This was also interesting: "CS-4 delivers more than 1,000 tokens per second on models exceeding 10 trillion parameters." Was it known that there were 10 trillion parameter models in use?
I think the frontier providers keep the size of their models carefully hidden.
You don't really need to train a 10T model to test cerebras against a 10T model. You can feed it an untrained (randomly initialized) model and benchmark it. Result will be gibberish but performance the same.
I'm confused. I thought Mythos 5 and Fable 5 were exactly the same model just with a different security layer in front of it.
Could they mean the Mythos 5 Preview?
> I believe this report has confused Opus (which is known to be around 5T) and Fable.
5T for Opus feels quite high though. DeepSeek V4 Pro is a mere 1.6T and often described as a match with Opus in overall quality. Even the largest open models in common use are around 2.8T.
pretty sure 10 trillion parameters is now the norm among closed ai labs, given that nvidia also references the same 10 trillion number for their nvl72 racks
If this is true, it's even more impressive that some of the open weight models that are <3.5T in size, approx 33% of its size, are within a few points of it in the artificial analysis leaderboard.
Not necessarily, there could be diminishing returns on mere parameters count .
There is nothing to say for example a 1 Quadrillion parameter model will be vastly more intelligent than current SOTA especially since new training data is largely synthetic today
I read it as it is impressive because smaller models 2.5T are squeezing similar returns as 10T models despite being 1/4th size not that there beyond 2T today the number or parameters do not have much meaning
I do not rely on any LLM of any size for general knowledge baked into the weights, they all hallucinate and that is the wrong way to hold them imo
I think there is some merit in that smaller models cannot memorize so much of the training data, i.e. that they are less likely to do copyright infringement, and by analogy not having memorized SDK / API surfaces that have since changed from the training data
> I do not rely on any LLM of any size for general knowledge baked into the weights
You have to rely on it to a certain level for agentic/coding work, presuming that's the general subject we're talking about here... For instance I recently encountered a project where it would have been a lot worse if the LLM didn't already know "what is" xterm.js and a bunch of its associated npm-related/node related software. If it was still smart but had to google and find results for everything it would have been a lot more time consuming and risked sending it down a wrong path.
The fact that Musk claims Opus is 5T to justify why Grok is far behind should be taken with a massive grain of salt given he's a recidivist mythomaniac.
Honestly if Opus is 5T parameters while being matched by the biggest open models that are at least twice smaller, it would mean that the US is already behind China in the AI race, despite a significant edge in compute.
> The cost to train and infer that would be insane, even by today's standards.
This assumption is likely what has led to the erroneous failure.
Enterprise compute per rack has scaled multiple fold in the last 3-5 years. Alongside the training efficiency gains & datacenter scale increases, even 50T+ is well within reach at the top end.
The raw margins on proprietary model inference are rumored to be quite high though (they have to successfully defray the entire investment into model training and datacenter capacity for inference, which is massive enough). The API cost you're paying for the model includes that raw margin.
AMD along with cerebras may probably compete with NVIDIA monopoly in near future. Also, NVIDIA will have competition form multiple companies. Just my prediction.
Maybe, but GPU is just one aspect of NVIDIA's dominance. If you are buying Vera Rubin GPUs, you're getting an NVL72 rack, which is only one of several racks that you're probably buying. You'll also need your NVIDIA racks with NVIDIA networking & storage gear, too. At the end of the day, they're "vertically integrated" for your accelerated computing data center (e.g. the "AI Factory"). This doesn't even count the software layer, where CUDA + CUDA-X (not to mention the software for all the sysadmin pieces) has a huge first mover advantage over anyone else.
If you are making a decision to spend 50B on hardware, would you use the proven tech stack or rely on engineers taking an unspecified amount of time vibecoding your software stack when the hardware sits idle?
How about in a month or so when you have to run a slightly different workload?
if you are developing your own hardware, you provide your own stack to avoid lawsuits with nvidia. i don't think it's a technical problem at all, but a legal one. this is probably why zluda was scrapped by AMD and Intel. Nvidia technically bans the creation of CUDA reimplementations in their TOS if i remember correctly
It can definitely create a software stack for you if you hold it right, but the software stack supported by a trillion dollar company with decades of expertise, that also uses AI to improve its stack is probably gonna be better.
If cerebars is performing well, why didn't its predecessor, server S-3, become the largest API token provider on OpenRouter, surpassing the official model releases?
Without having any inside information, one possible theory:
All or a vast majority of of the cerebras manufacturing capacity was going to a few companies that aren't publicly available inference providers on openrouter, for their own internal use.
or
The asking price of the S-3, no matter how speedy it might be, for small/medium size customers made it economically prohibitive to purchase and use to sell public inference vs. buying more common nvidia b200 or whatever.
it only takes ~445 GB300 NVL72 (about $22b) to run ALL of openrouter demand for a year. Microsoft rolled out $32b of DC 2026Q1.
imo the issue is that most openrouter demand is inauthentic activity (things that anthropic and openai models will refuse to do like pretend to not be bots when interacting with humans)
They do offer API services to individual users... though with a set of models that makes it unlikely that you want to use it. They are promising Qwen 3.8 27B any day now though*.
if you have the money as an "individual user" to purchase one of their racks... save your money and retire.
* Actually they sent out an email claiming they already have it, but I don't seem to have access, they're promising to release it to the "shared tier" any day now.
Now that this hypothetical person has retired, what are they gonna do all day? Just sit on the beach and drink Mai Tais? If that's what they wanna do, sure, but nerds gonna nerd, and if I had that kind of money to retire on, I'd totally buy some ridiculously expensive AI box for fun.
Ah but if you have the kind of money where this is a reasonable retirement hobby purchase, you aren't bothered by representing yourself as a "enterprise" :P
The Cerebras hardware is not locked to specific models / model families. Taalas is the company that's etching models into their silicon, locking it to that model forever.
And advanced geothermal. Fervo Energy let's us get energy that's not based on burning fossil fuels but is, instead, able to produce energy from the ground.
Indeed. You need 45 to 60 liters per second of cooling water flowing over a Cerebras wafer every minute to keep it under 90C. And that’s assuming the water leaves at 90C…
More realistically, you need much more cooling water.
Just a reminder for everyone that we are only several years and 3 or 4 iterations into hardware being optimized for LLMs. We should all expect orders of magnitude improvement in speed and/or cost over the next 5 years. Then we can have fun conversations about "unlimited" "intelligence" and about what the price wars and profit margins of consumer AI products are when your average ChatGPT user costs the company $0.10 per month.
> CS-4 delivers more than 1,000 tokens per second on models exceeding 10 trillion parameters
And, the software side isn't finished being optimized, either. We've seen with Qwen 3.8 27B and DeepSeek V4 Flash 0731 and GLM 5.3 that quite small models can pack a punch. Intelligence density will improve, efficiency of kernels will improve, efficiency of KV caching and MTP will improve, algorithms for splitting workloads across compute units will improve.
It'll all be as cheap as DeepSeek was before the price hike. And, it'll become more and more realistic to run near-frontier intelligence on personal devices.
Taalas will be one of the great disaster investments of the early AI era. It'll be a near total write-down.
The absolute worst market time to etch a model to a chip is right now (very rapid iteration). There is no scenario where they can keep up. The Taalas approach will be viewed as comically foolish within just a few years.
Cerebras will win in terms of approach.
It's 1998: hey, I can drastically speed up your web service, let's etch it right to silicon.
> The absolute worst market time to etch a model to a chip is right now
Slightly disagree. It really depends on the price-point at which they can do that etching. ~1k usd / ~30B model in a hdd-sized case that fits on your desk? I'd buy one right now, even knowing that I'm "stuck" with whatever model of the day is.
Time to market also matters a ton. If they can start shipping chips <1 month after the weights drop that's much more compelling than if it's a 6+ month development pipeline.
1) waiting for another 3-5 generations of transistor improvements before it can fit into a single conventional chip, or
2) another generation before getting a monster of a chip (1000+ mm^2), and prices for flawless etching scale quadraticly (likely $1000+ for manufacturing costs alone).
Could happen, but it's a long shot for a market that could be satiated by specialized accelerators.
I just want to but hardware so I can run a model at home that is fast. I don't see myself installing a server that burns almost two hundred kilowatts but maybe a card which runs a 27B Qwen...
maybe AMD wants the IP to deploy it once ai model development slows down in a few years. Or, their large cloud customers do want to burn through silicon, basically paying rent to AMD for models etched on silicon.
This is part of why I think the data center build-out is a bubble. We've barely scratched the surface when it comes to hardware optimization. We'll see exponential improvements in energy efficiency and speed over the next decade. Exponential, not linear.
GPUs really aren't that great for AI. They just happen to be the best chips we have in mass production right now for this work load, and it takes time to field new designs. Basically every chip engineer on the planet is working on this right now.
Whether it's a bubble or not depends on how much the demand for compute and the type of workload keeps growing, though.
If AI tends to be something used mainly in ideation and development, which is how a lot of people use it today, then once consumer hardware gets good enough you could see a bunch of the current data centre workloads move onto consumer devices.
But if AI starts being used more in repeatable, operational workloads I think it makes sense to have significant cloud infrastructure for it. TBH I haven't seen much of this, and I've been skeptical about people using agents for much of anything when it can be done with just software. But we are starting to see more of this kind of workload, like the taggable Claude in your slack etc that people seem to really love.
True, I have to agree with you. The AI giants might be investing a huge amount of money in generation 1 technology. There might be a much better way to do it just around the corner. They might know this and thus the hurry to IPO.
A rough analogy would be if the first generation of ISP's spent billions on dial-up exchanges, when fibre could be invented next year.
A cursory estimate courtesy of ChatGPT suggests that there is a grand total of one order of magnitude or less of power efficiency improvement available compared to current Blackwell if the entire system’s power consumption outside the ALUs went all the way to zero.
If you want three orders of magnitude improvement, you probably need to find two of those orders of magnitude somewhere else: process improvements, different ALU design, model architecture changes, etc.
Congratulations! You have just realized that the AI data center build out is a total scam, built on both the insurmountable trillions of debt, and the assumption that only GPUs are all we need to continue scaling.
There exist other AI accelerators (TPUs, ASICs) that perfectly exceed the throughput that LLMs need to scale as well. But the true solution is more software optimizations. There's a tiny handful of them but more needs to be discovered so that we can reduce building hundreds of more data centers as the alternatives mature.
As better software becomes more useful for the alternative AI hardware for developers with LLMs running efficiently you then would have more choices of hardware to run your LLMs on rather than just only GPUs.
TPUs and ASICs run in data centers too. Your argument only holds true if there's some satisfied limit to demand for inference. If not, data centers will continue to spring up to host more and more agents. Even if agents were running on hardware and software as efficient as the human brain, its conceivable we want trillions of them running at any given time which would require data center scale.
Everything has some satisfied limit to demand, often depending on the price. If you assume there will never be any satisfied limit to demand for inference at any price you can justify any investment.
Yeah, but there's certainly a part of the curve where price drops by X OOMs and demand increases by much more than X OOMs. (Presumably some of that is substitution and some of that is new use cases.)
I wonder what this looks like in 5 years... Will there be a massive push to repurpose these giant boxes into housing? Will they get turned back into the farm land from where they came? When a data center goes bust, what happens to the parts left behind?
I'd think the infrastructure would tend towards factories, smelters, and so on. Industrial things that have reasonably high power demands, can use the building, and don't care about the lack of windows.
They're typically not built where you want housing, and the buildings are distinctly the wrong shape.
If you can't use the power infrastructure profitably my next thought would be warehousing.
But also... we've seen a pretty continually increasing demand for compute. Even if AI busts a bit (or becomes a bit more efficient) I bet most data centres stay data centres, just less profitable ones.
Huh, why I'm not surprised that HN is full of opinions confidently stated without any numbers or resources to back up?
> built on both the insurmountable trillions of debt, and the assumption that only GPUs are all we need to continue scaling.
Insurmountable according to whom? And who assume that only GPUs are all we need to continue scaling? Google, Amazon, Microsoft, Meta and OpenAI, all have or plan custom non-GPU AI chips. Do they plan to use them not for scaling?
Information about RAM type/size and connection topology of the RAM to be used for context cache seems to be conspicuously absent from the slick looking marketing materials.
44GB on-chip-sram * 3 chips. Per chip: 43.2 PB/s memory access + 53.5 PB/s on-chip fabric bandwidth + 2.4 Tbits/s "IO" bandwidth (I think that means their RoCE v2 RDMA over Ethernet interface).
I suspect there might be a certain amount of customization for how much RAM they attach when you order it.
They have managed to make the external link 300GBps/2us. Cs3 was 150/5.
This is 1/3rd blackwells nvlink c2c bandwidth already. Not too bad. We can make KV cache offload work with that I suppose.
If magically KV cache was not an issue, pipeline parallelism on cerebras can be quite pleasant. As for the KV cache offload, I have hopes their CPO solution they're trying with that canadian company ends up bearing fruit.
> Introducing the all new Cerebras CS-4, a revolutionary rack-scale solution that delivers upto 30x faster inference compared to GPUs, enhanced economics, and a simple path todeploy [sic] hyperscale capacity.
If they had ask Claude it would probably look like this: Introducing the all new Cerebras CS-4, a revolutionary rack-scale solution that delivers up to 30x faster inference compared to GPUs, enhanced economics, and a simple path to load-bearing hyper scale capacity.
The comparison seems incomplete. CS‑4 is a full rack-scale system with three wafer-scale processors, but the exact GPU models, GPU count, power consumption, price information are not disclosed.
We still don't know if buying a multi-GPU rack (or racks) is cheaper and/or more efficient in power.
The fact that they didn't disclose these numbers makes me believe that the numbers are not in their favor. And personally, makes me see them as disingenuous.
Cerebras should slowly also move to dgx/ryzen market for a desktop version for masses at affordable price yet providing substantial tokens/second on desktop
OpenAI needs to immediately move to acquire Cerebras.
Nvidia's extreme margin is the opportunity for OpenAI's cost reduction. Buying Cerebras would pay for itself and they should take all of its future production (after filling required contracts).
Right now China's models have no silicon moat. Cerebras as a drastic speed-up / cost-reduction potential, can assist in building a competitive moat. And every time a Cerebras pops up, OpenAI or Anthropic should eat them if at all possible.
There's no stand-alone frontier AI company of great scale in the near future that doesn't have a large silicon advantage in-house. Apple knew it in smartphones, Google figured it out a long time ago as well.
OpenAI is partnering with Cerebras while simultaneously investing in their own silicon play. Hedged bets.
After sitting thru their keynote today, it makes sense. The main throughput speedups they tout are an obvious evolution of the GPU that all companies will be building in the next year. Wafer-scale interconnected memory and compute is just going to beat out mountains of network cabling any day on both cost and performance metrics.
Five years from now, I don't know why anyone will still be using Nvidia for inference. Note that Cerebras is for inference only, not for training.
I understand that Cerebras has competition, but this bodes even more poorly for Nvidia for inference. Nvidia may still have a role to play for training, however.
Nvidia is at this time a pretty well run company tech wise. They are going to keep iterating on the inferencing hardware stack over the next five years too.
Is it just me or is it bizarre that they're advertising old open-weight models.
GLM 4.7 (December 2025) not 5 (Feb) 5.1 (April) or 5.2 (June). 5.3 (4 days ago) is, to be fair, not open weights yet... but there's a lot since 4.7.
Kimi K2.7 (April) not K2.7-code (June) or K3 (July).
Gemma 4 (April), Llama (April), and gpt-oss (August 2025) are up to date, but old (for models).
Meanwhile the closed source GPT 5.6 sol is up to date (June)...
Should potential purchasers take away from this that they're not going to be able to run recent models unless they front the cost of developing software or something?
I mean the product is a server rack and while there's no advertised price I would assume it's six figures. So yes, an enterprise product.
But even an enterprise is going to care about the difference between "we can run the model we want with support from the manufacturer" and "we have to purchase the product, and then spend another 6 figure sum having developers port a recent model to the product to use it".
This was also interesting: "CS-4 delivers more than 1,000 tokens per second on models exceeding 10 trillion parameters." Was it known that there were 10 trillion parameter models in use?
I think the frontier providers keep the size of their models carefully hidden.
> According to FT, industry estimates say Anthropic's most advanced Mythos 5 has about 8 trillion parameters and Fable 5 about 5 trillion
https://www.reuters.com/technology/bytedance-targets-mega-ai...
I believe this report has confused Opus (which is known to be around 5T) and Fable.
Other reports say 10T. See for example https://eu.36kr.com/en/p/3760679047267075?ref=explainx where Musk talks about the models being trained on Colossus2
5T for Opus feels quite high though. DeepSeek V4 Pro is a mere 1.6T and often described as a match with Opus in overall quality. Even the largest open models in common use are around 2.8T.
It's not. Idk about who has more T's but, unfortunately, DS4 pro is not a match to Opus, at least not Opus 4.8.
There is nothing to say for example a 1 Quadrillion parameter model will be vastly more intelligent than current SOTA especially since new training data is largely synthetic today
https://pastes.io/r8F1AY8h
I think there is some merit in that smaller models cannot memorize so much of the training data, i.e. that they are less likely to do copyright infringement, and by analogy not having memorized SDK / API surfaces that have since changed from the training data
You have to rely on it to a certain level for agentic/coding work, presuming that's the general subject we're talking about here... For instance I recently encountered a project where it would have been a lot worse if the LLM didn't already know "what is" xterm.js and a bunch of its associated npm-related/node related software. If it was still smart but had to google and find results for everything it would have been a lot more time consuming and risked sending it down a wrong path.
The cost to train and infer that would be insane, even by today's standards.
Eg: https://www.reuters.com/technology/bytedance-targets-mega-ai...
That reports Mythos as 8T and Fable as 5T, but I think they mean Opus as 5T, which is widely known, eg: https://eu.36kr.com/en/p/3760679047267075?ref=explainx
Both Grok and Bytedance are training 10T models.
Honestly if Opus is 5T parameters while being matched by the biggest open models that are at least twice smaller, it would mean that the US is already behind China in the AI race, despite a significant edge in compute.
This assumption is likely what has led to the erroneous failure.
Enterprise compute per rack has scaled multiple fold in the last 3-5 years. Alongside the training efficiency gains & datacenter scale increases, even 50T+ is well within reach at the top end.
How about in a month or so when you have to run a slightly different workload?
If it's patents that are the problem then presumably all these large semiconductor companies have defensive parent portfolios.
Oops did they just out GPT-5.6 sol’s parameter count?
It seems it takes some time to run a new model on all the benchies, not sure they run all models on all of them either
All or a vast majority of of the cerebras manufacturing capacity was going to a few companies that aren't publicly available inference providers on openrouter, for their own internal use.
or
The asking price of the S-3, no matter how speedy it might be, for small/medium size customers made it economically prohibitive to purchase and use to sell public inference vs. buying more common nvidia b200 or whatever.
imo the issue is that most openrouter demand is inauthentic activity (things that anthropic and openai models will refuse to do like pretend to not be bots when interacting with humans)
if you have the money as an "individual user" to purchase one of their racks... save your money and retire.
* Actually they sent out an email claiming they already have it, but I don't seem to have access, they're promising to release it to the "shared tier" any day now.
Now that this hypothetical person has retired, what are they gonna do all day? Just sit on the beach and drink Mai Tais? If that's what they wanna do, sure, but nerds gonna nerd, and if I had that kind of money to retire on, I'd totally buy some ridiculously expensive AI box for fun.
Can you imagine something radiating that much energy into a space in your home?
This is far beyond the practical maximums of like 10 to 15kW per 44U cabinet front to rear air cooling for 'regular' rackmount server stuff.
More realistically, you need much more cooling water.
> CS-4 delivers more than 1,000 tokens per second on models exceeding 10 trillion parameters
Wow!
It'll all be as cheap as DeepSeek was before the price hike. And, it'll become more and more realistic to run near-frontier intelligence on personal devices.
I'm honestly baffled they were not acquired by somebody else (sorry AMD).
The absolute worst market time to etch a model to a chip is right now (very rapid iteration). There is no scenario where they can keep up. The Taalas approach will be viewed as comically foolish within just a few years.
Cerebras will win in terms of approach.
It's 1998: hey, I can drastically speed up your web service, let's etch it right to silicon.
Slightly disagree. It really depends on the price-point at which they can do that etching. ~1k usd / ~30B model in a hdd-sized case that fits on your desk? I'd buy one right now, even knowing that I'm "stuck" with whatever model of the day is.
Qwwen3.5 122b was released 6 months ago and is still best in class overall 100-140 B param model.
....
:T
Considering the 8B model uses 53 billion transistors, that's 6.625 transistors per parameter.
https://taalas.com/products/
Assuming they can get it down to 3 (somehow), that's still 300 transistors, or 5.565 RX 9070s.
https://www.techpowerup.com/gpu-specs/radeon-rx-9070.c4250
You're looking at
1) waiting for another 3-5 generations of transistor improvements before it can fit into a single conventional chip, or
2) another generation before getting a monster of a chip (1000+ mm^2), and prices for flawless etching scale quadraticly (likely $1000+ for manufacturing costs alone).
Could happen, but it's a long shot for a market that could be satiated by specialized accelerators.
Yes, please!
GPUs really aren't that great for AI. They just happen to be the best chips we have in mass production right now for this work load, and it takes time to field new designs. Basically every chip engineer on the planet is working on this right now.
If AI tends to be something used mainly in ideation and development, which is how a lot of people use it today, then once consumer hardware gets good enough you could see a bunch of the current data centre workloads move onto consumer devices.
But if AI starts being used more in repeatable, operational workloads I think it makes sense to have significant cloud infrastructure for it. TBH I haven't seen much of this, and I've been skeptical about people using agents for much of anything when it can be done with just software. But we are starting to see more of this kind of workload, like the taggable Claude in your slack etc that people seem to really love.
A rough analogy would be if the first generation of ISP's spent billions on dial-up exchanges, when fibre could be invented next year.
In the scenario, engineering everything becomes so easy - so why not optimize everything? every component, every product, every system?
And maybe llm's could invent. So even more to simulate. And simulation is inherently compute-heavy.
So unless there are some other bottlenecks, we'll use a lot of simulation servers.
In that 5+ year timeline, the compute per watt could change by three orders of magnitude.
GPUs are to LLMs what CPUs are to gaming — not a good fit.
If you want three orders of magnitude improvement, you probably need to find two of those orders of magnitude somewhere else: process improvements, different ALU design, model architecture changes, etc.
There exist other AI accelerators (TPUs, ASICs) that perfectly exceed the throughput that LLMs need to scale as well. But the true solution is more software optimizations. There's a tiny handful of them but more needs to be discovered so that we can reduce building hundreds of more data centers as the alternatives mature.
As better software becomes more useful for the alternative AI hardware for developers with LLMs running efficiently you then would have more choices of hardware to run your LLMs on rather than just only GPUs.
Unlimited.
What has been the limit to electricity demand globally?
Unlimited.
We can't get enough and never will. Costs have to become pretty severe to turn back the demand as well.
They're typically not built where you want housing, and the buildings are distinctly the wrong shape.
If you can't use the power infrastructure profitably my next thought would be warehousing.
But also... we've seen a pretty continually increasing demand for compute. Even if AI busts a bit (or becomes a bit more efficient) I bet most data centres stay data centres, just less profitable ones.
> built on both the insurmountable trillions of debt, and the assumption that only GPUs are all we need to continue scaling.
Insurmountable according to whom? And who assume that only GPUs are all we need to continue scaling? Google, Amazon, Microsoft, Meta and OpenAI, all have or plan custom non-GPU AI chips. Do they plan to use them not for scaling?
What's the point of 1000tok/s if you have to do prefill on every agentic turn which at 100k depth would make it 1.5 min latency every turn?
44GB on-chip-sram * 3 chips. Per chip: 43.2 PB/s memory access + 53.5 PB/s on-chip fabric bandwidth + 2.4 Tbits/s "IO" bandwidth (I think that means their RoCE v2 RDMA over Ethernet interface).
I suspect there might be a certain amount of customization for how much RAM they attach when you order it.
This is 1/3rd blackwells nvlink c2c bandwidth already. Not too bad. We can make KV cache offload work with that I suppose.
If magically KV cache was not an issue, pipeline parallelism on cerebras can be quite pleasant. As for the KV cache offload, I have hopes their CPO solution they're trying with that canadian company ends up bearing fruit.
Did nobody proofread this?
It’s still incredibly obvious.
Nvidia's extreme margin is the opportunity for OpenAI's cost reduction. Buying Cerebras would pay for itself and they should take all of its future production (after filling required contracts).
Right now China's models have no silicon moat. Cerebras as a drastic speed-up / cost-reduction potential, can assist in building a competitive moat. And every time a Cerebras pops up, OpenAI or Anthropic should eat them if at all possible.
There's no stand-alone frontier AI company of great scale in the near future that doesn't have a large silicon advantage in-house. Apple knew it in smartphones, Google figured it out a long time ago as well.
OpenAI is partnering with Cerebras while simultaneously investing in their own silicon play. Hedged bets.
After sitting thru their keynote today, it makes sense. The main throughput speedups they tout are an obvious evolution of the GPU that all companies will be building in the next year. Wafer-scale interconnected memory and compute is just going to beat out mountains of network cabling any day on both cost and performance metrics.
I understand that Cerebras has competition, but this bodes even more poorly for Nvidia for inference. Nvidia may still have a role to play for training, however.
GLM 4.7 (December 2025) not 5 (Feb) 5.1 (April) or 5.2 (June). 5.3 (4 days ago) is, to be fair, not open weights yet... but there's a lot since 4.7.
Kimi K2.7 (April) not K2.7-code (June) or K3 (July).
Gemma 4 (April), Llama (April), and gpt-oss (August 2025) are up to date, but old (for models).
Meanwhile the closed source GPT 5.6 sol is up to date (June)...
Should potential purchasers take away from this that they're not going to be able to run recent models unless they front the cost of developing software or something?
But even an enterprise is going to care about the difference between "we can run the model we want with support from the manufacturer" and "we have to purchase the product, and then spend another 6 figure sum having developers port a recent model to the product to use it".
A single AI server with a mere 8 GPUs from Nvidia is already mid 6 digits. A rack system from Nvidia is mid 7 digits.
There’s some info out there that suggests the CS1 had an 8 digits price tag, so it wouldn’t be surprising to see that here.