't count the fan.
I used to work in IT recycling, and let me put you at ease, RAM goes all the time.
Common scenario:
200 PCs at least 6 years old, often over a decade old, come in on pallets.
We pick up the first 50, stack them on benches, and power them on.
At least 10 wont post.
Of that 10, reseating memory and other little jobs fixes 5.
Of the remaining 5, replacing or removing memory gets 4/5 them to post.
Testing those sticks in other PCs confirms the RAM is dead.
Into the scrap it goes.
The remaining PC is dead mobo, PSU or needs a new CMOS battery.
I've had a few RAM sticks fail, but the ECC usually caught it first. Also had one case that somehow the stick slightly unseated itself; dunno how but it was out of service for months until someone opened it up and reseated the stick.
...granted though I've also seen a backpane on a server fail. I didn't think that was even possible until that point.
You may literally be the only one.
Interesting, I've had two fail already in much shorter time. One was brand new, another quite old.
I've seen plenty of DIMMs fail ¯\_(ツ)_/¯
[dead]
before the memory cartel, I was automatically putting in 64 or 128GB for my work flow as the prices just seems absurd not to.
I would absolutely love if some startup came up right now to manufacture and sell RAM at normal prices to everyone else and put Crucial et al out of business for price-gouging everyone just because they can.
DRAM fabs are multi-billion dollar investments which take years to come online. That's the whole reason we're in this situation, there is no quick fix available.
Surely the chinese are spinning up fabs as we speak?
DDR4, and from some comments I have seen around here of only marginally acceptable reliability. It may still take a few years if until they have all the production processes fully under control.
I thought CXMT is selling DDR5 now?
They are, I have some in my system. They are fairly common in Australia.
China is making central decisions for its RAM makers (and who can complain now that the US isn't a legitimate free trade defender). I think their strategy could go in any direction to either get a small portion of still gouging or a large portion of a market with somewhat stale specs.
I wonder if supply chain constraints currently represent an inherent scepticism of manufacturers that demand will last. It feels to me like a side effect of the self-dealing / incestuous financing that is going on is that manufacturers are not willing to bet on it all materialising and therefore are not scaling up capacity nearly as much as they would if they thought it was certain.
It's funny you mention that because this exact same issue has come up in war in Iran and the closure of the Strait of Hormuz.
What a producer will do now is hedge their risk. They do this by forward-selling their product, in this case oil from an oil well. The peak for spot prices, depending on the particular flavor of oil, peaked at >$170/barrel while future prices (which are the prices we see) didn't really get above $100 while long-dated futures were significantly lower, rarely exceeding $70/barrel by much.
Now this is a problem because there's no incentive to invest in extra production when you can only sell it at $70/barrel, despite the spot price being up to $100/barrel higher. So new wells and new production in general is down despite the supply disruptions and high prices.
I imagine RAM is in a similar position. It takes a long time to bring on more supply and suppliers don't expect those prices to last. Likely a large amount of new supply will probably cause the next crash in the boom-bust cycle.
But there's another problem with all this. A lot of existing manufacturers are switching from DDR5 (and soon DDR6) to HBM because it's way more profitable and there's seemingly insatiable demand for high end GPUs. I expect that in the short-to-medium term we're both going to see no increase in supply and the DDR prices are going to keep rising to the point where in a year we'll probably look back on current prices fondly
All I can say is I'm happy I upgraded my laptop to 64GB (for $200) and my PC to 64GB (also for $200) last year. .
> A lot of existing manufacturers are switching from DDR5
Nobody is "switching" old process node fabs to new processes, they just aren't building new fabs for old processes.
AI and high density tasks will use memory fabricated with newer processes, freeing the old supply for more elastic, price-sensitive tasks.
The vast majority of memory that has price impacts for most consumers isn't in personal computers, it's in embedded applications that were also impacted by the overall squeeze.
9-Aug-2026: SK Hynix to invest $38 billion building new memory chip plants as demand soars [1]
"In the AI era, technological competitiveness alone is not enough and the ability to supply the required volume at the exact moment customers need it is the ultimate competitive advantage," SK Hynix said in a press release. "We reached this investment decision after a thorough review of market demand."
[1] https://www.msn.com/en-us/money/general/sk-hynix-to-invest-3...
They will do so in steps, tho. One cleanroom at a time so they don't eat a big loss if this crazy demand for a commodity goes away overnight.
They also took quite some time to decide to ramp up, as this comes a year or more late after they announced they were fully booked for 2026 and 2027.
And, their stock sold off on the news, so investors are not having faith either on the demand to justify capex expansion.
Honestly, it sucks to be a ram company, you are either buying into the hype and then spending years losing money, or you don't and leave a lot of money on the table while competition ramps up and eats your lunch (and then spends years losing money)
The party's over when the CXMT freight train will pull up and unload insane amount of volume on global markets
What stopped them before the current demand spike?
Western sanctions & the simple need to catch up to the capacity of the current memory manufacturing cartel.
[dead]
Manufacturer skepticism =/= "welp, guess we won't build more factories then"
Facility buildouts will continue at their own pace, independent of current market demands.
It's exactly this - the memory market has gone through exactly this boom/bust since I can remember - there was a great RAM price stabilization around Windows 95 (price stayed the same or went up per megabyte over years) and then a glut of oversupply.
RAM follows the logic of cyclical commodities, going through cycles of under/over investments that lead to spikes and crashes.
Counterpoint: I drove past TSMC in Surprise AZ and I counted at least 18 large construction cranes on that single site.
TSMC doesn’t make DRAM though. Their biz model is manufacturing complex logic chips (eg cpus and asics), which doesn’t have the same commodity cycle.
TSMC in the US is a bit of a special case .... wouldn't surprise me at all if they were virtue signalling to the Trump admin more than anything with what they are doing there.
How so? They are diversifying their production chain so that all their eggs aren't in one basket. Taiwan might be threatened one day but their US, Japan, and German plants will much less likely be in the same danger.
That seems wrong to me - the idea of the "Silicon shield of Taiwain" is that alle the critical cutting edge production is all physically in Taiwan & rigged to be destroyed if Taiwan is ever attacked.
If they do diversify, they are compromising a key policy protecting their homeland, it makes no sense!
"Fiduciary duty" of TSMC execs? And can Taiwan government compel them not to build out in the US?
Using tariff threats, Trump forced them to build a plant in the USA because… America First ™
TSMC US federal plant funding was under Biden not Trump (CHIPS act)
Trump did do the initial deal though, but TSMC had to cough the money up for that
TSMC wouldn’t put tens of billons of dollars in irreversible investment just to virtue signal, the scale is too large
You can’t “undo” a fab
Building new fabs is a very different bet from selling into a temporary shortage: you're committing billions based on what demand will look like several years from now
Manufacturers know demand will last to the extent that big companies are willing to commit to purchases over a given period of time, but supply today is still constrained by capacity built for the pre LLM era. The 500 billion put in by Samsung/SKHynix will increase modern DRAM by roughly 50% but it wont be fully online until 2030.
It might be fully online by 2030 if the current demand lasts. IMHO the demand is unsustainable & build on circular financing deals and hidden debt with no way to be ever profitable.
The memory cartel would be crazy to actually invest into new capacity at this point.
If you want even more bad news: display panel makers also want to raise prices
"Due to the continued rise in the prices of core components globally, the production cost of monitors has increased significantly, and we can no longer absorb this pressure on our own."
https://respawnfirst.com/we-can-no-longer-absorb-this-ktc-wa...
It would strange if just when a lot of articles mention computing / social networks are detrimental, governments start passing laws to forbid platforms and/or phones for kids, and now price gouging flares up everywhere, this would entice people to just stop using these devices and maybe enjoy natural taper-off.
I like this idea! Perhaps folks will game less on PCs and consoles. But would they shift to playing on their phones? Or play less video games overall?
Also, when OpenAI's new device comes out, would people shift to talking to these small devices?
People do not want to talk to their devices because they are in the room with other people and they dont want to disturb them. I think their device is going to be a flop.
It's going to flop because everyone is already able to talk to Google Assistant to manage things, and outside of very specific circumstances (e.g. driving) everyone is still tapping screens instead
Doesn't solve the present shortage, but perhaps this is a good opportunity for software developers to start caring about memory usage again, since consumers may be forced to keep their existing hardware for a few more years.
Yea here's hope'in. Memory management as been really inefficient for a long time.
Honestly? Is it even going to be a drop in the bucket? This push is from massive data needs. The amount of ram code uses becomes an even smaller percentage of the growing pool.
Who the hell needs more than 640 KB anyway? :-)
Jokes aside, I've been extremely frugal with memory in my latest project, in the hope that it can fit in a few megabytes of RAM. But we have come a long way since the eighties, and now an audio codec (OPUS) takes the best part of a megabyte, and so do "frugal" scripting engines such as Lua and even SQLite \o/.
I almost upgraded my computer a year or two ago, but decided to hold off for a bit longer.
I've always been fine with being "behind" on the latest and greatest. I don't mind watching a movie, playing a game, buying a cell phone, etc, that was released a few years prior. If it was really all that great when it came out it should still be good later on.
Buying slow usually saves a.good bit of money when prices go down over time... but this time they went way up and now I'm kicking myself on it. Not upgrading the computer for now, but it is a bit overdue since I had already put it off for many years.
100% what I do. I buy a few years behind the leading edge, and keep the device until it either dies or will no longer do what I need it to do.
My main desktop computer is Haswell era. Still does fine, for what I do.
Except staying on the edge costs exactly the same as not, when you take the resell value of the components on the second hand market into account. The problem is that some can afford to lock their capital in the hardware (via one-time investment) and some do not.
R9700 will lose less than 30% of value in the upcoming 3 years. Give or take. Yet people buy ancient cards for way more than 30% of R9700's cost.
> ...when you take the resell value of the components on the second hand market into account.
When taking the resale value of used kit on the second-hand market into account, one must also take into account the hassle of selling on the second-hand market.
For me? Personally? That hassle absolutely eliminates any money I would get from any plausible sale price. If I'm getting rid of used kit, it's going to be as donations to people I personally know who will come and pick it up at the agreed upon time.
Yes and I seriously doubt gps insinuation that resale could make up for the price premium of top of the line components when they were consistently beat by the next years significantly cheaper mid range components.
I thought about upgrading my macbook, but I have the M1 Max and it still doesn't feel slow, so I decided it'll probably last another two years. That said, now I feel like this price raise is maybe permanent. If so it will definitely eat into computer makers profits, people will not be upgrading nearly as often.
I ordered a new CPU on sale a while back, right when the memory prices started spiking. I didn't even know RAM was going up in price. Realized the CPU needed a new motherboard and ram so I ordered the motherboard, started looking at ram and realized it would cost almost as much as the Mobo and CPU put together. Cancelled the orders because I wasn't intending a big expensive upgrade.
What I have is still fine, though I do regret cancelling now.
I upgraded a year ago and I bought a bunch of extra RAM because it was "so cheap", why not. Man did I get lucky, I got it right before the price hikes started.
Same, my desktop is a top end full of ram beast that I thought I overpaid for because motherboards had gone up. Bought everything before the surge. Past me was damn lucky. Wish I had bought some stock tho.
There is no law that states memory must be a consumer product. I truly believe the era of personal computing was a beautiful blip that our masters are beginning to take away from the general public. It could be terminals and time sharing for the foreseeable future through "smart" phones.
Personal computing has been in a sad state for a while now. Only "PC enthusiasts" were gamers. Rest of the population uses their phone 18hrs a day and seldomly use PC for Excel tasks at work.
Apple is making Macs and iPhones more powerful every year. Macs are more affordable and powerful that other computers since the M series. People are buying them to run their own private on device AI. This is a signal for personal computing.
> Macs are more affordable and powerful that other computers since the M series.
Apple has been raising prices on Macs and other products.
Just so I make it clear. I'm deeply opposed to this change.
This is the ultimate dystopian outcome. How do we stop this?
[dead]
Is the end result of this is that anything that is useful to AI going to be permanently more expensive, as the value a human hobbyist can get out of a certain amount of RAM will always be far less than the value a frontier AI model can get out of it?
If demand remains as elevated as it is you will definitely see new fabs breaking ground to meet it. China is expanding a lot as well. Even if AI doesn’t go anywhere there is no way prices remain this elevated or climb forever
More likely AI companies go bankrupt and market is flooded by repurposed hardware sold for pennies by creditors AND Chinese memory from their new fabs, hopefully also killing the existing memory manufacturing cartel.
So basically utopia. ;-)
Just had an HDD in my array die. Thought I'd replace it and add another spare. The prices are up 300% from 2 years ago. For spinning disks. Why?
My theory is demand substitution. NVME/SSD is up so much people at the margin are switching to spinning rust, which nobody was prepared for, so now they too are drowning in demand.
(1) Because they can. Implied
(2) If you're WD, Seagate, etc. you may shift all your fabs to higher growth areas which means you create less supply for less demand products (spinning disks)
I've resorted to buying refurb HDD's for my NAS (which I backup to rsync.net) but like damn!
Boy am I feeling this. I had my eye on building a new Threadripper machine, which requires DDR5 RDIMMs. Much to my surprise while happily building out a parts list one fine evening I learned that these are going for between $500 and $1000 for one single 16GB module right now. I think I'll hold off and squeeze a few more years out of DDR4 and PCIe 4.0.
I just laid an arm and a leg for 128GB of DDR4. More than I normally would have ever spent, because the alternatives were far worse.
I'll go hug my 128GB ECC DDR4 from 2019. With eletrostatic-safe clothing of course.
$580 aud for a single 32gb module last month, that really hurt
These unintended (or intended?) affects of the AI boom are yet another cause for hate among a key demographic (people who use/buy/build PCs) who might otherwise be pro-AI but because of this are quite anti-AI.
Seconded. The mad dash for everything is screwing up future expansion opportunities, because everyone in the market wants a future slice at any cost.
Inflation, cost of living, purchasing power... I feel like life is getting really difficult right now for most people in ways it hasn't been in decades. I don't think memory prices going up this much this fast is good. I think it's a sign of how bad things are and that we are heading in a terrible direction. I think we're going to get a BIG wave of inflation in 2027.
Life on earth today has objectively never been better in the history of our entire species. Try to retain some perspective.
On a macro level you're correct. On a micro level, I miss the pre-pandemic, pre-AI era before stuff got expensive, before Russia invaded Ukraine, before the Israel/Palestine strife went into overdrive, and before the US declared war on Iran, Canada, Greenland, and Minnesota.
Almost everything that is upsetting is you, is being intentionally shoved in front of your face to make you feel mad. You are being played.
Your great-grandparents had like a 40% child mortality rate. You’ll be ok that some people around the world are fighting which has been a human constant throughout all of history.
At the same time hundreds of thousands of children die annually from diseases which would cost pennies to treat. From starvation.
The world is still a horrible place for many people, and there's no real reason why it has to be that way.
What are you doing to solve it? How much are you donating towards medical supplies?
Oh, right.
Yet… violence has plummeted world wide and there has never been a safer, healthier, more prosperous time in human history.
But, you could understand the hedonic treadmill, read some history books… OR… cry about it into your thousand dollar smartphone.
Im not saying "it was better 100 years ago" I'm saying we're in a local dip in the upward curve and it was better like 7 years ago.
> Life on earth today has objectively never been better in the history of our entire species. Try to retain some perspective.
The typical relative privation argument trap
By what metric? Houses, healthcare, cars, college... everything that matters is way more expensive for current generations when you compare to multiples of income or ounces of gold. We're working harder for less to show. Yes, there are FSD vehicles and rockets that can land and be re-used... but steak is up 30% from a year ago. In most ways, life is going in the wrong direction. The only thing getting 'better' for most Americans is that you can get a bigger TV now than ever.
Your grandparents almost certainly had siblings or knew of their friends siblings dying from now completely preventable diseases.
The quality, safety, comfort of low income housing in the United States is exceptional in historical terms. The fact that you CAN be anything you want is an unheard of concept even 60 years ago.
The worst cars on the market today are modern marvels of engineering compared to decades ago. Let alone throughout the history of the country.
College, lol, you literally have a secured loan to do study whatever you want with absolutely no required direction from anyone else… to receive the worlds knowledge if you want it.
Oh, things are more expensive? Yes, they’re massively higher quality and larger.
You’re mad about things, OK. They’re objectively the best they’ve ever been. This is indisputable, but again, I get that you WANT to be mad.
Okay Ben Shapiro, but you skipped an entire generation to make your argument and are ignoring what's happened in the last 5/10/15/30 years.
I don't WANT to be mad and that shows you're arguing in bad faith. I want to stop checking my bank balance before I buy toilet paper and paper towels. Life wasn't like this until very recently... We're destroying the middle class and making home ownership virtually unaffordable. The average first time home buyer is 40. The economy is not healthy and has not been this disconnected from the stock market since probably 2007.
> By what metric? Houses, healthcare, cars, college... everything that matters is way more expensive for current generations when you compare to multiples of income or ounces of gold.
Check the price of food. Also, check what the price of a movie resolution film camera cost 20 years ago and realize that you're getting that in the price of your iPhone. Plus of course the corresponding computing power.
> but steak is up 30% from a year ago.
Well that is an extremely specific cherry pick.
It's crazy if you think that some years ago were the cheaper part of a personal computer compared to motherboard, cpu, gpu and so on.
last year i mis-bought a ssd on amazon (wrong size or model i can't recall) so I asked for a return, surprisingly amazon wanted to save on transport that they offered to pay me to keep it (i don't want to spread stupid claim but maybe the full item price), i was confused and worried it would create issues down the road so i sent it back thinking i'd buy the proper item the month after.. and now i'm running duplicate search programs to save space on drives.
Any new fabs on the horizon or do the powers that be expect the music to stop sooner than that?
They are building fabs on track to get online towards the end of 2027, full capacity production expected for 2028. These are serious pieces of very complex technology that take years, not months, to build.
CXMT is a memory fab that just opened in China. They achieved a 90% yield on DDR5. Hopefully that could ease up the supply squeeze.
CXMT has been in a game for a while, at least since 2025, because my SBC is using their LPDDR4 RAM.
It's very unlikely to have any meaningful effect. CXMT DRAM will also be hoovered up the moment it hits the market at the same inflated prices. Even if they had the capacity (they don't) why flood the market when you can just print money?
[flagged]
I believe they've been building them like crazy since covid, but it takes a looong time for them to come online (like 5-10 years?)
My Dell rep quoted me $736 for a single 16GB RDIMM for the newer Dell Pro 16 laptop, which retails at its baseline configuration with 16GB for ~$1000.
Dell wants $4,507 (web price) for each 32GB RDIMM in a PowerEdge 6625 (16GB is showing as unavailable)
Storage has more than doubled as well. The same SSD (4TB M.2) I bought in September 2025 for £229 is now £500.
I also had another order where I paid £30 for a 1TB SSD (sata though) and it is now more than £100...
Not just SSDs, HDDs are also way more expensive. End of 2024 I bought 4 14TB drives for ~$250 CAD each, $400+ for even a refurb drive now.
The possibility that these fail does concern me.
And that's the really fun part of all this, that weak substitutes will also become more expensive. But on the bright side, I've got dozens of sticks of DDR3 RAM that I'm hoarding like Pokémon cards. Any day now, we'll get desperate enough to start using them again!
I would be very surprised if the cost of the inputs being hoovered up by AI drops after this. Without ever forming a real cabal, I think the functional behaviour of the disparate people in this supply chain are going to hold on to the profit margin because there isn't a clear basis to stop them.
Sure, some imaginary perfect market would drop. This one is far from perfect, and it may drop by marginal percentages. It won't drop back to what it was before it went up 4x in price.
I would be very surprised if a market for older, superceded chips takes off, not because it isn't a brilliant idea, but because in order to use them, you have to make boards and systems which can consume them, and right now, the people who do that for a living are making a good living selling stuff as inputs to the people making a good living selling memory at inflated prices to the AI factories.
It's turtles all the way down.
The weird part is that RAM used to be the component you barely thought about when budgeting a build
If this was a competitive market some suppliers would be ramping up pc memory and avoiding the AI buildout. But instead we see all of them diverting to GPU memory and cutting production of PC related parts universally. This sure looks like a coordinated move.
I think there is a real point people are missing here. Memory is going sky high because most of it is being shifted to GPU specific ram for server build out. This means that when this ram hits EOL it -won't- flood the consumer market in a meaningful way. It is disposed of because most people won't put a server GPU into anything. It is like a used big rig, built for purpose and when it reaches end of life it is truly at end of life. Consumer DRAM will stay very high while the cartel retools and moves back into that space. This means that the cartel wins and pays no real price for it when they shift back to producing consumer ram, they step immediately into a massive demand for pc RAM and keep raking in $$$. But they all have to do it in lockstep so that they have a market to come back to. Otherwise someone would become the king of PC ram and they couldn't re-enter easily if/when the AI server market fails. Being in lock-step means they rake in more money and mitigate the risk of an AI bubble, all at the cost of the consumer.
At least that is what this looks like to me. They know they could make a lot of money ignoring the AI server build out but they all magically didn't. A competitive market would have seen a spike in price when this much demand came in, but the lockstep movement of the big three shows that they are coordinated and not competing.
It is a competitive market, all the market is selling to the most competitive bid price.
When there are massive costs to entry, then this is what happens on a huge demand surge that can pay 10x more than everyone else.
> This means that when this ram hits EOL it -won't- flood the consumer market in a meaningful way. It is disposed of because most people won't put a server GPU into anything. It is like a used big rig, built for purpose and when it reaches end of life it is truly at end of life.
This was true until the last year.
P40, P100, V100, A40, etc are selling like hotcakes and prices have been rapidly increasing.
Not sure about that, V100s are all selling for less than $1k despite being 24-32GB VRAM around 1TB/s memory bandwidth, which is the same price point that 4090s and 5090s are commanding $4k-$5k for. Hardly hotcakes.
CXMT is almost certainly going to eat the DDR5 market (not that will change anything in the next year or two)
500% in a year is just insane. I knew RAM prices were going up, but $3k+ for 128GB of DDR5 makes me wonder how much of this is actually AI demand, and how much is this just manufacturers taking advantage of it. They have done this in the past after all
I don't understand the distinction you are making. If you have a massive excess of demand ready to pay any price, and a limited supply, of course the sellers will take advantage, that's how any market works. They can do that because of the lack of supply.
Over the long term demand will rationalise and supply will adjust. But in between there is a squeeze, and the right price is the maximum price buyers are willing to bid for the limited supply. Sucks for other buyers, on the other hand it ensures demand destruction, i.e. that the limited supply goes to the buyer who need it the most, other buyers will delay their purchase and do with what they have.
> If you have a massive excess of demand ready to pay any price, and a limited supply, of course the sellers will take advantage, that's how any market work
The question is how quickly the supply adjusts. In a free market, the suppliers would respond to the price signal by quickly expanding their production to meet the elevated demand.
But a market captured by a cartel will be much slower to expand the supply.
So, the test isn't "how much is price going up" it's "how aggressively are the RAM producers expanding supply production"
Memory fab build out times are years not months. The existing ones already run 24/7. If these companies could pull more supply out of thin air they would, but they cannot.
Would they?
Cartels aren't known for massively increasing production to satisfy all demand. Far from it I'd say, they're more likely to expand supply to such a level that they can still reap massive profits. Very few countries/companies have the knowhow and willingness to be able to build more.
OPEC is a perfect example of this behaviour. They set hearly production goals to keep prices high. Or the Quebec Maple Syrup Producers organization that limits the amount of syrup a farmer can produce and sell in a year to keep prices high.
OPEC's aim isn't to keep prices as high as possible. If they go too high, people will start buying electric cars, investing in renewable energy, etc. Basically they want to make oil as expensive as possible while preventing it from getting so expensive that it's disruptive.
OPEC's goal is as much profit as possible. They're not thinking in the short term but long term. If as much possible means over the span of multiple centuries then that's what they're going for.
I think we're in agreement here. I don't think memory manufacturers are interested in a market disruption even though profits are incredibly high for a few years. Make it too profitable for too long and others will arise but keep the supply at a steady pace with decent margins and they will remain the only ones on the block because of massive upfront cost.
China is, to my knowledge, the only country that has so far had the willingness to seriously invest. Not necessarily for profits but for independence reasons. I don't see many (or any really) other countries following suit.
> OPEC is a perfect example of this behaviour.
It really is. First thing I searched for in the thread.
> Quebec Maple Syrup Producers organization
Can you just not join the organization?
Maple trees are living things and have a right to reproduce, you can't fault a bird for shitting maple seeds all over my farm planting maple trees everywhere.
It looks like you're forced to participate by law once you sell over a certain volume.
> Backed by the Canadian civil courts, the federation has the monopoly for selling Quebecois maple syrup on the wholesale market, and for exporting it outside the province. Producers are only allowed to sell independently cans of less than five litres or 5kg to visitors to their farm. https://www.bbc.com/news/business-35028380
I saw a documentary(forgot the name, sorry) about that that said, no you can't choose not to join. I forgot if it was about this organization but it was about maple production in Canada where a farmer wanted to do their own thing and they were intimidated and all sorts of nasty things happen to them because they didn't think the org was just and wanted to sell freely. Very interesting documentary and an eye opener that even farmer unions can act like organized crime when there's money involved.
> you can't fault a bird for shitting maple seeds all over my farm planting maple trees everywhere.
If only building a ram FAB was that easy :).
The best hope is that cxmt gets going quicker than expected and is able to satisfy a decent chunk of the Chinese demand for RAM which would - in theory - reduce the demand for ram from the few other producers.
Can you list out the precise "cartel" company, their supply chains and the board of directors, their financial and political ties?
Lot of unsubstantiated claims with low fidelity meme evidence
There's details in the class action lawsuit
https://stream-ai.s3.us-west-1.amazonaws.com/class-action-co...
https://en.wikipedia.org/wiki/DRAM_industry_price_fixing
> Cartels aren't known for massively increasing production to satisfy all demand.
Don't worry they will leave money on the table and lobby for protectionism when china comes and eats their lunch.
You're correct in theory. I'm just not sure we have a cartel situation. Micron is a US company, and it controls 25% of the global market. I see no evidence that they nor SK hynix nor Samsung are intentionally limiting production.
I do see people arguing, "they had years to increase production - why didn't they?" They've been remarkably candid about this: because RAM production has been losing them money in the past. Because they don't have a cartel, prices were on the floor. Investing in more production would have been corporate suicide.
Even now, there is considerable risk that the bubble will burst, and all this investment in new production will be flushed down the toilet. RAM production is a brutal business.
But are the memory chip producers acting like a cartel? In the past they have not. When there was a shortage, they increased production as rapidly as they could build the new fabs, crashing the price of memory.
As others have noted they've been found colluding previously, more than once. Even admitted guilt and some people went to jail over it.
And the reason the market is so small is because of players who did as you propose, the DRAM market crashed on them (not due to them but due to broader market dynamics), and they went out of business.
So there is, quite perversely, strong incentive to collude if possible and if not to abide by an unstated agreement to be extremely slow to expand DRAM capacity.
That said...
SK Hynix has two fabs under construction, one set to open later this year. One for 80k wpm, another 360k wpm. A third was recently approved to start construction.
Micro has an overhaul of Fab 6 in Virgina near done and slated to produce first wafers this year. It also has a large expansion in Boise nearly completion for 300k wpm. They're also building a "megafab" site in New York (my guess is >1m wpm) and expanding at their Japanese plants (40k wpm).
Samsung has two under construction, P4 for 100k wpm.
Nanya - historically a secondary player focused only on producing non-leading-edge DRAM is making a play for the leading-edge market with a new fab dedicated to DDR5 and HBM.
It appears they've all decided that the demand is sufficiently sustained that building new fabs is worth doing... but it takes time. If my guess is right Micron is making the largest play with their New York site at 1-1.5m wafers per month. That would give them a large chunk of the market.
(wpm = wafers per month)
>In the past they have not.
Well guess what :) They've done it before, they will do it again.
"To date, five manufacturers have pleaded guilty to their involvement in an international price-fixing conspiracy between July 1, 1998, and June 15, 2002, including Hynix, Infineon, Micron Technology, Samsung, and Elpida.[1]"
[1]https://en.wikipedia.org/wiki/DRAM_industry_price_fixing
That was 24 years ago. Half the people commenting on this board aren't even that old. How many of the shareholders on the boards or executives at the helms of those companies do you think are still there?
Wasn't that bad for business last time? I could see them erring on slower build up because of how it went before, if they're not sure how long this demand will last especially
Which is exactly what's happening. There used to be a lot more then 3 makers but up till very recently memory was a comedically low margin business you didn't go into.
I didn't say the build out time was months.
I said:
> the test isn't "how much is price going up" it's "how aggressively are the RAM producers expanding supply production"
There are several Chinese fabs that they could contract out production to. In fact that's exactly what Apple tried to do, before the U.S. government went full protectionist and added the two memory manufacturers to the banned imports list. There's also the Texas Instruments fab that's underutilized right now. The extra capacity exists, it's just Micro and SK Hynix are throwing their money around to stifle anything that could jeopardize their spike in profits that are historic not only for their industry, but for manufacturing as a whole.
Large barriers to entry are anathema to competitive markets, so the conclusion here is still that firms are benefitting from the lack of a competitive market.
Memory fabs take years to build, they cost a lot and there is no guarantee that the demand will stay as high years in the future. It's all very complex.
Yep, the hard part is not going out of business in the slow years.
I'm curious what the limiting factor is for "years not months" when building these fabs. DRAM is a commodity, there are only threeish major manufacturers, all have huge amounts of capital, multiple nations bend over backwards for new semiconductor production projects, and there seems to be unlimited money going into purchasing these things.
Even when it comes to manufacturing the tooling/testing infrastructure, it's just a question of money aiui.
It's not "just a question of money", time is absolutely a factor. Getting nine women pregnant doesn't give you a baby in a month.
And yet, memory manufacturers have been caught (in legal cases and fined) many times to collide and keep prices high artificially in the past.
> But a market captured by a cartel will be much slower to expand the supply.
Oh yea, governments can fix by putting 100 billion in memory fabs right now, and when demand crater in decade or so, let tax payer eat the cost of those fabs. At least it will benefit poor computer users like me.
Free Market doesn’t mean new ram production can “just be stood up”
And given that these are independent actors in the free market, they’ve seen the history of boom/bust cycles in this market
Free market doesn’t mean “everything just magically happens”
As has already been pointed out further down, OPEC is the perfect example of this. Things don't really work that way when the time scales are so extended. What we need is fracking, but silicon.
This system is too complex, the few individual players too big, with too many artificial distortions for the adjustment to equalize out in a way that's anything but years-long (it may have shock-waves that extend past the decade), and highly destructive to everyone who isn't OPEC, or their largest customers, in this analogy.
The brunt of it will be paid by the people further down the food chain.
Even if the memory market was completely competitive the time it takes to expand supply is measured in years. There was no way for them to serve the AI demand without reducing supply given to everyone else in the short run
> There was no way for them to serve the AI demand without reducing supply given to everyone else in the short run
I didn't say that there was a way to do it in the short run. I laid out what the test is to determine if there's cartel-like behavior or competitive-market-like behavior.
I will say, in a market with only 3 players and very high cap-ex cost, my prior is that you're likely to see more cartel-like behavior than competitive-market-like behavior, but that's still to be determined.
> In a free market, the suppliers would respond to the price signal by quickly expanding their production to meet the elevated demand.
Why would they?
And you claim to have historic examples where they did?
Generally speaking, if demand stays high, and existing suppliers don't increase supply, the high prices create an incentive and opportunity for new market entrants to increase the supply.
Yes everyone took Econ 101.
Now what do you think happens when developing new supply is substantially constrained by lead times?
Coz a whole lot less of you have ever dealt with manufacturing logistics.
Prices go up, until new supply comes online or demand wanes. Econ 101 is not invalidated just because you are having to go without due to high prices.
Some supply and demand curves move in days, some in months, some on years.
Yes that would be my exact point.
Yes, "quickly" is obviously industry-dependent. If they were fabricating something extremely easy to manufacture, "quickly" might mean months-to-a-year. Fabricating something very complicated like memory "quickly" might mean years.
That said, even though we might have to wait years for new supply to hit there market, we probably don't have to wait years to see signals of that. We will likely know that manufacturers are starting to build new supply long before we see the chips hit the market.
In the current environment unless people start decommissioning datacenters en masse supply will be locked up till the early 2030s IMO.
Even if you started a new fab today, it's 5 years till it comes online. The great hope here is that China's DDR5 yields represent a high reject rate of an enormous amount of new capacity that they'll dump into the market to get a sovereign capability in that area.
> In a free market, the suppliers would respond to the price signal by quickly expanding their production to meet the elevated demand.
That’s a good one!
[dead]
> I don't understand the distinction you are making.
20 years ago: https://en.wikipedia.org/wiki/DRAM_industry_price_fixing
> According to the one-count charge filed in San Francisco's federal court on Thursday, Park conspired with unnamed employees from other memory makers to fix the price of DRAM sold to computer makers from April 1, 2001, to June 15, 2002. The government says the move directly affected sales to U.S. computer makers Dell, Hewlett-Packard, Compaq, IBM, Apple Computer, and Gateway.
etc
"that the limited supply goes to the buyer who need it the most" - I would say that the limited supply goes to the buyer who can afford to pay higher than others. Someone may need the memory more, but may not have the deeper pockets to pay for it like NVidia and Apple might.
I would have framed it as "buyer who can get the most leverage from it", which equates to being able to spend more. It's an economic argument, not a values argument.
You suppose what you'll do with the ram has an incidence on your capacity to buy it which is only true if you are part of those able to get financing.
Gamers may need the memory more, but AI labs can use the memory more productively
Profitably-for-the-fabs*
>that the limited supply goes to the buyer who need it the most
Everything else you said is accurate, but this is not. Supply will go to those most able to pay, not those who need it the most. It is not a moral allocation, which your language implies.
Nobody really "needs" RAM immediately though, like a heart or a lung or something.
That’s a bold statement to make in the 21st century when nearly everything is run on computers, including cars, planes, hospital networks, etc etc.
Not really. They can get RAM just fine, and delay upgrade/refresh cycles to stretch existing component lifecycle, which was already very long if they needed to.
They can get RAM at drastically more expensive rates. And not all critical applications have unlimited budgets to absorb the increase. It’s the reverse Jevon’s paradox.
What if it’s for the machines that support this surgery?
Someone please put RAM into Maslow's hierarchy of needs
"limited supply goes to the buyer who need it the most"
No? It goes to the buyer with the deepest pockets.
it is probable that you don't get the deepest pockets by wasting money. anyone that pays so much for memory means they expect to get even bigger value from that.
"It is probable"
So RAM manufacturers will turn down company B's $2M offer for company A's $1M offer because company A "needed it more"?
I'm actually done discussing this...
My comment is about B offering 2M when normally the price of that RAM was around 1M. Most probably B has a way to get > 2M value from that RAM.
This is a misunderstanding of bubble markets to an extreme degree.
In a bubble market where B has massive amounts of capital allocated to it in a way that will likely collapse in the future than it's very likely they will never recuperate the 2M they put in the RAM. Instead buyer A that would get at least 1M of usefulness if not more, gets far less, crippling their business. Then when B collapses a ton of RAM is dumped on the market for far less than 1M cratering the RAM providers business for years.
> I don't understand the distinction you are making.
It's pretty simple. In a correctly regulated market the sellers are in competition with each other so there is a limit to how much they can increase prices, since they'll just lose business to their competitor.
In a poorly regulated market that allows competitors to collude with each other to raise prices by the same amount, they'll inflate prices much further.
Real history here on this subject is the LCD panel price fixing scheme: https://en.wikipedia.org/wiki/TFT-LCD_(Flat_Panel)_Antitrust...
[dead]
>I don't understand the distinction you are making. If you have a massive excess of demand ready to pay any price, and a limited supply, of course the sellers will take advantage, that's how any market works. They can do that because of the lack of supply.
The question is how much of this demand is actually based on AI, or just an excuse to fleece regular buyers (and AI is not really the actual demand driving such prices).
Memory manufacturers have driven up prices via collusion many times in the past (and paid fines for it), without AI or some similar high demand vertical market existing.
But I don't think it's the case now, see my answer elsewhere here for why.
> or just an excuse to fleece regular buyers
In a market with healthy competition (no collusion), this generally can't happen. Every single vendor is incentivized to lower their prices slightly at all times, with a floor at cost plus margin.
>In a market with healthy competition (no collusion), this generally can't happen.
We're talking about the memory market here.
The "healthy competition (no collusion). Every single vendor is incentivized to lower their prices slightly at all times, with a floor at cost plus margin" is the fairy-tale version they teach at 101s.
why plus margin? the floor is cost
Floor is below cost if you think you can put one of your competitors out of business before you run out of money, and you'll make the money back after doing so. And this includes even after fines.
that's true, also see "loss leaders" like milk in supermarkets
I'm sure there will be no large scale leverage campaign against a country who is enabling profiteering at this level. No defense agreements removed, troops removed from the country, or large tariff increases around failed investments.
I'm sure it will only suck for the buyers. This has no effect on inflation and probably will be overlooked by the current admin.
They fixed prices in the past and now they have an easy cover, like egg producers had with eggs after the cull a few years ago
The distinction is that on the latter case you have a cartel that decided to fix prices and fuck everyone in the ass.
Byron the bulb checking in.
Google: Cartel
What is cartelization for 500%, Alex.
I'm not sure how turning into crabs is going to help us download more RAM.
In an absolutely free market? sure. But we don't live in an absolutely free market, the memory manufacturers have refused to budge on creating new fabs until recently out of caution and to be able to create a demand excess.
They have slow rolled rollouts and everything, if there wasn't CXMT they would probably still wouldn't have budged on building new fabs for a while longer.
You can absolutely price fix/manufacture a shortage. Ever heard of merchants setting fire to crops after hoarding to sell at a higher price?
The issue is the distributer and producer are the very same people this essentially means producers have no incentive to rationalize supply, they can make much more money of the long drawn out shortage.
I think everyone is aware now that you don't build a fab by snapping your fingers, and you need to amortise it over many years. The said manufacturers are just out of a massive slump in demand during covid, when they were selling at a loss, and these cycles happened many times in this industry. It is understandable that they don't jump all-in on the AI bandwagon, which, god knows where it is going and for how long.
If you want evidence that no one predicted this, look at WD. They disinvested from sandisk up to early this year. Now sandisk is worth more than WD. Doesn't seem consistent with a concerted strategy to gauge prices.
> You can absolutely price fix/manufacture a shortage. Ever heard of merchants setting fire to crops after hoarding to sell at a higher price?
Fine. Show me which manufacturer is cutting production to corner the market. I don't see that.
Yes when there are large fixed costs to expanding production, that will happen more slowly, after determining that the high demand isn't just a short term spike.
It's not like wheat, where you just can just plant more at marginal cost (I'm sure someone in farming will now explain why that isn't exactly true either).
I'm not super familiar with farming, but I'm pretty sure there is an upper bound on how much wheat you can plant dictated by the amount of land you own.
I feel like I and HN live in different worlds,
https://www.blocksandfiles.com/ai-ml/2026/01/20/samsung-sk-h...
Big manufactures have absolutely cut production of lower margin products and it's not like they can re-purpose all of the equipment used for it for HBM yet. edit They can definitely focus more resource because of it on HBM though.
Things have improved a bit since govts are starting to get angry.
Tldr; no one is cutting the production of high margin stuff, but low margin stuff can go and die, because it's not a free market it's a stock market these companies want to show off better margins in every unit there is no reason to keep inefficient things running when not doing it pays more.
They just don't want 30-40% margins anymore. That is not free market that is quite literally serving their own interests because they know there is a monopoly.
All big memory manufacturers are heavily supported by the govt, that's what has made them the monopolies they are Micron supported by USA (they crushed Japan's memory industry for it), and Korean giants are more tangled up with the govt than well most people in US would realize.
Somehow people think free market means anything goes, which honestly is where we stand today so I don't blame anyone.
But using govts to crush competitions and build yourself up on subsidies and similar isn't really free markets. China hasn't been able to grow a significant memory chip industry yet because of lack of EUV machines and similar technologies blocked by the US, which is also not "free market".
As for WD, and Sandisk separating I feel like this was more mutual and they new HDDs or SSDs being together was a distraction for both. Which feels more strategic, think of it like this, the shareholders who had the shares how hold shares in both and together they might have likely been worth less so shareholders made money.
Aka good for free market. I don't see how that relates to the actual memory crunch issue more broadly.
Margins pre-AI were more like 5-15%. Even if the margin was 5%, if prices increased 5 times but product cost didn't move, then margins are now 80%. That's not a 30-40% product, your argument doesn't make any sense.
Feel free to spin up a fab and undercut the current prices, I'd certainly appreciate it!
I don't know if I would personally fund such a venture though, for roughly the same reason existing manufacturers aren't expanding production.
How about I threaten you with mandatory 10 Electron applications / AI agent harness on your computer? Now you must be ready to fund my venture :)
If you did, then the big manufacturers would just sell a competitive product, at a loss if necessary, until you are driven out of business.
CXMT is about to reel in Apple as a customer and I don’t see Samsung doing a whole lot about it.
> the memory manufacturers have refused to budge on creating new fabs
That's EXACTLY how a free market is supposed to work. The existing memory manufacturers are FREE to not build new factories, and anyone else is FREE to build those factories in their place.
Then perhaps we’re seeing one of the downsides of that. When 0->1 is prohibitively expensive, existing players can continue to raise prices *very* high because cost-of-entry is extreme, until it’s not sufficiently extreme.
And then we have bubble issues because the introduction of a new, cheaper player, causes the bottom to fall out and for the circumstances that allowed the new manufacturer to exist to completely go away.
This is also caused by the FREE market - consolidation, all markets tend to consolidate around a few companies due to economy of scale.
The fix is regulation - we are going to have NOT-FREE-MARKET electricity, NOT-FREE-MARKET water supply, .... for vital services.
I'm not talking about you, it's funny to see anti-regulation HN complain about a FREE MARKET working as such. Suddenly it's not so fun where you are at the losing end of the free market when it turns out some are willing to pay way more for RAM than you are.
Everybody needs electricity and water, but nobody needs more RAM except for companies. And they can pay for it.
You can travel to the ends of the earth, to a village down miles and miles of dirt roads and still see someone bent over their phone. People buy RAM.
> but nobody needs more RAM except for companies.
I'm amazed you could write this in apparent seriousness.
Be so kind then and tell me what a normal citizen would need more than 8GB of RAM for, if it isn't for professional work or a specialized hobby?
A Mac with 8GB of RAM is blazingly fast for any task your average person would need for it to do. Which doesn't include gaming, mining crypto, or heavy AI models.
Maybe I've missed something? So please tell me exactly what task requires more than 8GB of RAM which is so necessary for the average citizen that the government should subsidize it?
A Windows 11 machine with 8GB of memory is borderline unusable.
We can go in circles about how developers need to optimize their applications, etc but truth of the matter is that 8GB is considered pretty low for a lot of "average" use cases. MacOS is an exception to the rule and even 8GB on MacOS is limited with multiple Electron apps.
And why should the government and the tax payers be held responsible and be forced to pay for the bad decisions of Microsoft and their customers?
MacOS is not limited in any way for the needs an average citizen has from computers. The average citizen is even fine with an iPad. 8GB is considered "low" by people here, just like a 170 horse power hatchback is considered "low" by muscle car enthusiasts and mining companies looking for dump trucks. But those uses have no bearing on reality for most people.
> And why should the government and the tax payers be held responsible and be forced to pay for the bad decisions of Microsoft and their customers?
Ah, a motte and bailey argument! I always love those.
https://en.wikipedia.org/wiki/Motte-and-bailey_fallacy
I don't find anything hard to defend about my argument. It origins in response to people arguing on the lines that the government should subsidize RAM because it is expensive. But it is not expensive for a customer unless they're getting a lot of it. My argument has been the same all through this thread, while hackers are arguing all over the place.
The most absurd argument is that any consumer would need more than 8GB, when it is proven beyond all doubt in physical reality that 8GB gives you extreme amounts of power in for example an Apple device - at consumer friendly prices.
> I don't find anything hard to defend about my argument.
Of course. Because "I need more than 8 GB of RAM" and "the government should subsidize Microsoft" are logically equivalent statements. If you're insane.
You have provoked yourself into a rage, because you didn't understand the arguments being made here. Whether you have the capacity to understand, I don't know. But I'll assume so, and explain:
1. Nobody has argued either for or against for the government to subsidize Microsoft.
2. However people are arguing that the government should subsidize RAM makers. And their extended argument for it is that people choose to purchase Microsoft Windows. A well known low quality product.
That's a very weak argument. If you choose to make bad decisions and use sub-par solutions when good solutions are readily available, that isn't something that the government should take to account when deciding on subsidies.
Could you explain how you chose the 8GB number? Would you still have used 8GB as the baseline 10 years ago? 20? 30? No? Then at what point would say 16GB will become acceptable?
What reasoning are you using to figure out what exact amount of RAM is acceptable at a given time period?
Hello!? Us consumers need it too.
Really depends how you're defining "need".
Will you die without it?
Will your life as you know it end if your next phone has 8GB instead of 12GB?
Those questions can’t get to the root of how the GGP is using the word for two reasons: 1. Companies, not being alive, can’t possibly need in this sense without analogy. 2. The poster included electricity as something everybody needs but there are people surviving without it.
You need 8GB at the very most, which isn't out of reach for any consumer. If you need more, then you are a professional and can afford more. Or you are a hobbyist, and your hobby doesn't deserve any subsidies from the hard working tax payer.
> You need 8GB at the very most
I call total bullshit on that. I need more.
Why would you need more compute power than other people? An 8GB machine is already way beyond capable for anything a consumer - and most professional - can throw at it.
And why in the world should the government force tax payers to subsidize your extraordinary computing needs?
Have you ever used a computer?
No, have you?
In addition of being expensive, it's also a gamble. If it wasn't risky then I don't think financing it would be that big of an issue.
> But we don't live in an absolutely free market, the memory manufacturers have refused to budge on creating new fabs until recently out of caution
...that's how the market works. This is called capital discipline and is standard across industries. If the increase in demand is temporary, they would be spending 5-10 years of profit on facilities which will only lose money. The same works in non-capitalist economies too, you don't want to waste resources on stranded assets and many still believe this RAM demand is temporary and the house of cards will collapse.
FWIW I have a friend who makes hardware and it's extremely difficult for him to secure memory at any price. Because AI people have locked up so much of the upcoming supply it's often not a question of how much it will cost but of finding someone who will sell you memory at all. If this were price gouging I would expect he'd have no trouble finding allocations at exorbitant prices but that doesn't seem to be the case.
same here, even when we went to larger hardware producers, they were basically we would sell you, but we cant actually get any RAM at all either.
There are never shortages in a free market, just higher prices.
I have some bad news about how closely Econ 101 resembles the real world…
[flagged]
At a longer timescale than very temporary supply shocks, I believe what I said is correct. The basic point is that if supply contracts, demand must contract as well, which in a free market happens via price increases and in an unfree market happens via shortages (i.e., the inability of buyers to obtain a good at a market-clearing price). If high barriers to entry are relevant in this context (obviously they very much exist in the memory industry!), then please enlighten me how, if you are capable of doing so without unconstructive snark.
But currently we are in a very temporary supply shock, and its called a shortage.
> in a free market happens via price increases and in an unfree market happens via shortages
Since people have earmarked purchases in advance the market isn't free, I'm not allowed to buy the ram another person already bought. Its the same how the texas electricity market works, prices reach infinity since there are more consumers that have already bought electricity than there is electricity left to provide, and you get a shortage.
Free markets with contracts aren't free since contracts creates constraints that results in shortages like this.
Great explanation, thanks.
There are only 3 manufacturers (about to be 4 soon) total. This is a demand crunch the likes of which the market hasn't seen before, prices are doing the natural thing.
You make it sound like the crunch is caused by manufacturers, my understanding is that it is caused by the abnormal demand. I am not aware that RAM production is at a historically low level.
It's actually both. DRAM manufactures operate on a boom bust cycle and by 2023 there was significant oversupply. This meant even leading up to that they had already decided not to invest in further capacity explicitly to recover profit on their existing lines.
So a lack of builds coupled with AI boom at the same time created a perfect storm with no new capacity coming online for years after demand rapidly escalated.
Demand crunch comes from the customer. All HBM and DDR5 wafer capacity is sold out through 2027 and all three vendors are booking 2028 capacity.
This was not created by the manufacturers.
The manufacturers are not doing a whole lot to fix the problem l and I don’t blame them. It’s expensive to build out and they know demand will slow down and don’t want to be stuck with excess capacity. Better to be sold out and high prices.
Samsung said they were “moderating demand” which doesn’t sound like they want this problem to be solved either.
If you are talking CXMT Apple recently learned the hard way that even with price fixed RAM from the big 3, CXMT is struggling to provide cheaper prices as they are using much older design/manufacturing systems.
https://www.msn.com/en-xl/money/general/apple-s-cxmt-memory-...
> There are only 3 manufacturers (about to be 4 soon) total.
...3 manufacturers with a long history of price fixing...
https://en.wikipedia.org/wiki/DRAM_industry_price_fixing
Uhm, your link states that they were price fixing over 20 years ago, and were in court for almost a decade. There was another accusation in 2018, but it was dismissed.
That's not a long history of pricing fixing, just accusations.
The difference between 2002 and now is that it is so much easier to follow each other's leads on pricing without explicitly communicating about it and leaving a paper trail. But they never stopped acting indistinguishable from a cartel.
who is the 4th?
Intel is also looking into returning to making RAM, but it's probably the Chinese that will enter the market faster
Probaby CXMT
What about YMTC?
Tiny market share compared to top 4, but also looking to expand into HBM in collaboration with CXMT
https://chipletsummit.com/proceeding_files/a0qVV0000032cp3/2...
I remember the cost of RAM in the 80s and 90s. If I remember correctly my mom paid ~$1000 for 32mb of RAM in 94/95. Something or other like that.
While they have spiked this year, the trend is towards more RAM for cheaper.
https://ourworldindata.org/grapher/historical-cost-of-comput...
Micron for example discontinued their consumer business because of the demand from enterprise business.
>and how much is this just manufacturers taking advantage of it
"Manufacturers taking advantage of it" wouldn't take it to 500% without AI, because it's way beyond the price elasticity points where they make up lost sales at higher prices.
Those prices lose most consumer sales, and are only viable if there's constant high demand that's relatively insensitive to price (so, not regular folks, but e.g. AI companies).
To take advantage they'd have to do some form of price fixing which in most places is a very big footgun.
I honestly don't think they need to fix their prices, and it's possible that selling to consumers is a secondary market for them compared to datacenters, like Nvidia selling GPUs to gamers. We know that's not where the money is.
Seeing some companies actually exit the consumer market is evidence for this.
> To take advantage they'd have to do some form of price fixing which in most places is a very big footgun.
When parent commenter says "They have done this in the past after all" that's exactly what they're talking about. Micron, Samsung, Hyinx (prior to SK Group acquisition), and others all participated in RAM price fixing in the early 2000's.
https://en.wikipedia.org/wiki/DRAM_industry_price_fixing
If it's a highly profitable footgun why not?
https://en.wikipedia.org/wiki/DRAM_industry_price_fixing#202...
If something is illegal, and the punishment is a fine, it's not illegal it just has a cost.
Is it worth the cost? Yup.
Just to be clear the big Memory giants are famous for their price fixing historically speaking.
It could be the economically rational thing to do if they think there is a bubble
Forming a cartel to rig prices is economically viable with or without a bubble.
Same thing happened with eggs in the USA with bird flu used as an excuse.
Which reversed in about 6 months and now eggs are are their lowest prices in nearly a decade.
The average hatching egg costs a few dollars. Can go as high as nearly $20 if it's something remarkably rare and useful in terms of breed/quality.
An egg takes 21 days to hatch.
Then it's 18-24 weeks to start laying their own eggs. That's ~$5-15 in feed over the period, maybe 40-50 pounds. There's utilities, labor, land, etc., but your incremental cost to add a hen if you have the space is minimal. My numbers won't reflect a lot of operators, as scale matters.
Out of a typical 1.5-3 year lifespan that had a startup cost of a few dollars per, paying back its return by about month 6-7, that's not too bad. It's only worth doing financially in this economy at ludicrous, arguably inherently inhumane scales, but that's what it is on paper.
Not quite the same path to course-correction as a silicon shortage.
The same thing will happen with RAM. A glut will come eventually. It might take a year, maybe two.
Isn't next year's capacity already sold out? Is there any reason to hope for significant additional capacity to come online during that term?
Eventually 2028 and 2029 will come around and at some point today's data center builds will be (economically) full.
In addition to that there are extra fabs build and (for example) China can use the high margin to catch up with their processes without breaking the bank.
We can dream. I would do unspeakable things with 512GB of RAM.
I had 512 in my desktop a decade ago
thank goodness the adults are in charge now
What is also interesting is no one predicted this in 2022-2025 as LLMs were gaining widespread use. It just suddenly happened--"boom!" memory shortage in mid 2025 to early 2026,and it caught everyone by surprise. It was understood that the biggest bottleneck was "compute", not the memory. Even experts, such as popular AI bloggers, industry leaders, and top podcasters failed to anticipate this. No one on twitter either. I spend a lot of time online and I don't recall, afik, anyone before 2026 saying to buy memory stocks.
Sam Altman laying claims on 40% of the world’s RAM wafer production is what took everyone by surprise. The supply shortage likely would have ramped up more slowly otherwise.
More importantly that did take subterfuge: there's no way a contract that large would've been agreed to by both manufacturers if they knew the other was signing it.
The actual deals were hours apart and secret for that reason.
It's hilarious watching the conspiracy theorists do their usual thing when there was an actual conspiracy, it was perpetuated by the inner circle of a US company and announced and publicized in full once it had achieved its goal.
I think Leopold of "situational awareness" fame predicted this, tried to make it widely known on podcasts and x, and still started billions worth of hedge funding based on this.
> I think Leopold of "situational awareness" fame predicted this, tried to make it widely known on podcasts and x, and still started billions worth of hedge funding based on this.
Ironically enough it's because of its bet on RAM companies that all of Situational Awareness' publicly trading stocks got bought by the highest bidder (Citadel) when the (leveraged) shit hit the proverbial fan and Situational Awareness created a hole of tens of billions.
Situational Awareness was not forced to sell its share in Anthropic for, by contract, investors in Situational Awareness have no right to force the sale (before this month of September I think) of shares in companies not publicly trading yet.
Citadel bought those RAM stocks right at the end of July, during the dip. Amazing timing for for example MU bounced back from $746 back to above $1000 (now around $950).
Now Situational Awareness may still do exceptionally well: they've got $5 billion or so left in Anthropic AIUI.
Everyone in cloud knew this was going to happen, the contracts for all parts of server stuff were measured in years.
Demand and manufacturers not increasing capacity.
Eventually someone will manufacture more to undercut
The problem is the scaling cycles for demand and supply are at different timescales. It takes a lot longer to build capacity than these bubbles take to form and pop and there is a SIGNIFICANT history of silicon manufactures scaling during a bubble then going bankrupt because their increased capacity never had the chance to make them any money.
Taking advantage of Trump admin, AI hype, and setting themselves up as "too big to fail" for bailouts when demand eventually collapses
See the tariff thing where Bessent bought rights to refunds
It's all a really low tech con perpetuated by weak elders who need blue pills to simulate virility. Odd how they complain about the kids virtual reality.
Oh no it's not either; it's intentional petty bullying
What’s dumb is even older tech has gone up. Look at microsd and usb thumb drives. There’s a lot of price gouging just because they can. There’s a real shortage and then there’s just “it’s memory so price go up”
Tried to buy old DDR4 memory recently? I've been revitalizing some old laptops, and I foolishly paid nearly $200 for a Crucial 32GB Kit of DDR4 PC4-19200.
I bought 32GB of DDR4 SODIMMs (16GBx2 Corsair 3200MHz) April last year for $50.
128GB 3600Mhz… $600 just last week.
I've just compared prices for an SSD and a rotary disk drive I bought both about 2 years ago. The same SSD is now 3x as expensive and the rotary drive is twice as expensive.
This is insane & needs to stop. Also those responsible should be punished for undermining modern society (which is heavily dependent on IT equipment) in such a manner for personal gain.
If this really happened, then someone is clearly benefiting from it. Basically, it’s speculation and an artificial shortage. But that’s just how the game works.
So... blame capitalism?
Capitalism gave you cheap hardware and then took it away.
PS: Don't worry, memory is a boom/bust business.
I was at the Marietta Microcenter, yesterday, and truly experienced sticker shock on the DDR5 aisle. hlyfcknsht I'll just stay with 32gb RAM on my main machine (6000MT was a few hundred dollars I don't even remember because it was not feasible).
So I went to the DDR4 aisle (similarly: wow!), and realized that I'm sitting on a gold-mine of reclaimed DDR4, from several 8th-10th-gen Intels (probably a few hundred gbs, total, mostly as 8gb modules).
Incredibly, this was all e-waste (given to me "if I took it all"). Eureka.
If anybody wants any DDR2/DDR3, I'm in SE Tennessee (and have even more of that... not worth too much, pay for shipping).
----
Anybody shopping for RTX 5050 or 5060 (for new OEM, 8GB both, non-Ti), WalMart has the best prices (probably hasn't recalibrated recently?), beating Microcenter by $85 and $115, respectively. I ended up getting a 5070Ti from MicroCenter (for a great price; my local WalMart does not carry this xx70Ti version), so the trip wasn't wasted.
After RAM prices climbed 200% between August and November of last year, a store associate at Micro Center encouraged me to buy some RAM sticks if I needed them because "prices might continue to rise for a few years". I told him that was ridiculous and nobody would tolerate it if prices got higher. How wrong I was...
In short, this is how life is designed. Nobody can ever predict with a 100% accuracy what will come in the future.
If Laszlo Hanyecz knew that, after 10 years, 1 BTC will exceed $50K, he wouldn't trade his 10,000 Bitcoins (~$41) for two pizzas!
I was facing a similar situation when I upgraded my system about a year and a half ago. I wanted to upgrade from 32GB RAM to 64GB. The kit (DDR5 6000) was $179. I knew that I might want to add an additional 64GB later on, so I had to decide if I should wait for prices to drop further.
I decided to get the extra RAM. Boy am I glad now. If I thought this was just a temporary spike, I would be tempted to sell the extra memory (I don't really need it yet), but I don't see prices falling anytime soon.
I built a new main rig in November because I saw the memory prices going up. I overpaid about AUD$110 above its RRP. The same kit is now AUD$610 above RRP from the same retailer.
Here's hoping I can wait out the current shortage before any upgrades (I typically only upgrade every 5-7 years or so anyway) or replacements, because I do not plan to pay the new price.
If it makes you feel better. The US government looks like it might start removing some of the 28k troops that are stationed in South Korea over this. No reason to keep them in a country that enables this sort of behavior.
Do you have a source for this being the reason they might remove some troops?
An admin upset about inflation driven negative economic sentiment and a country that is slow rolling a couple hundred billion in investments. Current events.
What else do you know about the inner workings of the current administration… and seriously reflect on why your answer isn’t “lol what? Absolutely nothing!”
You can stick your head up a cows ass but I think I’ll just take the butchers word.
Probably for the best. No hosting nation of imperialism has ever benefited and now we see what the US is willing to throw away because their corporations are upset.
If there is something Korea doesn’t lack, it’s manpower.
The US troops are more likely there as a political shield. Indiscriminate bombing would kill them too, hence drag the US into war so any enemies of S Korea have to act judiciously.
If you think the US public has any appetite for more wars you're straight up delusional. Iran War is going to cost the GOP the house + senate, something that wasn't really a sure then say even 7 months ago.
If anything more foreign bases will be closed in the upcoming years to reign in the reckless spending + endangerment by the DoW. It's not only politically easy, foreign bases don't have representation in Congress, but it's a great to sell alternative governing visions (less on foreign imperialism vs more domestic spending).
> No hosting nation of imperialism has ever benefited
That's why Japan and Germany are some of the worst places to live in the world. Oh wait, I had that quality of life chart upside down.
As opposed to the behavior the US enables?
I guess it's high time to stop using memory inefficient programming languages for any serious work.
Been long past that time, probably 15-20 years past. It's not even the languages its the 500 different layers of abstraction and abominations like Electron.
It turns out supply and demand is a thing, and can go both ways. Remember that in 2001, memory prices dropped like a rock, from $300 to $80. A ton of memory capacity came online at the same time, the market was flooded, so prices sank. The loss of revenue then hurt the memory industry.
I’m even more confused by the economics of the frontier model businesses. Even if they got preferential prices, they are still paying an inordinate amount for their infrastructure.
If it wasn’t economically feasible without VC/Nvidia money 3 years ago, how is it possibly economically feasible now at 10x prices for things like memory?
OpenAI (Stargate) kicked off the price increases a year ago by locking up 40% of the global supply of RAM output.
https://www.tomshardware.com/pc-components/dram/openais-star...
And watch … $5 says they have nothing to do with the ram because even if there was enough power there’s no use case for this much spend
If I had to guess they’re just buying up ram to keep others from having it same way meta, Google, etc hired up engineers to keep others from having them
>even if there was enough power there’s no use case for this much spend
According to them, demand for tokens already exceeds supply. Coding agents are very popular but burn through millions of tokens per hour.
The reasoning models that have powered breakthroughs in AI for math, coding, etc require substantially more compute than base LLMs.
Let's see Paul Allens IPO filling
Also imagine people having their own cheap 2TB rigs at home for their own usage/inference, or a multitude of small providers serving/sharing capacity thanks to cheap available hardware.
But we cant have that, as the big guys are hoarding everything for themselves, so that you have to pay them instead.
I don’t think there were some grand conspiracy to buy up the world’s ram to keep other people out, it is probably more likely that they just really thought they would need it.
However, I do feel like this is the type of conspiracy that unfiltered Claude would come up with.
I’ve been wondering this if you buy up all the memory you can ensure the competition cannot compete with huge investment raising the bar for competition. Of course this all sounds all conspiracy theory. Though without a place to power this or put it makes me wonder this.
It also would be a huge win for the memory companies to agree to this, considering they will make more money with less supply. Seems interesting to me.
next they will become successful business by reselling RAM
It's possible to be economically feasible. It requires the ability to pay off the capex. Not only do they have to pay off their loans in record time (prob 2-3 years), they also have to keep spending every 3 years because the failure rate of GPUs is something like 20%. In addition they need to complete building out the stuff they've started, which means not failing to acquire land, energy, and water (the most dangerously rare and absolutely necessary resource for AI), not dealing with collective bargaining, nor any increased shortages or price hikes in materials. So there is a lot of risk.
Ballpark that they need to make around 110 billion a year, each, to break even on these investments. Let's estimate 550 billion a year in necessary profit required for the major frontier companies. That means there needs to be 550 billion of money, available to customers today, that isn't being spent on anything else, that they will now spend on AI. Maybe some of that comes from increased value, efficiency, or layoffs. But 550 billion is not a small amount of money.
Spread over the whole globe, the cash is there. But it is a hunt for cash, combined with a battle to successfully complete their buildouts, keep them running, and make bank, before the bookie comes knocking.
The railroad panics of the 19th century (and subsequent depressions) happened because they over-leveraged private capital without the ability to profit from it quick enough. So it really is a question of 1) can they really build it all, and 2) will people really pay for it all. If either answer is No, we are looking at economic catastrophe.
It's pulling the ladder up behind them. OpenAI bought up more co tracts for wafers than they could even use. But now that means everyone else has to fight over what's left. This makes it too expensive to compete with OpenAI. No upstart can really afford to build infrastructure.
There is a reason NVIDIA and the hyperscalers are injecting so much money into them
I've got a dead older GPU (RTX titan, 24GB of GDDR6). It's been a paper weight for a while, core is damaged, but I realize I could probably sell it or just the memory chips for a decent price on eBay.
Still, it's a nice reminder of my hubris.
I've inadvertently stocked piled a bunch of DDR4 memory over the years. Usually bought them on an impulse when I saw a good deal, hoping to use it later when someone in the family/friend wanted a new gaming PC or for my own homelab projects. I'm happy with just using DDR4 at this point. The motherboards/CPUs are still relatively cheap either new or on secondary markets. With a decent GPU, the performance difference between DDR4 vs DDR5 is marginal. Most people won't even notice it. My point is, if cost is a factor for a new build, then stick with older hardware. Especially if your use case is gaming or home server.
Intel should probably restart their Optane and also get back in RAM.
I have written a llm compression quantization library called glq because of the high ram prices. Glq uses qtip Trellis quantization which are efficient at low bpw 2-4 bits. As a gamer and ai developer I want to squeeze more out of the same hardware. Glq runs with vLLM.
Open source https://github.com/cnygaard/glq
Pypi glq https://pypi.org/project/glq/
This is "the memory shortage that no one saw coming." To the best of my knowledge, no podcasters, bloggers, or experts in AI predicted this or were sounding the alarm of a memory shortage. No one was saying pre-2025 to buy memory/data storage stocks. This goes to show how hard it is to predict. Everyone was saying to buy NVDA, which had done well from 2022-2024, and then stalled out, only to be surpassed by memory stocks like SanDisk, Western Digital, and Micron, which took many by surprise.
There were some. Read SemiAnalysis.
I picked up some Micron last year anticipating memory demand, but then the stock went down for a while. I waited, it went up and I took 25% profit as I figured they couldn't add any new capacity and would have sold all the capacity from their existing fabs on long term contracts. Then prices went crazy, sigh.
I'm not sure why prices have gone boom now rather than before given we're years into the LLM boom, but probably it's to do with capacity lockups expiring or something similar.
the memory kit i bought earlier this year has been within ~10% of the price for the past 8 months. do with that as you may. a simple ddr5 stick of 6000 mt/s cl30 if wondering.
ever since i had my own money, gpu or general hardware had some excuse to fleece us. however, with ram the blatant collusion throughout history does not invite confidence.
this looks as dramatic as it does because the recent trend happened right after a major crash in prices. throughout the years, the usual price for an 8gb stick would hover around 50-80 usd (albeit laptop one, might be lower for desktop). ddr5 also started being 50-100% more expensive than the incumbent when it first dropped. it took a better part of a year to be in a similar price bracket as ddr4.
all this to say, this situation sucks but those of us with 64+ gigs, bought during the last crash, are eating good but got used to using much more memory than they really need.
while it is insane to pay a decent cpu price for 16 gigs today, getting a "good enough" computer still makes more economic sense than a max plan from a frontier lab.
I still hope all this will come crashing down, there will be an oversupply of components and RAM will be cheaper than paper.
While I share your optimism, I still don't fully understand why a crash will necessarily lead to an oversupply and lower prices for regular consumers. The components / RAM you are referencing are for enterprise use / data centers which in my limited understanding is not a 1:1 swap for the kind of components that your smartphone / car / fridge / gaming rig use. I would love for someone more informed than me to spell out the rationale here.
It seems the chokehold started from the very beginning of the supply chain (wafers), so if a crash were to occur I don't see how we wouldn't be doubly screwed. Lots of worthless enterprise gear being shopped on eBay + a scarcity of consumer grade components leading to sustained high prices.
"I've seen gluts not followed by shortages, but I've never seen a shortage not followed by a glut." - NNTaleb
My understanding is that it's fundamentally a fab capacity limit. At a given node size/technology (think 7nm, etc) they can only output so many wafers. If AI data centers are generating enormous demand for memory, then that's what those wafers are going to be used to for, unless they can be outcompeted, hence higher prices.
I'm not actually sure if enterprise designated memory is equivalent or not to consumer designated. Could be they are the same just packaged differently and/or with better binning. Either way, it's the same supply/demand dynamic at work.
Typically (not always!) server memory is 'registered' so won't work with consumer CPUs which run on un-registered.
Does 'registration' introduce anything physically different? Or could they be completely interoperable?
Yeah a RCD on the memory stick, it's PCB is designed around it. The actual DRAM chips themselves are not different though, which is why I mention it would be super easy to produce more consumer unregistered and less server registered at assembly time :)
Thanks for the clarification, that makes a lot of sense!
Memory fabs make a mix of DDR/LPDDR/GDDR/HBM and they continuously adjust the mix to market demand. If demand for HBM drops they will start making more consumer RAM.
The way I understand it is, these fabs have massive overheads so not selling is not an option. It is also relatively easy to pivot production from enterprise to consumer. Now if they all start doing that at the same time to stay in business during the down cycle, an oversupply in consumer chips is possible.
Datacentre NVMe disk is still usable in a standard desktop. Sever memory less so, but essentially if there were no 'new' server market whatsoever then you could redirect 100% of memory production to consumer chips without enormous upheaval.
By the way I personally think this scenario is very unlikely to happen.
> If there were no 'new' server market whatsoever then you could redirect 100% of memory production to consumer chips without enormous upheaval.
> By the way I personally think this scenario is very unlikely to happen.
Can you explain why? I imagine pivoting to a much less lucrative consumer market when the time comes would lead to financial ruin?
We are starting from a frankly completely unimaginable hypothetical that there's 0 new "datacentre demand".
In that scenario producing consumer chips is still profitable, much less profitable but still much better than giving up and going home.
You will also unlock latent demand by the mere fact that consumer prices are reasonable again. Those who may have postponed a new MacBook or gaming PC will likely buy as soon as lower prices return.
> you could redirect 100% of memory production to consumer chips without enormous upheaval
They would never do that. By then the market has changed and they would produce at the limit of expected demand, certainly not intentionally causing oversupply.
Sounds to me like we'll all be using enterprise servers as home computers! Hell yeah!
The net effect of a crash would be far worse than the benefit of getting a bit of a discount on RAM and storage.
This is why the companies should be aiming for a soft landing to this fad (if they know that they can't make the ROI that makes the industry worthwhile).
I bought an RX 6600 for $450 during the GPU shortage a few years ago. I know the AI boom is different than the crypto/COVID boom, but still. Don't be me.
I got a 3060 ti for like $900 at that time <melting-face-emoji>
My PC built February 2025, with a 5090 and 64GB of ram, is literally worth more than double than what I paid for it. I'm actually concerned what my home insurance would do if it was stolen, since the cost of replacement is so much higher than my invoices say.
Even with AI pressure I'm willing to bet there's no need for prices to be this high. There has to be some price gouging/fixing going on. Same thing for nand and GPUs. Or else they're doing it to try to pressure people not to use local compute instead of the "cloud"
Memory manufacturers have gone through several cycles of lawsuits for market collusion at this point. It's gone from an open secret to just outright flaunting in our face.
Until the penalties involve actual jail time or a fine 10x larger than the profit they made from doing it, they couldn't care less.
While I regret not getting 64GB, I do not regret getting an additional 32GB stick for my PC, for a total of 48GB before the ram crisis kicked off. That's probably in the range of $600+ today for just $120.
With such profit margins, any new players entering the market will want those. This means prices are unlikely going to go down unless demand for memory vanishes, even if new suppliers come to the table.
I already decided I won’t be upgrading any hardware for a few years and I’m going to ride this out. As right now we’re seeing that all prices of everything electronics is going up, my biggest worry is that once memory prices go down, the rest of the electronics will stay high. Eg after the price hike of all Apple electronics, will they go down again in prices in a few years after RAM goes down?
If history can predict anything it’s that the price hikes are permanent.
> If history can predict anything it’s that the price hikes are permanent.
There was a massive spike of RAM prices from 2016 to 2018 followed by a price crash in 2019...
But if we learn anything from big corporations, they'll keep the difference as profits. IOW, Apple won't be lowering their prices.
Nope. Apple wants to maximize profits. They clearly make more at lower prices. Source: the fact that they had lower prices until they were forced to increase them. They will be hoping to return to a more optimal price point as soon as they can.
And when they don't, copium will suggest that it is due to market conditions.
You will not believe the amounts of ram I will buy once the whole thing crashes. Like my raspberry pi clone will be running on a terabyte of ram.
Ram, hdd, ssd, even blu-ray writers are expensive. I purchased mdisc blu-ray writer for $70 in 2021, I see people are trying to sell the same device for $500+ wtf
Is it a Pioneer model? They stopped making their blu-ray drives which are at least perceived to be better quality than the competition, so NOS prices have gone way up.
I'm really happy that I bought 32G of RAM and a 2T M.2 SSD for my desktop two years ago. In retrospect, of course, I should've bought 64G and 4T.
Really happy I got a new PC last year with 128 GB of RAM, 4 TB SSD, and an RTX 5090. This would have been prohibitively expensive now.
Congrats. Happy for you. Nice.
Yeah, I built mine in 3090 times and told myself I'd upgrade every 2-3 GPU generations. I do have 128G ram and I'm thankful for that but I'm definitely not on the same upgrade schedule any more.
I have 32GB and an aging 5800X3D and I have not had any issues running games yet. Honestly would love to know what games need more powerful RAM/CPU outside of stuff like Flight Simulator, or running at 160fps without framegen.
Honestly I have a steam deck and that has gotten me accustomed to not sweating the lower frame rates. That and the ridiculous number of great indie games that aren't heavy on the graphics, and I'm not too worried about upgrades as much as I am worried about replacements if/when the time comes.
Are we observing corresponding work to make things use less memory though?
I've switched a lot of things to WASM and use way less RAM.
or rather, get more done with the same RAM
But what about Apple, with their memory-carrying SOC, one would think they are in a very cool position?
They still need to buy said RAM from all the same players.
What do I read to quickly get a grasp of how economics works, from a high level perspective?
200 pages https://www.goodreads.com/en/book/show/3028.Economics_in_One... There are free PDFs if you don't want to buy it.
Seems solid rec. For lazy person, I think a couple of things that will give you quick intuition for present affairs are:
* The Fifth Discipline - read the beer game chapter and AI summaries for the rest, helps build intuition for chained systems
* Chip War - a quick read but AI summaries will suffice, helps build knowledge for semiconductor industry orientation
* The Goal - classic business book on just-in-time, useful to overturn standard human reactions. read if you're a quick reader, AI summary otherwise
Nice thanks
I cross my fingers that my 10 years old laptop and desktop will hold until the crazyness ends
Ditto; my laptop is slowly losing ports on the right side, but it has 32GB of ECC ram...
same. still rockin a 2016 era dell xps 13. I was about to pull the trigger on a few laptops but I always psych myself out because I assume prices will drop and I can pick them up secondhand. big mistake...
This is more concerning than anything else happening in the AI space.
We had last bought 24TB+ (4x 24 racks, and some miscellaneous) of DDR5 4800 and 6400 last year while finishing a move off of AWS entirely. I tried to be responsible and prioritize a cold and warm spares of all our equipment and over allocation to ensure proper redundancy capacity. At the time it was a hard pill to swallow looking at our budget; but a few months ago we've started to offload some DIMMs on various channels. What went from a cost analysis between a cloud provider, to having to manage transit, circuits, equipment, a IP blocks, etc etc and all the operational obligations that come with it.. it ultimately covering our hosting costs entirely for the next ~18months *at least* (while "only" coming at the cost of having slightly less cold spares sitting in the bottom of our cages).
Given we're all 9004/9005 now, and that (from what I understand) Venice will support 6400. Will hopefully buy back in a glut comes ready for the next bubble. It's been wild, I even posted a small 32GB DDR5-4800 that was not getting recognized for $100 on eBay and it was literally purchased within ~15min of being live.
I'm waiting for China to start selling cheaper memory. They can have my money. The greed of US billionaires need to be lanced with a hefty kick in their wallets.
AI Datacenter proponents do not get to complain about this.
Well this seems sustainable.
No hope in US but I hope EU regulators are taking note of this working on fix as we talk.
Time for memristors?
I noticed that resellers on ebay of used server equipment that I keep an eye on in Germany pulled all their DDR4 from surplus and used server racks and now list it separately for sale, at jacked up prices. Where you used to be able to see a used Gigabyte server listed with "128GB" "256GB" etc listed in the title, they now all say "0GB". They took extra effort, to pull used ram from older servers, and list it separate, just out of the greed of potentially capitalizing on this market.
The problem is... now for anyone looking to pickup a slightly older used/surplus server rack - you have the go and buy the (now marked up) RAM separate.
So... did they list the "0GB" racks with removed ram cheaper you ask? Nope. Many of them seem to be basically the same price as they were or even higher than they were, except now you're paying for RAM separate.
I bought a surplus Gigabyte server about 2 years ago off ebay for a home project, and I just went and the same seller still has some for sale, even older now than they were when I grabbed one, but they are 2x the price, and they don't come with any RAM, whereas I got one already populated as they were from the factory. Once you add that back in literally looking at 150-200% markup in the last 2 years on what was just surplus computer equipment that is 5+ years old.
Why would I voluntarily choose to sell the hardware I have for less than the going price? The odds of it getting picked up by a scalper and resold are extremely high. If I was sitting on a goldmine of old server hardware, I would absolutely be ripping out the RAM and selling it per stick in this market.
> So... did they list the "0GB" racks with removed ram cheaper you ask?
Could this be an indication that there's a supply shortage/price increase of the remaining components in the rack? Or do you interpret it as people just trying their luck out of sheer greed?
Greed? Resellers aren't charities, good for them if the value of their stock goes up. Besides, people buying server equipment aren't exactly starving, I bet most of them can either simply not buy the server, or cut some other expense.
The fake economics of the AI sector are damaging the real economy. This is madness.
Deep down it's all just people telling stories. If the AI sector's story is setting the price, and yours is not, then theirs is the real real one, not yours.
It's a totally broken way to practice economics, but alas, it's ours until we change it.
I hate that I am no longer certain I don’t need to buy a new machine just in case.
[dead]
[dead]
[dead]
[flagged]
Not every single post of yours has to be a link to your "blog".
[flagged]
No, this is not good, in any way, except for AI shareholders (of which I assume you are one, else not rational to support this position).
Paperclipping here, greedily optimizing for shareholder value above all else. Please stop.
Some people really so need the /s huh
To be fair, the AI boosters are saying the exact same thing but mean it seriously. There is no way to distinguish without a tone indicator
> There is no way to distinguish without a tone indicator
And yet many of us aren't having any trouble distinguishing.
Given the number of upvotes my comment got and the other comments where people missed the sarcasm, I wouldn’t be too sure that many people have no trouble distinguishing
"sarcasm" is a cop-out.
The writer can simply say "it was a joke!", cuz we all know someone who uses this pattern regularly when they are trying to squeeze in a non-popular but highly lucrative (to the messenger) position to test the waters.
Either it lands (in which case, it was always meant to be taken literally), or it doesn't land (and can always be walked back as "it was meant sarcastically" or "can't you take a joke?"
The cost proposition to sending this message payload is low, since the messenger can always wiggle their way out (therefore, a "cop-out").
Therefore the need to be more explicit (i.e., /s) in a medium with little nuance affordances for tone and the significantly-increasing risk of such non-popular messages being taken literally (particularly by the also significantly-increasing audience of robots, who themselves may not have the capacity for nuance or irony detection).
Tl;Dr
It's a cop-out, and should be stopped.
sadly, Poe's law scales exponentially on HN
If the shareholders are backing outcomes that make life worse for the rest of us, they deserve to lose money.
what the heck
did you drop a '/s'?
> and the shareholders.
Your comment is an illustration of exactly what is wrong with this website. Given his preceding statements, I sincerely do not believe it is sarcasm.
I'm quite sure he's being sarcastic.
"The best good ever is to maximise shareholder profits" is a tired and cliched caricature that some people unfortunately take seriously. But this is too on the nose.
The fact that, on HN, you can't be sure of that, tells you all you need to know.
I’ve been on HN long enough to know that a disturbing number of people do actually believe that line and act accordingly.
I've come to the same conclusion. Some believe that grownups shouldn't care about RAM prices going up.
I have the innate right to take what is said literally. Moreover, his preceding statements attempt to legitimize his concluding statement.
You also have the innate right to be off base, btw.
It's sarcasm, and the main tell is I'm not even bothering with pretense; just serving up the dystopia straight and acting like that's good.
But one of the "fun" things about HN is SV is so fucked up you can't even be sure someone doesn't actually sincerely believe that shit. And most of them are probably dumbasses with nothing to gain.
I appreciate sarcasm when it's clear. Marking it as "/s" prevents this whole line of discussion. As you noted in your last paragraph, it's hard to be sure without a marking.
and FOR WHATTTT