I guess gigawatts is how we roughly measure computing capacity at the datacenter scale? Also saw something similar here:
> Costs and pricing are expressed per “token”, but the published data immediately seems to admit that this is a bad choice of unit because it costs a lot more to output a token than input one. It seems to me that the actual marginal quantity being produced and consumed is “processing power”, which is apparently measured in gigawatt hours these days. In any case, I think more than anything this vindicates my original decision not to get too precise. [...]
I think for the same model wall time is probably a more intuitive metric; at the end of the day what you’re doing is renting GPU time slices.
Large outputs dominate compute time so are more expensive.
IMO input and output token counts are actually still a bad metric since they linearise non linear cost increases and I suspect we’ll see another change in the future where they bucket by context length. XL output contexts may be 20x more expensive instead of 10x.
> I think for the same model wall time is probably a more intuitive metric; at the end of the day what you’re doing is renting GPU time slices
This is a bit too much of a simplification.
The LLM provider batches multiple customer requests into one GPU/TPU pass over the weights, with minimal latency increase.
The LLM provider may in fact be renting GPUs by the second, but the end user isn't. We the end users are essentially timesharing a pool of GPUs without any dedicated "1 vGPU" style resource allocation. In such a setting, charging by "GPU tick" sounds valid, and the various categories of token costs are an approximation of cost+margin.
As a customer, it's nice that I can quantize and count the units of cost in an understandable way.
For Anthropic, as a business bleeding money, it's probably nice to have value-based pricing, for the tokens, so innovation (like computation efficiency improvements) can result in some extra margin. If they exposed the more direct computation cost, they could never financially benefit from any improved efficiency, including faster hardware!
Gigawatts seems like more a statement of the power supply and dissipation of the actual facility.
I’m assuming you can cram more chips in there if you have more efficient chips to make use of spare capacity?
Trying to measure the actual compute is a moving target since you’d be upgrading things over time, whereas the power aspects are probably more fixed by fire code, building size, and utilities.
Measuring data centers in watts is like measuring cars in horsepower. Power isn't a direct measure of performance, but of the primary constraint on performance. When in doubt choose the thermodynamic perspective.
I mean a single nuclear reactor delivers around 1GW, so if a single datacenter consumes multiple of those, it gives a reasonably accurate idea of the scale.
It's not really a stable measure of compute, but it's a good indication of burn rate as energy cost is something we closely track in economies and it actually dominates a lot of the cost of operating data centers. At least short term. Over time we'll get more tokens per energy unit and less dollars for the hardware needed per energy unit. Tokens currently is too abstract for a lot of people. They have no concept of the relation ship of numbers of tokens per time unit and cost. Long term there's going to be a big shift from op-ex to cap-ex for energy usage as we shift from burning methane and coal to using renewables with storage.
That these data centers can turn electricity + a little bit of fairly simple software directly into consumer and business value is pretty much the whole story.
Compare what you need to add to AWS EC2 to get the same result, above and beyond the electricity.
That's a convenient story, but most consumers' and businesses' use of AI is light enough that they could easily run local models on their existing silicon. Resorting to proprietary AI running in the datacenter would only add a tiny fraction of incremental value over that, and at a significant cost.
Sure but where the puck is going is long-running reasoning agents where local models are (for the moment) significantly constrained relative to a Claude Opus 4.6.
All of big tech (except Google obviously) is pushing hard for Claude Code internally. I’m talking “you all have unlimited tokens and we’re going to have a leaderboard of who used the most” kind of push.
"we’re going to have a leaderboard of who used the most"
Yeah I've seen stuff like that and it's a bit bewildering for me. Feels a bit like AWS is new and we're competing to see who can deploy the most EC2 instances.
It’s the crudeness of available management methods at play. Quite exposing for the profession, really (remember lines of code as measure of productivity?).
Their disclosed run rate was 14bn around the time of those filings IIRC, they started showing meaningful revenue around start of 2025, so if you just linearly extrapolate up that would give you ~7bn-ish actual revenue over that period. The more the growth is weighted towards the last few months the more that number goes down
So I don't think those numbers are really in tension at all
If your revenue doubles every month, then in the first month where you make $2.5B, your total lifetime revenue has been $5B ($2.5B this month, $1.25B the month before, etc. is a simple geometric series). But your current revenue run rate for the next year will be $2.5B x 12 = $30B.
They're not quite growing that fast, but there's nothing inherently inconsistent between these claims... as long as the growth curve is crazy.
1) It's in their interest to distort numbers and frame things that make them look good - e.g. using 'run-rate'
2) The numbers are not audited and we have no idea re. the manner in which they are recognising revenue - this can affect the true compounding rate of growth in revenues
The numbers are certainly audited by their investors. Anthropic isn't foreign to PR talk, but investors know what to look for in their book. They aren't stupid unlike how they are viewed on HN.
There are more investment money than Anthropic need. They can pick and choose.
I do, and I do trust the numbers. I doubt Anthropic is pursuing fraud given that they already don't have enough compute to serve demand. What is the point of lying to the public, investors and risk going to jail?
I think you can argue that AI is going to explode and take over the economy, and it’s still a bubble.
I think one possible route is that cloud capacity just becomes totally commoditized and none of the hyperscalers will be able to extract the kinds of profit margins that would allow them to make a good return on their investment (model makers will fall victim to this too). Ultimately, what may happen is that market competition for everything explodes since AI and robots can do all the work, prices for everything (goods, services, assets) collapses, and no one is really any richer than anyone else.
Even if the AI frontier becomes "totally commoditized" it will still be reliant on a scarce factor, namely leading-edge chips. Chipmakers will ultimately capture that value, because competing it away would require expanding the industry and that's a very slow process involving billion-dollar expenses planned far in advance (multiple years, and that lead time can only expand further as the required scale gets even larger).
AI being a bubble it's not mutually exclusive of being a real and useful technology and the existence of non-snake oil companies.
Cisco was a bubble in the dot com crash, despite being a company that provide real value and profit, just not at the level of the crazy expectations from the time.
I’m not sure your numbers are accurate, they raised $13bn in funding in September last year. Also do note that a lot of the money is cross-subsidized by Google who is funding the TPUs as an investment, so I wouldn’t be so confident that they are returning money quite yet (though it does seem that Anthropic might make it).
Interesting to see Anthropic investing in compute infrastructure. The bottleneck I keep hitting is not
raw compute but where that compute lives — EU customers increasingly need guarantees their data stays
in-region. More sovereign compute options in Europe would unlock a lot of enterprise AI adoption.
I think it’s also important to add the context that Broadcom’s CEO, Hock Tan, went on CNBC in October and had a vacuous conversation with Jim Cramer about their OpenAI “deal” at the time [0]. Nothing of substance was said, it was just endless loops about the opportunity of AI. It is now 6 months later and there has been nary a peep from Broadcom about any updates.
I think Anthropic is a more grounded company than OpenAI because Sam Altman is insane, but it is still playing the same game.
And and Broadcom designs a huge part of the chip. They take Google's (mostly) logical design and providing everything TSMC need to physically make the chip (including imports g IP such as serdes, PLLs, and test).
The VMware s/w rental market has no relevance to this deal, any more than the IBM role in data processing in germany in the 1930s had any relevance to their business in Israel in the 60s, or Oracle's failure in the DC market impacts licencing of the database product.
It's just not material. Broadcom make devices they need, and Broadcom want to sell those devices and exclude another VLSI company from selling, so the two have an interest in doing business. That's all there is to it.
About the most you could say is that the lawyers drafting whatever agreement they sign to, will reflect on the contract in regard to future changes of pricing and supply, in the light of what Broadcom did with VMWare licencing costs.
Because all the variables that go into performance / efficiency measurement of a model (processing power, algorithm efficiency, parallelization, etc) boil down to cost per token input and token output. And the tangible cost for a datacenter is power consumed. Of course, amortized capex costs are also part of the game.
It's easy to think about. Google reported a global average power consumption of 3.7GW in 2024, so you can think of this deal as representing an expansion of something like 10-15% of that 2024 baseline, if you assume 50% capacity utilization.
There's at least a decent argument to be made that the limiting factor is actually the physical silicon itself (at least at cutting-edge nodes) not really the power. This actually gives AI labs an incentive to run those specific chips somewhat cooler, because high device temperatures and high input voltages (which you need to push frequencies higher) might severely impact a modern chip's reliability over time.
Power is the limiting layer above physical chips. You can add more chips and them at lower clock or add more efficient chips later on, but you can't really change the power of a data center easily.
It will nonetheless be vastly cheaper to build a new datacenter and arrange for powering it than to fab the amount of leading-edge chips and compute systems that are going to ultimately eat that power. So the chips themselves are still the meaningful constraint.
Surely, there should be some more critical questions posed by why just buying a bunch of GPUs is a good idea? It just feels like a cheap way to show that growth is happening. It feels a bit much like FOMO. It feels like nobody with the capital is questioning whether this is actually a good idea or even a desirable way to improve AI models or even if that is money well spent. 1 GW is a lot of power. My understanding is that it is the equivalent to the instantaneous demand of a city like Seattle. This is absurd.
It feels like there is some awareness that asking for gigawatts if not terrawatts of compute probably needs more justification than has been proffered and the big banks are already trying to CYA themselves by publishing reports saying AI has not contributed meaningfully to the economy like Goldman Sachs recently did.
kinda complicated though when you consider it fully. Power consumption only measures the environmental impact really, we come up with more clever ways to use the same amount of power daily.
it's kind of like an electrical motor that exists before the strong understanding of lorentz/ohm's law. We don't really know how inefficient the thing is because we don't really know where the ceiling is aside from some loosey theoretical computational efficiency concepts that don't strongly apply to practical LLMs.
to be clear, I don't disagree that it's the limiting factor, just that 'limits' is nuanced here between effort/ability and raw power use.
"Do you realize that the human brain has been liken to an electronic brain? Someone said and I don't know whether he is right or not, but he said, if the human brain were put together on the basis of an IBM electronic brain, it would take 7 buildings the size of the Empire State Building to house it, it would take all the water of the Niagara River to cool it, and all of the power generated by the Niagara River to operate it." (Sermon by Paris Reidhead, circa 1950s.[1])
We're there on size and power.
Is there some more efficient way to do this?
Maybe it's just because the specifics on FLOPs are more complicated, especially given how many different floating point formats are floating around in ML. Even NVIDIA has like 6 different FLOPs numbers on their GPUs nowadays.
And you know Nvidia can't be constent with one format for FLOPs within a single graph, 1,000,000x faster but comparing FP32 to FP8 or NVFP4 and acting like it's the same.
Some of it might be market-signaling to the broader energy industry: "hey would you PLEASE build more power plants and power lines? Look at all this money we have, we will pay for it!"
Interesting timing given the quantum computing timeline pressure from this week's cryptography discussions. $30B run-rate and gigawatts of TPU capacity — and meanwhile the most interesting AI work I've seen lately runs on a phone in Termux with no cloud dependency at all. Both things are true simultaneously.
I don’t think Opus and Sonnet are significantly better than Gemini or ChatGPT. Am I missing something?
It looks to me that Anthropic is one or two Gemmas away from a lot of people using Opus for 20% of hard use cases and letting on-device LLM rip through the code base on a Mac Mini or Studio and OpenCode.
Once Claude Code is not the only game in town and Cowork is made redundant by Google pulling their finger out on integration with Workspace, what else is there for Anthropic?
On-device agentic use is orders of magnitude harder than simple chatting (which is still slow for SOTA), it uses up a huge amount of context and tokens on reading code and reasoning through it. It's sort of viable if you just set it to work overnight on some completely vibe-coded stuff, but that has very middling results. Giving feedback to the model interactively is completely out of the question.
Where open models can make a difference for agentic use is with third-party inference at scale, which can actually be fast enough for reasonable workflows.
That’s mostly because current open models are way too heavy and too general for coding. If the model is compact enough to run well on normal Apple Silicon, local stops being an overnight-only toy and you also lose the third-party dependency. That’s the bet behind rig.ai.
How so? Opus and Sonnet are frontier models which cannot easily be replicated. Compute has real physical constraints which require appropriate procurement at this scale. At least those two points seem like pretty strong moats against the majority of companies.
You don't need to "replicate" Opus and Sonnet, you just need to match their overall performance at lower cost. That's been absolutely doable so far, with a steadily decreasing lag time.
That's a fair response. But I'm not aware of any metrics supporting the point that the lag time is decreasing. The discourse I've seen has more focused on the ways Claude/OpenAI/Google have pulled away from the rest of the pack.
To be clear, I accept you might be right, but I think the crux is whether lag time is increasing, steady or growing.
You're right and the your reasoning is great. Anthropic should fold and give up their $30 billion ARR just announced in the OP. Shut it all down, no moat here.
How is compute shortage to satisfy demand manifested? Obviously they never close sign-ups, so only option is to extended queues? But if demand grows like crazy, then queues should get longer, yet my pro claude plan seems snappy with only occasional retries due to 429.
There's no limit to the algorithms. People dont understand yet. They can learn the whole universe with a big enough compute cluster. We built a generalizable learning machine
the question is will we experience resource constraints before we get there? what if the step up to post-scarcity is gated by a compute level just out of our reach?
Edit: What we have built is a natural language interface to existing, textually recorded, information. Transformers cannot learn the whole universe because the universe has not yet been recorded into text.
Transformers operate on images and a variety of sensor data. They can also operate completely on non-textual inputs and outputs. I don't know what the ceiling on their capabilities is, but the complaint that they only operate on text seems just obviously wrong. There are numerous examples but one is meteorological forecasting which takes in a variety of time series sensor inputs and outputs e.g. time-series temperature maps. https://www.nature.com/articles/s41598-025-07897-4
100% agreed. Sadly, lots of people out there with the "trust me bro, just need more compute". Hopefully we don't consume all the planet's resources trying.
I reevaluated my priors long ago when I saw that scaling laws show no sign of stopping, no sign of plateau.
Strangely some people on HN seem to desperately cling to the notion that it's all going to come to a halt. This is unscientific. What evidence do you have - any evidence - that the scaling laws are due to come to an end?
I’d like to see something that indicates models are getting better without the need for more training data. I would expect most gains are coming from more and better labeled data. We’re racing towards a complete encyclopedia of human knowledge. If we get there that’s only a drop in the bucket of all knowable things.
Bro the planet is literally experiencing a climate disaster and you think the solution is to create more systems that are misaligned with the planet's ecosystem for humans?
I guess the great filter is a real thing and not just a thought experiment.
I assure you that voluntary meat consumption because "taste buds go brr" is a much bigger problem than AI that results in actual productivity gains (and potentially solve the very climate crisis you complain about.)
I suspect it's not that people do not see the progress, they fail to fully trust laws not truly backed by physics like the transistor laws. We empirically see that scaling works and continue to work.
The issue people have isn’t some interpretation of scaling laws, it’s whether the planet’s ecology is goi g to be able to sustain this endeavour.
I shouldn’t have to say this out loud, but if the environment collapses, we will die, and no amount of “just a bit more scaling bro, just think of the gains” will matter.
People's voluntary dietary choices cause far more suffering and ecological damage than AI, and for much less return or economic output. But you tell people to switch to plant based foods and they lose their shit.
Well diffusers are trained unsupervised on raw pictures. I don't know how they train multi-modal LLMs on images, but yes obviously they are consuming other media than just text. I don't think, but would be happy to be corrected, that models glean much of their "knowledge" from non-textual training data.
Please tell me more. When I ask an LLM a question, and get a text response, can that response incorporate non-textual information from visual training data?
I guess gigawatts is how we roughly measure computing capacity at the datacenter scale? Also saw something similar here:
> Costs and pricing are expressed per “token”, but the published data immediately seems to admit that this is a bad choice of unit because it costs a lot more to output a token than input one. It seems to me that the actual marginal quantity being produced and consumed is “processing power”, which is apparently measured in gigawatt hours these days. In any case, I think more than anything this vindicates my original decision not to get too precise. [...]
https://backofmind.substack.com/p/new-new-rules-for-the-new-...
Is it priced that way, though? I assume next-gen TPU's will be more efficient?
> but the published data immediately seems to admit that this is a bad choice of unit because it costs a lot more to output a token than input one
And, that's silly, because API pricing is more expensive for output than input tokens, 5x so for Anthropic [1], and 6x so for OpenAI!
[1] https://platform.claude.com/docs/en/about-claude/pricing
[2] https://openai.com/api/pricing
I think for the same model wall time is probably a more intuitive metric; at the end of the day what you’re doing is renting GPU time slices.
Large outputs dominate compute time so are more expensive.
IMO input and output token counts are actually still a bad metric since they linearise non linear cost increases and I suspect we’ll see another change in the future where they bucket by context length. XL output contexts may be 20x more expensive instead of 10x.
> I think for the same model wall time is probably a more intuitive metric; at the end of the day what you’re doing is renting GPU time slices
This is a bit too much of a simplification.
The LLM provider batches multiple customer requests into one GPU/TPU pass over the weights, with minimal latency increase.
The LLM provider may in fact be renting GPUs by the second, but the end user isn't. We the end users are essentially timesharing a pool of GPUs without any dedicated "1 vGPU" style resource allocation. In such a setting, charging by "GPU tick" sounds valid, and the various categories of token costs are an approximation of cost+margin.
As a customer, it's nice that I can quantize and count the units of cost in an understandable way.
For Anthropic, as a business bleeding money, it's probably nice to have value-based pricing, for the tokens, so innovation (like computation efficiency improvements) can result in some extra margin. If they exposed the more direct computation cost, they could never financially benefit from any improved efficiency, including faster hardware!
They already bucket when context goes above 200k
No longer
Gigawatts seems like more a statement of the power supply and dissipation of the actual facility.
I’m assuming you can cram more chips in there if you have more efficient chips to make use of spare capacity?
Trying to measure the actual compute is a moving target since you’d be upgrading things over time, whereas the power aspects are probably more fixed by fire code, building size, and utilities.
Measuring data centers in watts is like measuring cars in horsepower. Power isn't a direct measure of performance, but of the primary constraint on performance. When in doubt choose the thermodynamic perspective.
Gigawatts are units of power, gigawatthours are units of energy.
The equivalent of cars would be pricing by how much gas you burned, not horsepower.
1 horsepower = 745.7 watts
I mean a single nuclear reactor delivers around 1GW, so if a single datacenter consumes multiple of those, it gives a reasonably accurate idea of the scale.
This conversation is confusing because OP didn't use the same units as the person in the quote.
It's not really a stable measure of compute, but it's a good indication of burn rate as energy cost is something we closely track in economies and it actually dominates a lot of the cost of operating data centers. At least short term. Over time we'll get more tokens per energy unit and less dollars for the hardware needed per energy unit. Tokens currently is too abstract for a lot of people. They have no concept of the relation ship of numbers of tokens per time unit and cost. Long term there's going to be a big shift from op-ex to cap-ex for energy usage as we shift from burning methane and coal to using renewables with storage.
We need a Moore's law for tokens, and energy.
That these data centers can turn electricity + a little bit of fairly simple software directly into consumer and business value is pretty much the whole story.
Compare what you need to add to AWS EC2 to get the same result, above and beyond the electricity.
That's a convenient story, but most consumers' and businesses' use of AI is light enough that they could easily run local models on their existing silicon. Resorting to proprietary AI running in the datacenter would only add a tiny fraction of incremental value over that, and at a significant cost.
Sure but where the puck is going is long-running reasoning agents where local models are (for the moment) significantly constrained relative to a Claude Opus 4.6.
I'm looking forward to running a Gemma 4 turboquant on my 24GB GPU. The perf looks impressive for how compact it is.
I often get a 10x more cost effective run processing on my local hardware.
Still reaching for frontier models for coding, but find the hosted models on open router good enough for simple work.
Feels like we are jumping to warp on flops. My cores are throttled and the fiber is lit.
$19B -> $30B annualized revenue in a month?
Feels like the lede is buried here!
All of big tech (except Google obviously) is pushing hard for Claude Code internally. I’m talking “you all have unlimited tokens and we’re going to have a leaderboard of who used the most” kind of push.
"we’re going to have a leaderboard of who used the most"
Yeah I've seen stuff like that and it's a bit bewildering for me. Feels a bit like AWS is new and we're competing to see who can deploy the most EC2 instances.
It’s the crudeness of available management methods at play. Quite exposing for the profession, really (remember lines of code as measure of productivity?).
Also, very very recently they said in a court filing that their lifetime revenue was "at least" 5 billion. Which is it?
Their disclosed run rate was 14bn around the time of those filings IIRC, they started showing meaningful revenue around start of 2025, so if you just linearly extrapolate up that would give you ~7bn-ish actual revenue over that period. The more the growth is weighted towards the last few months the more that number goes down
So I don't think those numbers are really in tension at all
If your revenue doubles every month, then in the first month where you make $2.5B, your total lifetime revenue has been $5B ($2.5B this month, $1.25B the month before, etc. is a simple geometric series). But your current revenue run rate for the next year will be $2.5B x 12 = $30B.
They're not quite growing that fast, but there's nothing inherently inconsistent between these claims... as long as the growth curve is crazy.
The reality is
1) It's in their interest to distort numbers and frame things that make them look good - e.g. using 'run-rate' 2) The numbers are not audited and we have no idea re. the manner in which they are recognising revenue - this can affect the true compounding rate of growth in revenues
The numbers are certainly audited by their investors. Anthropic isn't foreign to PR talk, but investors know what to look for in their book. They aren't stupid unlike how they are viewed on HN.
There are more investment money than Anthropic need. They can pick and choose.
"The numbers are certainly audited by their investors."
Hahaha.
Mate nobody cares about that nor trusts it. Everyone is waiting in anticipation for the S-1 filing.
I do, and I do trust the numbers. I doubt Anthropic is pursuing fraud given that they already don't have enough compute to serve demand. What is the point of lying to the public, investors and risk going to jail?
Money? Bankman-Fried wasn't the only one.
Curious - what’s this court filing?
Too lazy to pull up a url, but it was a filing by Anthropic's CFO in the Anthropic v Department of War case.
Doesn't that beat openai in revenue?
[flagged]
I think you can argue that AI is going to explode and take over the economy, and it’s still a bubble.
I think one possible route is that cloud capacity just becomes totally commoditized and none of the hyperscalers will be able to extract the kinds of profit margins that would allow them to make a good return on their investment (model makers will fall victim to this too). Ultimately, what may happen is that market competition for everything explodes since AI and robots can do all the work, prices for everything (goods, services, assets) collapses, and no one is really any richer than anyone else.
Even if the AI frontier becomes "totally commoditized" it will still be reliant on a scarce factor, namely leading-edge chips. Chipmakers will ultimately capture that value, because competing it away would require expanding the industry and that's a very slow process involving billion-dollar expenses planned far in advance (multiple years, and that lead time can only expand further as the required scale gets even larger).
You don’t think open AI models will eventually be able to design and build chips and fabs and all their components?
Except you're neglecting the fact that LLMs can become more efficient.
The magical thing about software is that efficiency gains can come pretty quickly relative to other industries.
We're already seeing this with Qwen 3.5 and Gemma 4. They're better than GPT-3.5 and they run on smartphones and old laptops.
AI being a bubble it's not mutually exclusive of being a real and useful technology and the existence of non-snake oil companies.
Cisco was a bubble in the dot com crash, despite being a company that provide real value and profit, just not at the level of the crazy expectations from the time.
I'm literally talking about the fact that Anthropic is making $30B Annual Revenue, which is the result of less than $10B investment two years ago.
The public hasn't seen the insane ROA/ROI on GPUs. So all AI adjacent stocks are massively undervalued.
I’m not sure your numbers are accurate, they raised $13bn in funding in September last year. Also do note that a lot of the money is cross-subsidized by Google who is funding the TPUs as an investment, so I wouldn’t be so confident that they are returning money quite yet (though it does seem that Anthropic might make it).
[flagged]
[flagged]
Interesting to see Anthropic investing in compute infrastructure. The bottleneck I keep hitting is not raw compute but where that compute lives — EU customers increasingly need guarantees their data stays in-region. More sovereign compute options in Europe would unlock a lot of enterprise AI adoption.
[flagged]
[dead]
I’m surprised Anthropic wanted to partner with Broadcom when they have such a negative reputation with antics such as their VMWare acquisition.
I think it’s also important to add the context that Broadcom’s CEO, Hock Tan, went on CNBC in October and had a vacuous conversation with Jim Cramer about their OpenAI “deal” at the time [0]. Nothing of substance was said, it was just endless loops about the opportunity of AI. It is now 6 months later and there has been nary a peep from Broadcom about any updates.
I think Anthropic is a more grounded company than OpenAI because Sam Altman is insane, but it is still playing the same game.
[0] https://www.youtube.com/watch?v=pU2HhJ3jCts
Broadcom builds the TPU chip. Google designs it. You can’t avoid partnering with Broadcom if you want TPUs in significant volume .
Broadcom designs it as well [0], though GCP also works on design as well.
[0] - https://www.broadcom.com/products/custom-silicon
TSMC builds the TPU chip. Broadcom does the rest of the electronics (motherboard, networking, etc...)
And and Broadcom designs a huge part of the chip. They take Google's (mostly) logical design and providing everything TSMC need to physically make the chip (including imports g IP such as serdes, PLLs, and test).
Broadcom makes the TPU. If you want TPUs, you are working with Broadcom whether you want to or not.
The VMware s/w rental market has no relevance to this deal, any more than the IBM role in data processing in germany in the 1930s had any relevance to their business in Israel in the 60s, or Oracle's failure in the DC market impacts licencing of the database product.
It's just not material. Broadcom make devices they need, and Broadcom want to sell those devices and exclude another VLSI company from selling, so the two have an interest in doing business. That's all there is to it.
About the most you could say is that the lawyers drafting whatever agreement they sign to, will reflect on the contract in regard to future changes of pricing and supply, in the light of what Broadcom did with VMWare licencing costs.
On a tangential note: It seems the whole theater with the DoD is over for now, am I seeing this right?
Can someone explain why everything is being marketed in terms of power consumption?
Because all the variables that go into performance / efficiency measurement of a model (processing power, algorithm efficiency, parallelization, etc) boil down to cost per token input and token output. And the tangible cost for a datacenter is power consumed. Of course, amortized capex costs are also part of the game.
It's easy to think about. Google reported a global average power consumption of 3.7GW in 2024, so you can think of this deal as representing an expansion of something like 10-15% of that 2024 baseline, if you assume 50% capacity utilization.
Because that’s the limiting factor
There's at least a decent argument to be made that the limiting factor is actually the physical silicon itself (at least at cutting-edge nodes) not really the power. This actually gives AI labs an incentive to run those specific chips somewhat cooler, because high device temperatures and high input voltages (which you need to push frequencies higher) might severely impact a modern chip's reliability over time.
Power is the limiting layer above physical chips. You can add more chips and them at lower clock or add more efficient chips later on, but you can't really change the power of a data center easily.
It will nonetheless be vastly cheaper to build a new datacenter and arrange for powering it than to fab the amount of leading-edge chips and compute systems that are going to ultimately eat that power. So the chips themselves are still the meaningful constraint.
I feel like that’s a bit glib?
Surely, there should be some more critical questions posed by why just buying a bunch of GPUs is a good idea? It just feels like a cheap way to show that growth is happening. It feels a bit much like FOMO. It feels like nobody with the capital is questioning whether this is actually a good idea or even a desirable way to improve AI models or even if that is money well spent. 1 GW is a lot of power. My understanding is that it is the equivalent to the instantaneous demand of a city like Seattle. This is absurd.
It feels like there is some awareness that asking for gigawatts if not terrawatts of compute probably needs more justification than has been proffered and the big banks are already trying to CYA themselves by publishing reports saying AI has not contributed meaningfully to the economy like Goldman Sachs recently did.
kinda complicated though when you consider it fully. Power consumption only measures the environmental impact really, we come up with more clever ways to use the same amount of power daily.
it's kind of like an electrical motor that exists before the strong understanding of lorentz/ohm's law. We don't really know how inefficient the thing is because we don't really know where the ceiling is aside from some loosey theoretical computational efficiency concepts that don't strongly apply to practical LLMs.
to be clear, I don't disagree that it's the limiting factor, just that 'limits' is nuanced here between effort/ability and raw power use.
Somehow we must be doing this wrong.
"Do you realize that the human brain has been liken to an electronic brain? Someone said and I don't know whether he is right or not, but he said, if the human brain were put together on the basis of an IBM electronic brain, it would take 7 buildings the size of the Empire State Building to house it, it would take all the water of the Niagara River to cool it, and all of the power generated by the Niagara River to operate it." (Sermon by Paris Reidhead, circa 1950s.[1])
We're there on size and power. Is there some more efficient way to do this?
[1] https://www.sermonindex.net/speakers/paris-reidhead/the-trag...
pretty sure evolution spent more time and energy getting there then we ultimately will
I'd imagine one day there will be a limiting factor of cash to burn as well.
We're getting close. The first big AI bankruptcy can't be far off.
Lol well OAI is falling apart at the seams.
Simo takes a medical leave. And there appears to be friction between the CEO and CFO.
Big Gov will bail out the big guys if/when necessary
[dead]
It's more meaningful to most people than FLOPS/other measures of actual computing power.
Maybe it's just because the specifics on FLOPs are more complicated, especially given how many different floating point formats are floating around in ML. Even NVIDIA has like 6 different FLOPs numbers on their GPUs nowadays.
And you know Nvidia can't be constent with one format for FLOPs within a single graph, 1,000,000x faster but comparing FP32 to FP8 or NVFP4 and acting like it's the same.
Some of it might be market-signaling to the broader energy industry: "hey would you PLEASE build more power plants and power lines? Look at all this money we have, we will pay for it!"
Interesting timing given the quantum computing timeline pressure from this week's cryptography discussions. $30B run-rate and gigawatts of TPU capacity — and meanwhile the most interesting AI work I've seen lately runs on a phone in Termux with no cloud dependency at all. Both things are true simultaneously.
I don’t understand Claude Code’s moat here. What can it do that opencode can’t or couldn’t fairly easily implement?
The moat is in:
1. Opus and Sonnet.
2. Compute capacity. Anthropic has much more of it than your average coding startup.
3. The developing ecosystem around Claude Code.
I don’t think Opus and Sonnet are significantly better than Gemini or ChatGPT. Am I missing something?
It looks to me that Anthropic is one or two Gemmas away from a lot of people using Opus for 20% of hard use cases and letting on-device LLM rip through the code base on a Mac Mini or Studio and OpenCode.
Once Claude Code is not the only game in town and Cowork is made redundant by Google pulling their finger out on integration with Workspace, what else is there for Anthropic?
On-device agentic use is orders of magnitude harder than simple chatting (which is still slow for SOTA), it uses up a huge amount of context and tokens on reading code and reasoning through it. It's sort of viable if you just set it to work overnight on some completely vibe-coded stuff, but that has very middling results. Giving feedback to the model interactively is completely out of the question.
Where open models can make a difference for agentic use is with third-party inference at scale, which can actually be fast enough for reasonable workflows.
That’s mostly because current open models are way too heavy and too general for coding. If the model is compact enough to run well on normal Apple Silicon, local stops being an overnight-only toy and you also lose the third-party dependency. That’s the bet behind rig.ai.
none of the three are even remote moat
How so? Opus and Sonnet are frontier models which cannot easily be replicated. Compute has real physical constraints which require appropriate procurement at this scale. At least those two points seem like pretty strong moats against the majority of companies.
You don't need to "replicate" Opus and Sonnet, you just need to match their overall performance at lower cost. That's been absolutely doable so far, with a steadily decreasing lag time.
That's a fair response. But I'm not aware of any metrics supporting the point that the lag time is decreasing. The discourse I've seen has more focused on the ways Claude/OpenAI/Google have pulled away from the rest of the pack.
To be clear, I accept you might be right, but I think the crux is whether lag time is increasing, steady or growing.
absolutely! I am sooooo confused why people think either claude or openai have any sort of moat outside of mom&pop only heard about those on the tv
The moat is in training on the most high value user sessions. What Altman referred to as the "data flywheel."
You're right and the your reasoning is great. Anthropic should fold and give up their $30 billion ARR just announced in the OP. Shut it all down, no moat here.
/s
The moat is Anthropic's legal team and their ability to make legal threats.
Code is not the moat, it's the gateway drug to their subscription (hence why they just locked other harnesses from using their subscription).
And the subscription is not Anthropic's moat either since it's likely heavily subsidized. They're just using it to acquire customers.
The moat is locking you into Anthropic's model particularities (extended thinking, getting you into their "mindset", etc.)
Claude Code can be lower cost.
OpenCode: you pay per token.
Claude Code: you pay a flat fee.
Claude Code personal or Team: you pay a flat fee
Claude Code Enterprise: you pay per token
How is compute shortage to satisfy demand manifested? Obviously they never close sign-ups, so only option is to extended queues? But if demand grows like crazy, then queues should get longer, yet my pro claude plan seems snappy with only occasional retries due to 429.
They have several levers for demand destruction. From Anthropic's POV, I suspect this is least to worst bad
- reducing the surface area of "acceptable use" (e.g., blocking third-party tools OpenClaw)
- tighter usage limits and more subscription tiers
- increasing existing subscription prices
- moving to usage based model completely
- taking away compute from training next gen models (future demand destruction)
Reduce quality? Like quantizing models or context cache
Or reducing default reasoning levels?
There's no limit to the algorithms. People dont understand yet. They can learn the whole universe with a big enough compute cluster. We built a generalizable learning machine
the question is will we experience resource constraints before we get there? what if the step up to post-scarcity is gated by a compute level just out of our reach?
human ingenuity will solve this
Or we'll have ecological collapse.
NFTs will solve this
There are limits to algorithms. AI won't solve the halting problem nor will it solve EXPTIME problems in polynomial time.
Not sure if this is satire.
Edit: What we have built is a natural language interface to existing, textually recorded, information. Transformers cannot learn the whole universe because the universe has not yet been recorded into text.
Transformers operate on images and a variety of sensor data. They can also operate completely on non-textual inputs and outputs. I don't know what the ceiling on their capabilities is, but the complaint that they only operate on text seems just obviously wrong. There are numerous examples but one is meteorological forecasting which takes in a variety of time series sensor inputs and outputs e.g. time-series temperature maps. https://www.nature.com/articles/s41598-025-07897-4
Based on a glance at their other comments: not satire.
100% agreed. Sadly, lots of people out there with the "trust me bro, just need more compute". Hopefully we don't consume all the planet's resources trying.
I reevaluated my priors long ago when I saw that scaling laws show no sign of stopping, no sign of plateau.
Strangely some people on HN seem to desperately cling to the notion that it's all going to come to a halt. This is unscientific. What evidence do you have - any evidence - that the scaling laws are due to come to an end?
All the curves have been levelling off as expected. Not really sure what you're talking about.
They have not, every successful pre-train as of late has had performance increases greater than what the scaling laws predict.
Those gains are arch based, data quality based, etc. Scaling laws only relate to data volume and compute, holding other factors constant.
I’d like to see something that indicates models are getting better without the need for more training data. I would expect most gains are coming from more and better labeled data. We’re racing towards a complete encyclopedia of human knowledge. If we get there that’s only a drop in the bucket of all knowable things.
Bro the planet is literally experiencing a climate disaster and you think the solution is to create more systems that are misaligned with the planet's ecosystem for humans?
I guess the great filter is a real thing and not just a thought experiment.
I assure you that voluntary meat consumption because "taste buds go brr" is a much bigger problem than AI that results in actual productivity gains (and potentially solve the very climate crisis you complain about.)
Completely agree. Meat should be priced to include externalities. People can get used to beans. Beans are great!
I suspect it's not that people do not see the progress, they fail to fully trust laws not truly backed by physics like the transistor laws. We empirically see that scaling works and continue to work.
https://en.wikipedia.org/wiki/Neural_scaling_law
Why should we have strong priors in either direction? Maybe it will keep scaling for decades like Moore's law. Maybe not.
The issue people have isn’t some interpretation of scaling laws, it’s whether the planet’s ecology is goi g to be able to sustain this endeavour.
I shouldn’t have to say this out loud, but if the environment collapses, we will die, and no amount of “just a bit more scaling bro, just think of the gains” will matter.
People's voluntary dietary choices cause far more suffering and ecological damage than AI, and for much less return or economic output. But you tell people to switch to plant based foods and they lose their shit.
Yes. There's more than one thing that needs to change if we're going to make it through this.
AFAIK the data does not need to be text.
Well diffusers are trained unsupervised on raw pictures. I don't know how they train multi-modal LLMs on images, but yes obviously they are consuming other media than just text. I don't think, but would be happy to be corrected, that models glean much of their "knowledge" from non-textual training data.
you couldnt be more wrong
Please tell me more. When I ask an LLM a question, and get a text response, can that response incorporate non-textual information from visual training data?
It’s more than likely not.
Poe's (c)law?
Poe’s (C)law: The more absurd AI-generated content becomes, the more likely people are to believe it is real.
TPU architecture explained
https://news.ycombinator.com/item?id=47637597
This is not a good sign for Nvidia. They might have to step up their TPU game and price competitiveness.
is anthropic interested in TPU?
[flagged]
[flagged]
[dead]