they're trying to compare at iso-power? I just want to see their box vs a box of 8 h100s b/c that's what people would buy instead, and they can divide tokens and watts if that's the pitch.
furthermore usually NVLink connected within the box (SXM instead of PCIe cards, although the physical data link is still PCIe.)
this is important because the daughter board provides PCIe switches which usually connect NVMe drives, NICs and GPUs together such that within that subcomplex there isn't any PCIe oversubscription.
since last year for a lot of providers the standard is the GB200 I'd argue.
Feels like theres some amount of (software) orchestration for making data sit on the right drives or traverse the right NICs, guess I never really thought about the complexity of this kind of scale.
I googled GB200, its cool that Nvidia sells you a unit rather than expecting you to DIY PC yourself.
GPU -> PCIe switch -> PCIe switch (most likely the CPU, with limited bw) -> PCIe switch -> GPU
NVLink comes into the picture as a separate, 2nd link between the GPUs: if you need to do GPU-to-GPU, you can use NVLink.
you never needed to DIY your stuff, at least not for the last 10 years: most hardware vendors (Supermicro, Dell, ...) will sell you a complete system with 8 GPUs.
what's nice on GH200/GBx00/VR systems, is that you can use chip-to-chip NVLink between the CPU and GPU, so the CPU can access GPU memory coherently and vica versa.
Yeah they are defining a "rack" as 15kW, though 3x H100 PCIe is only a bit over 1kW. So they are assuming GPUs are <10% of rack power usage which sounds suspiciously low.
I think Llama 3 focus mostly reflects demand. It may be hard to believe, but many people aren't even aware gpt-oss exists.
The 8B models are easier to run on an RTX to compare it to local inference. What llama does on an RTX 5080 at 40t/s, Furiosa should do at 40,000t/s or whatever… it’s an easy way to have a flat comparison across all the different hardware llama.cpp runs on.
It still kind of makes the point that you are stuck with a very limited range of models that they are hand implementing. But at least it's a model I would actually use. Give me that in a box I can put in a standard data center with normal power supply and I'm definitely interested.
But I want to know the cost :-)
That's 86 token/second/chip
By comparison, a H100 will do 2390 token/second/GPU
Am I comparing the wrong things somehow?
Targeting power, cooling, and TCO limits for inference is real, especially in air-cooled data centers.
But the benchmarks shown are narrow, and it’s unclear how well this generalizes across models and mixed production workloads. GPUs are inefficient here, but their flexibility still matters.
Seems like it would obviously be in TSMCs interest to give preferential taping to nvidia competitors, they benefit from having a less consolidated customer base bidding up their prices.
Anyone not under some kind of export restrictions can scrounge together some GPUs to train a frontier model (hell, even DeepSeek which is under these restrictions could) but providing a service that can compete with OpenAI et al. will prove to be quite costly. 3x improvements in inference are therefore nothing to sneeze at IMO.
The story with Intel around these times was usually that AMD or Cyrix or ARM or Apple or someone else would come around with a new architecture that was a clear generation jump past Intel's, and most importantly seemed to break the thermal and power ceilings of the Intel generation (at which point Intel typically fired their chip design group, hired everyone from AMD or whoever, and came out with Core or whatever). Nvidia effectively has no competition, or hasn't had any - nobody's actually broken the CUDA moat, so neither Intel nor AMD nor anyone else is really competing for the datacenter space, so they haven't faced any actual competitive pressure against things like power draws in the multi-kilowatt range for the Blackwells.
The reason this matters is that LLMs are incredibly nifty often useful tools that are not AGI and also seem to be hitting a scaling wall, and the only way to make the economics of, eg, a Blackwell-powered datacenter make sense is to assume that the entire economy is going to be running on it, as opposed to some useful tools and some improved interfaces. Otherwise, the investment numbers just don't make sense - the gap between what we see on the ground of how LLMs are used and the real but limited value add they can provide and the actual full cost of providing that service with a brand new single-purpose "AI datacenter" is just too great.
So this is a press release, but any time I see something that looks like an actual new hardware architecture for inference, and especially one that doesn't require building a new building or solving nuclear fusion, I'll take it as a good sign. I like LLMs, I've gotten a lot of value out of them, but nothing about the industry's finances add up right now.
The acquisitions do. Remember Groq?
Most M&As arent done by value investors.
Nothing about the industry’s finances, or about Anthropic and OpenAI’s finances?
I look at the list of providers on OpenRouter for open models, and I don’t believe all of them are losing money. FWIW Anthropic claims (iirc) that they don’t lose money on inference. So I don’t think the industry or the model of selling inference is what’s in trouble there.
I am much more skeptical of Anthropic and OpenAI’s business model of spending gigantic sums on generating proprietary models. Latest Claude and GPT are very very good, but not better enough than the competition to justify the cash spend. It feels unlikely that anyone is gonna “winner takes all” the market at this point. I don’t see how Anthropic or OpenAI’s business model survive as independent entities, or how current owners don’t take a gigantic haircut, other than by Sam Altman managing to do something insane like reverse acquiring Oracle.
EDIT: also feels like Musk has shown how shallow the moat is. With enough cash and access to exceptional engineers, you can magic a frontier model out of the ether, however much of a douche you are.
they use Chinese open LLMs, but Chinese companies have moat: training datasets and some non-opensource tech, and also salaried talents, which one would need serious investment for if decide to bootstrap competitive frontier model today.
As for the business side, I’ve yet to hear of a transformative business outcome due to LLMs (it will come, but not there yet). It’s only the guys selling the shovels that are making money.
This entire market runs on sovereign funds and cyclical investing. It’s crazy.
It is, however, actually funny how bad e.g. the amazon chatbot (Rufus) is on amazon.com. When asked where a particular CC charge comes from, it does all sorts of SQL queries into my account, but it can't be bothered to give me the link to the actual charges (the page exists and solves the problem trivially).
So, maybe, the callcenter troubles will take some time to materialize.
I don't know who needs to hear this, but the real break through in AI that we have had is not LLMs, but generative AI. LLM is but one specific case. Furthermore, we have hit absolutely no walls. Go download a model from Jan 2024, another from Jan 2025 and one from this year and compare. The difference is exponential in how well they have gotten.
GP was talking about commercially hosted LLMs running in datacenters, not free Chinese models.
Local is definitely still improving. That’s another reason the megacenter model (NVDA’s big line up forever plan) is either a financial catastrophe about to happen, or the biggest bailout ever.
But yes it will write you a flawless, physics accurate flight simulator in rust on the first try. I've proven that. I guess what I'm trying to say is Anthropic was eating their lunch at coding, and OpenAI rose to the challenge, but if you're not doing engineering tasks their current models are arguably worse than older ones.
Is this the second most abused english word (after 'literally')?
> a model from Jan 2024, another from Jan 2025 and one from this year
You literally can't tell the difference is 'exponential', quadratic, or whatever from three data points.
Plus it's not my experience at all. Since Deepseek I haven't found models that one can run on consumer hardware get much better.
I've been wondering about this for quite a while now. Why does everybody automatically assume that I'm using the decimal system when saying "orders of magnitude"?!
Unless you've explicitly stated that you mean something else, people have no reason to think that you mean something else.
Anyway, there are 10 types of people, those who understand binary and those who don't.
I don’t think that’s true. I think both my mother and my mother-in-law would start to complain pretty quickly if they got pushed back to 4o. Change may have felt gradual, but I think that’s more a function of growing confidence in what they can expect the machine to do.
I also think “ask how long to boil an egg” is missing a lot here. Both use ChatGPT in place of Google for all sorts of shit these days, including plenty of stuff they shouldn’t (like: “will the city be doing garbage collection tomorrow?”). Both are pretty sharp women but neither is remotely technical.
I did. The old one is smarter.
(The newer ones are more verbose, though. If that impresses you, then you probably think members of parliament are geniuses.)
For api models, OpenAI's releases have regularly not been an improvement for a long while now. Is sonnet 4.5 better than 3.5 outside pretentius agentic workflows it's been trained for? Basically impossible to tell, they make the same braindead mistakes sometimes.
Not likely since TSMC has a new process with big gains.
> The story with Intel
Was that their fab couldn’t keep up not designs.
Instead, they couldn't get 10nm to work and launched one low-power SKU in 2018 that had almost half the die disabled, and stuck to 14nm from 2014-2021.
Based on what?
Their measured performance on things people care about keep going up, and their software stack keeps getting better and unlocking more performance on existing hardware
Inference tests: https://inferencemax.semianalysis.com/
Training tests: https://www.lightly.ai/blog/nvidia-b200-vs-h100
https://newsletter.semianalysis.com/p/mi300x-vs-h100-vs-h200... (only H100, but vs AMD)
> but nothing about the industry's finances add up right now
Is that based just on the HN "it is lots of money so it can't possibly make sense" wisdom? Because the released numbers seem to indicate that inference providers and Anthropic are doing pretty well, and that OpenAI is really only losing money on inference because of the free ChatGPT usage.
Further, I'm sure most people heard the mention of an unnamed enterprise paying Anthropic $5000/month per developer on inference(!!) If a company if that cost insensitive is there any reason why Anthropic would bother to subsidize them?
I mean the amount of money invested across just a handful of AI companies is currently staggering and their respective revenues are no where near where they need to be. That’s a valid reason to be skeptical. How many times have we seen speculative investment of this magnitude? It’s shifting entire municipal and state economies in the US.
OpenAI alone is currently projected to burn over $100 billion by what? 2028 or 2029? Forgot what I read the other day. Tens of billions a year. That is a hell of a gamble by investors.
This means the bigger questions are whether you believe the labs are compute constrained, and whether you believe more capacity would allow them to drive actual revenue. I think there is a decent chance of this being true, and under this reality the investments make more sense. I can especially believe this as we see higher-cost products like Claude Code grow rapidly with much higher token usage per user.
This all hinges on demand materialising when capacity increases, and margins being good enough on that demand to get a good ROI. But that seems like an easier bet for investors to grapple with than trying to compare future investment in capacity with today's revenue, which doesn't capture the whole picture.
Basically, it strikes me as not really apples to apples.
The competition requiring them to spend that money on training and free users does complicate things. But when you just look at it from an inference perspective, looking at these data centres like token factories makes sense. I would definitely pay more to get faster inference of Opus 4.5, for example.
This is also not wholly dissimilar to other industries where companies spend heavily on R&D while running profitable manufacturing. Pharma semiconductors, and hardware companies like Samsung or Apple all do this. The unusual part with AI labs is the ratio and the uncertainty, but that's a difference of degree, not kind.
So if you ignore the majority of the costs, then it makes sense.
Opus 4.5 was released on November 25, 2025. That is less than 2 months ago. When they stop training new models, then we can forget about training costs.
That's not to mention that Dario Amodei has said that their models actually have a good return, even when accounting for training costs [0].
Do we know this is true for AI?
You spend the same amount on R&D whether you have one hobbyist user or 90% market share.
So I'll ask, how is that any different than fabs? From what I understand R&D is absurd and upgrading to a new node is even more absurd. The resulting chips sell for chump change on a per unit basis (analogous to tokens). But somehow it all works out.
Well, sort of. The bleeding edge companies kept dropping out until you could count them on one hand at this point.
At first glance it seems like the analogy might fit?
Invariably, there's going to be a collapse in the hype, the bubble will burst, and an investment deleveraging will remove a lot of money from the space in a short period of time. The bigger the bubble, the more painful and less survivable this event will be.
That sounds like “we’re profitable if you ignore our biggest expenses.” If they could be profitable now, we’d see at least a few companies just be profitable and stop the heavy expenses. My guess is it’s simply not the case or everyone’s trapped in a cycle where they are all required to keep spending too much to keep up and nobody wants to be the first to stop. Either way the outcome is the same.
OpenAI could put in ads tomorrow and make tons of money overnight. The only reason they don't is competition. But when they start to find it harder to raise capital to fund their growth, they will.
Yes and no. Some of it just claims to be "AI". Like the hyperscalers are building datacenters and ramping up but not all of it is "AI". The crypto bros have rebadged their data centers into "AI".
That the previous unsustainable bubble is rebranding into the new one, is maybe not the indicator of stability we should be hoping for
I'm more concerned about fully-loaded dollars per token - including datacenter and power costs - rather than "does the chip go faster." If Nvidia couldn't make the chip go faster, there wouldn't be any debate, the question right now is "what is the cost of those improvements." I don't have the answer to that number, but the numbers going around for the costs of new datacenters doesn't give me a lot of optimism.
> Is that based just on the HN "it is lots of money so it can't possibly make sense" wisdom?
OpenAI has $1.15T in spend commitments over the next 10 years: https://tomtunguz.com/openai-hardware-spending-2025-2035/
As far as revenue, the released numbers from nearly anyone in this space are questionable - they're not public companies, we don't actually get to see inside the box. Torture the numbers right and they'll tell you anything you want to hear. What we _do_ get to see is, eg, Anthropic raising billions of dollars every ~3 months or so over the course of 2025. Maybe they're just that ambitious, but that's the kind of thing that makes me nervous.
Yes, but those aren't contracted commitments, and we know some of them are equity swaps. For example "Microsoft ($250B Azure commitment)" from the footnote is an unknown amount of actual cash.
And I think it's fair to point out the other information in your link "OpenAI projects a 48% gross profit margin in 2025, improving to 70% by 2029."
OpenAI can project whatever they want, they're not public.
Private companies do have a license to lie to their shareholders.
It's worse than not contracted. Nvidia said in their earnings call that their OpenAI commitment was "maybe".
The economics of the entire setup are laughable and it's obvious that it's a massive bubble. The profit that'd need to be delivered to justify the current valuations is far beyond what is actually realistic.
What moat does OpenAI have? I'd argue basically none. They make extremely lofty forecasts and project an image of crazy growth opportunities, but is that going to ever survive the bubble popping?
Of course, a rational investor looks at this and discounts the fact that most of those promises are predicated on insane growth that has no grounding in reality.
However, there are plenty of greedy or irrational investors, whose recklessness will affect everyone, not just them.
For the AI company being bought: I wouldn't trust these shares or valuations, because the money invested is going on GPUs and back to Nvidia.
Nvidia is literally selling GPUs with 90% profit margin and still everything is out of stock, which is unheard of before.
Companies have wasted more money on dumber things so spending isn't a good measure.
And what about the countless other AI companies? Anthropic has one of the top models for coding so that's like saying there ins't a problem pre dot com bubble because Amazon is doing fine.
The real effects of AI is measured in rising profit of the customers of those AI companies otherwise you're looking at the shovel sellers
I haven't and I'd like to know more.
LLM's are probably a small fraction of the overall GPU compute in use right now. I suspect in the next 5 years we'll have full Hollywood movies being completely generated (at least the specialfx) entirely by AI.
The other thing to compare is the narrative quality. I find even middling books to be of much higher quality than blockbuster movies on average. Or rather I'm constantly appalled at what passes for a decent script. I assume that's due to needing to appeal to a broad swath of the population because production is so expensive, but understanding the (likely) reason behind it doesn't do anything to improve the end result.
So if "all" we get out of this is a 1000x reduction in production budgets which leads to a 100x increase in the amount of media available I expect it will be a huge win for the consumer.
I wouldn’t be that dismissive. Some have managed to make impressive things with them (although nothing close to an actual movie, even a short).
https://www.youtube.com/watch?v=ET7Y1nNMXmA
A bit older: https://www.youtube.com/watch?v=8OOpYvxKhtY
Compared to two years ago: https://www.youtube.com/watch?v=LHeCTfQOQcs
They have really advanced the coherency of real-time AI generation.
It used to mean psychedelic weird things worthy of the strangest dreams or an acid trip.
Then it meant strangely blurry with warped alien script and fifteen fingers, including one coming out of another’s second phalanx
Now it means something odd, off, somewhat both hard to place and obvious, like the CGI "transparent" car (is it that the 3D model is too simple, looks like a bad glass sculpture, and refracts light in squares?) and ice cliffs (I think the the lighting is completely off, and the colours are wrong) in Die Another Day.
And if that’s the case, then these models have covered far more in far less time then it took computer graphics and CGI.
It's not feature length movie but I'm not sure there's any reason why it couldn't be, and its not technically perfect but pretty damn good.
In 2001, there were something like 50+ OC-768 hardware startups.
At the time, something like 5 OC-768 links could carry all the traffic in the world. Even exponential doubling every 12 months wasn't going to get enough customers to warrant all the funding that had poured into those startups.
When your business model bumps into "All the <X> in the world," you're in trouble.
Didn't the Core architecture come from the Intel Pentium M Israeli team? https://en.wikipedia.org/wiki/Intel_Core_(microarchitecture)...
If you wanted to make that point, Itanium or 64-bit/multi-core desktop processing would be better examples than Core.
And I'm still convinced we're not paying real prices anywhere. Everyone is still trying to get market share so the prices are going to go up when this all needs to sustain itself. At that point, which use cases become too expensive and does that shrink it's applicability ?
Now that the model architecture has settled into something a bit more predictable, I wouldn't be surprised if we saw a little more specialisation in the hardware.
I think the software side of the story is underestimated. Nvidia has a big moat there and huge community support.
https://furiosa.ai/blog/tensor-contraction-processor-ai-chip...
I just want to buy ddr5 and not pay an arm and a leg for my power bill!
At this point, I don't even think they do the envelope math anymore. However much money investors will be duped into giving them, that's what they'll spend on compute. Just gotta stay alive until the IPO!
Google presented TPUs in 2015. NVIDIA introduced Tensor Cores in 2018. Both utilize systolic arrays.
And last month NVIDIA pseudo-acquired Groq including the founder and original TPU guy. Their LPUs are way more efficient for inference. Also of note Groq is fully made in USA and has a very diverse supply chain using older nodes.
NVIDIA architecture is more than fine. They have deep pockets and very technical leadership. Their weakness lies more with their customers, lack of energy, and their dependency on TSMC and the memory cartel.
People forget this is also a place of discussion and the comment section is usually peak value as opposed to the article itself.
Hence the relevance, maybe.
Maybe they are cheap.
whatever runs on typical investor/C-suite laptops and phones (so new iPhone/MacBook with "stock" Safari, maybe in corporate some cursed Windows setup with Chrome) is okay, and obviously they need to maxx out the glitter, it's the 2020s
probably they don't want this site to be scraped by LLMs which would be kinda ironic
Also, there is no mention of the latest-gen NVDA chips: 5 RNGD servers generate tokens at 3.5x the rate of a single H100 SXM at 15 kW. This is reduced to 1.5x if you instead use 3 H100 PCIe servers as the benchmark.
You can see them admit that RNGD will be slower than a setup with H100 SXM cards, but at the same time the tokens per second per watt is way better!
Actually, I wonder how different that is from Cerebras chips, since they're very much optimized for speed and one would think that'd also affect the efficiency a whole bunch: https://www.cerebras.ai/
Nvidia = flexible, general-purpose GPUs that excel at training and mixed workloads. Furiosa = purpose-built inference ASICs that trade flexibility for much better cost, power efficiency, and predictable latency at scale.
Edit: from comments and reading the one page that loads, this is still the 5nm tech they announced in 2024, hence the H100 comparison, which feels dated given the availability of GB300.
The reasons why this almost never works is one of the following:
- They assume they can move hardware complexity (scheduling etc, access patterns into software). The magic compiler/runtime never arrives.
- They assume their hard-to-program but faster architecture will get figured out by devs. It won't.
- They assume a certain workload. The workload changes, and their arch is no longer optimal or possibly even workable.
- But most importantly, they don't understand the fundamental bottlenecks, which is usually memory bandwidth. Even if you increase the paper specs, like FLOPS total, FLOPS/W etc. youre usually limited by how much you can read from memory. Which is exactly as much as their competitors. The way you can overcome this is by cleverness and complexity (cache lines, smarter algorithms, acceleration structures etc), but all these require a complex computer to run with all those coherent cache hierarchies, branching and synchronization logic etc. Which is why folks like NVIDIA keep going on despite facing this constant barrage of would-be disruptors.
In fact this continue to be more and more true - memory bandwidth relies on transcievers on the chip edge, and if the size of the chips doesn't increase, bandwidth doesn't increase automatically on newer process nodes. Latency doesn't improve at all. But you get more transistors to play with, which you can use to run your workload more cleverly.
In fact I don't rule out the possibility of CPU based massively parallel compute making a comeback.
Or it will get figured out in the niche fields where people are willing to figure out really hard stuff to squeeze out max performance (PE, hedge funds, intelligence)
Either way agree, it's hard to get mass adoption without the software ecosystem feeding back in
After I read the article :) The improvements in FuriosaAI's NXT RNGD Server are primarily driven by hardware innovations, not software or code changes.
