A nuclear bomb
"Based on this framework, we have achieved high-throughput, low-latency large-scale inference services for DeepSeek V4.1 Flash. Currently, the service throughput of a single cluster is 155 trillion tokens/day, and the overall revenue is $38 per card per hour (at off-peak pricing)"
It's really refreshing to put some cold hard numbers on this stuff, hopefully this quells some of the "AI is unprofitable everyone is spending a dollar to make fifty cents" arguments. Amortized cost of owning a DC full of GPUs is far less than $38/card/hr
Not to say it can't still be a bubble, just that there is legitimately profit to be had and perhaps some sustainable equilibrium in the future.
> hopefully this quells some of the "AI is unprofitable everyone is spending a dollar to make fifty cents" arguments.
I think everyone has known that inference is profitable overall for the major providers for a while (I'm sure not all user accounts, but in aggregate).
The major providers are unprofitable after you add in the training and research costs and other overheads.
This is why the open source models are such a threat if they become good enough - it will destroy the business model of just providing raw chat/intelligence.
Arguably that's also why a few of these providers are trying to diversify into whole-platform approaches rather than just relying on having their whole thing be a great model.
There has always been a lot of evidence that prices are only going down. The fact that you could already get DeepSeek or GLM inference from 20+ providers on openrouter.ai was already sufficient evidence that inference was profitable and being commoditized.
People that want to believe something different as a coping mechanism will not be swayed by this evidence either.
AI will end up being a commodity. That's also why you already see the first rounds of protectionism under the guise of national security. Obviously they do have a point but that point works against them just as easily. Anybody that integrates unreviewed model output into their workflow is asking for trouble.
non frontier intelligence will be a commodity. Frontier intelligence will be a high margin luxury good. this is how most markets work
Yesterday’s frontier is today’s commodity, so I don't think frontier AI is out of reach. With computing and solar energy costs continuing to drop, the current AI supply shortages are just a temporary blip. We’ll probably land right back in standard boom-and-bust cycles.
Either frontier AI will continue to get better steadily, until it's better than people at basically everything, and very weird things start to happen, or, it plateaus and the open weights models will eventually catch up.
There's a third possibility, which to me is also the most plausible. Models will keep getting better and better within certain domains, while never catching up with humans when it comes to other type of tasks. Like that elusive quality-without-a-name some have called "taste". Or knowing what UX is most intuitive to humans. Really, anything where we have to tap into our common humanity and our human experience in order to be of help to other humans and generally to life on earth. Which is an all-important skill, since the flowering of life on the earth is (or should be) the whole point of anything we do, after all.
We've spent a couple of millennia documenting that elusive thing called 'taste' in word, sound, image, lately in moving images. I'm pretty sure that even if you didn't have taste you could make a good simulation based on all that input.
I don't know, that sounds like a question of taste. :)
It can remix things people have already documented but can it synthesize an entirely new thing?
to synthesize an entirely new thing you have to start form the big bang. Everything else is derivative.
So what you're saying is that since there are so many examples of what humans have preferred in the past for these models to be trained on, surely there will come a time when they will know even better than us what we want and what we like.
... as long as they still care about what we want and what we like.
Everyone knows the path followed by every single technology mankind has ever developed - including the incredible run that was Moore's law. Sure, some will think "This time is different" - but is it?
It doesn't need to run indefinitely. I don't think anyone sane thinks it's going to run indefinitely. It just needs to run long enough. Moore's law was what, 50 years? If AI keeps improving at the rate it currently is for 20 years, the world will be very different when that curve flattens out, in a way that makes most of our current concerns are entirely irrelevant.
Those cycles have shortened considerably. Just look at the lifespan of media, integrated circuits, software, products. There have been so many big model releases just in the last three months I've lost count. We're now closer to weeks between steps.
Maybe if China stops releasing capable models the pace will slow down a little, but they have lots of incentives not to.
It is, because just like you mentioned, there's no exponential physics to back up growth. LLMs have to stop scaling up compute, because there's only so much resources.
But even if we think back to the previous decade or two (sans AI), most of tech has been a a red queen's race - developing and improving stuff, without any real substantial results.
Microservices and orchestration are not better than old monoliths. Electron apps are not better than native. Most programs people actually use have decades-old codebases.
Video games of today don't look substantially better or play different than ones from 15 years ago.
But it is not a constant point of "this % of people needs the frontier intelligence".
As thje models get better, the % of the pie gets smaller and smaller. So higher and higher cost of getting frontier model out gets divided onto smaller and smaller customer base as other models catch up
Yes, this is the crunchy bit and we're already well underway on that curve. I haven't touched a frontier model in months because I can do with smaller ones without wondering/worrying when I'll get banned next because what I'm doing is dual use. Qwen was the the first time I thought oh, ok, so this is 90% of what I need. Then DS4 made that 95% and with GLM 5.3 it's somewhere between 99% and 100%. Now, I'm not trying to solve any Nobel prize winning stuff here but if that's the kind of customer that will remain I don't think the frontier model providers are going to like the next six months or so.
That's the crux of the issue. You think that GLM 5.3 is all you need.
However, the expectations for AI are not to make your software engineering work 100% faster.
It's, for example, to build software more or less end-to-end with minimal input and feedback from you.
Then next, robotics and boosting manufacturing. Not by 10% but let's say 200%.
Such are the ambitions. Can GLM 5.3 do it?
The crucial question is whether eg. Fable 6 can, and with what lead over the competition.
Open-weight models still have not beat February's Mythos.
I am actually doing robotics stuff and yes, GLM 5.3 can do it.
As for the manual feedback: I want that manual feedback, but if I let it rip it will do just fine all by itself, but then I have no clue what is going on under the hood.
At some point you have to let go of trying for full understanding and control, delegate to the AI, and unlock the real productivity boost.
I have no clue about the state of robotics.
For software, I know that Opus 5.5 can handle more than GPT 5.6 Sol. I am willing to pay for the difference rather than use the cheaper DeepSeek Flash.
> At some point you have to let go of trying for full understanding and control, delegate to the AI, and unlock the real productivity boost.
The prerequisite here is that you can trust it not to do something that gives criminals access to your bank account or allows users to consume an unlimited amount of services it costs you money to provide for free or bankrupts you through legal liability etc.
It's basically this:
"A computer can never be held accountable, therefore a computer must never make a management decision."
Notice what this is actually saying. It's not saying that the computer can't generate a given output. It's saying that if the computer makes a mistake and you don't bother to check, you're the one who has to eat it.
> "A computer can never be held accountable, therefore a computer must never make a management decision."
That's a great way of putting it. So far I always used 'you can outsource work but you can't outsource responsibility'.
> "A computer can never be held accountable, therefore a computer must never make a management decision."
Legendary quote from an internal 1979 IBM training document [1]
https://simonwillison.net/2025/Feb/3/a-computer-can-never-be...
I'm getting plenty of mileage without that and anything with a motor in it requires that you do your utmost best to ensure it works within the design envelope.
We aren't going to get useful humanoid robots while understanding the implementation fully.
If you can't leverage AI to build a safe and useful robot without fully understanding the implementation, the model is not good enough.
There are so many assumptions in these two lines that I don't know where to begin. I wish you best of luck.
The humanoid robot was an example.
My argument is that if safety and liability make the use of AI without fully understanding the implementation impossible, the model is arguably not good enough.
A long time ago we decided as a society that in order for an engineer to sign off on physical designs, they have to take legal liability for them.
No such process exists for a model to take a design seriously that can kill people.
For serious engineering work not only physical designs, software also.
When we manually tested nuclear power plant controlling software, each test case was printed out on paper and tested and then signed by the tester. A lot of paper was produced and archived for each software release. (Archived for as long as the software was running in any nuclear power plant).
Everyone understood, that in a case of significant accident at the nuclear power plant, accident analysis would be done and if the root cause was an software error, this could have possible legal consequences for the developers and testers. (Of course nuclear power plants have many independent and redundant hardware controlling mechanisms, but incorrectly working controlling software could damage the power plant in a very costly way).
With a larger project, I do not think there is a requirement that the engineer signing off understands every implementation detail.
I imagine that the requirement is that after testing, you are confident that the product is safe.
Respectfully: you are out of your depth. Maybe pick another subject?
It was you who shifted the conversation away from software to make an argument that "anything with a motor" requires full understanding of the software implementation, and that for this reason GLM 5.3 is good enough.
I think you're entirely wrong, and we can leave it at that, seeing that you're clearly unwilling to engage constructively.
The practice of deliberately engineering software without full understanding of the implementation, especially when it comes to robotics, should be illegal. Technology serves society alone and it should progress at a pace that is sustainable for adaption. Your excitement for a “singularity” does not outweigh someone else’s health and safety.
Nah man, no you don’t. There are very few absolutes in the world and that certainly isn’t one of them.
And over a increasingly small slice of time, as, due to the commodification of these things, the frontier businesses will see their customer base walk to the next one that does any better on their task.
This is probably ultimately unsustainable.
Every time they throw someone out with a message about how this request is definitely against their safety guidelines they are clinching the deal on someone running something local or switching to a provider that just lets them get on with their day. Their own fearmongering is going to be part of their undoing.
A high margin luxury good with huge r&d costs is a death spiral.
A high per-user r&d means high costs per user. So then users leave to save a buck, qnd the price goes up again for the remaining users.
Maybe it works out OK if you want to signal wealth - the high price is the selling point. But luxury pure utility is a hard sell to consumers, and for luxury B2B it's about high touch sales (e.g. Oracle).
Agreed, but I I wonder what fraction of tasks really require frontier performance, either in terms of sheer difficulty or where speed is a matter of necessity. For many jobs, it's increasingly a question of whether you'd rather have one filet mignon or 20 good cheeseburgers (all from the same cow).
I’m not sure if this counts as a nitpick, but the definition of frontier automatically includes the cheapest models
Who can and is willing pay top dollar for frontier human intelligence, and fully utilize it to actually make money?
The only answer I could come up with is hedge funds. Everyone else fails some part of my criteria.
The models will be a commodity, the resources to run the frontier will be the luxury good. This is already the case nearly.
The market for "frontier intelligence" will be too small to serve outside of certain government/military applications that exist in spite of market forces, in the same way that the market for small, portable nuclear reactors is exclusively limited to submarines and aircraft carriers. When I call a plumber, I don't need a frontier plumber, I just need someone who's baseline competent.
> market for small, portable nuclear reactors is exclusively limited to submarines and aircraft carriers
SMRs are being commercialized. Likewise nuclear cargo ships have been seriously studies. Nuclear reactor powered (not just RTG) satellites have been built.
I know this is beside your point, but I think if we figured out how to commercialize this, demand would come.
> I just need someone who's baseline competent
But why wouldn't you want (and pay) for the very best plumber? A frontier plumber will optimize the time their customers poop, pee, bathe, or do dishes 10x to 200x, and with their freed-up time,they'll outcompete those who don't pay for frontier plumbing!
You forgot the /s
Sir, I take the impending plumbing singularity very seriously, and wouldn't want anyone to be left behind. When plumbing hits recursive self-improvement, we may see toilets that shit themselves more efficiently than any man can, and finally attain ASI (Auto-Shitting Infrastructure).
I can see you do. I beseech you though to think about the consequences for other industries. For instance, the ratio of toilet paper:organic product will need to be maintained somehow and I suspect that if we follow your pathway to its logical conclusion that there will be a cellulose crisis unlike anything seen in the world since the USSR went bust.
The next move in my investment strategy is obvious though, invest to the hilt in PVC tubing and I thank you for this invaluable insight. Fortunately this forum has no visitors so we can rest secure in the fact that the pigs won't realize it is too late to avoid being slaughtered. Power to the bulls, after all it is their product more than anything that is responsible for the plumbing revolution.
there is definitely an asymptote
Non-frontier intelligence has become a commodity already.
And the success of things like ESP32 shows that good price + good enough is a powerful combination.
> AI will end up being a commodity
what does this even mean, the commoditization... like a thing you can buy from a gas station? AI commoditization is the rock bottoming of pricing.
Usually we end up with cartels in such cases, see OPEC, mafias etc. But we have China throwing sticks into the wheels of the US providers that are attempting to make the cartel and "pacing the frontier"
> Usually we end up with cartels in such cases, see OPEC, mafias etc.
Whether a cartel forms is independent of whether a product is a commodity. To the extent that commodity markets are more likely to form a cartel, it's only because it's harder for them to form a monopoly.
The general problem is that if your legal system allows competitors to buy each other then they have the incentive to do that instead of competing. This is why there are supposed to be antitrust laws, and why we get problems whenever they're not actually enforced. But in a typical commodity market, even buying the competition doesn't quite work because someone new can always enter the market. Congratulations, you've cornered the market on landscaping services... until some unemployed fellow who already owns a lawnmower and a truck starts advertising their services.
OPEC is kind of a special case because the product is entirely fungible but it's also finite and consumed by usage. Landscaping services aren't meaningfully finite, if demand increased then supply could match it under any plausible scenario. Copper is finite but it isn't destroyed by usage, so if someone had a lock on copper mining then people could still get it by recycling existing products.
Whereas oil is finite but then the end customer burns it. If Saudi Arabia cuts production, Germany can't make up for it by extracting more domestically because they haven't got enough in the ground, but they also can't capture the exhaust from their cars or power plants and turn it back into petroleum because the product is really energy and it takes more energy to do that than you get from burning it to begin with. So if the largest exporters get together, they control enough of the market to control prices and there is no real way for anyone else to enter the market with enough production capacity to change that.
Until you realize that the product is really energy and you can get out from under their thumb by switching to electric cars and generating electricity without burning fossil fuels.
A commodity is an undifferentiated good. I expect gasoline to be practically the same at each provider. AI will be similar.
The funny thing is the most lasting contribution OpenAI makes to this field could very well be the protocol for the endpoint.
Not too long in the future, I fully expect OpenAI to rescucitate the API copyright aspects of Google v Oracle to pad its revenue in OpenAI vs Everyone profitable
I wish them good luck. I don't doubt that they will try.
China is doing all the heavy lifting for the free market in AI right now. I did not have that in my 2020s playbook.
> what does this even mean, the commoditization...
Like a utility. Not 'ACME intelligence' but just 'Intelligence'. With multiple providers selling you intelligence that is 'good enough' by the token.
>Usually we end up with cartels in such cases, see OPEC, mafias etc.
No we don’t. That’s quite rare.
Look at the collapse of the solar market around 2010 as the price collapsed.
Look at the collapse of the steel industry in the US as outsourcing to Asia took over in the 70s-80s.
Look at the entire computing boom from 1990-2020. Memory and storage followed such a steady line down and to the right that people are struggling to operate compute businesses in a normal market.
E-bikes, scooters, electric cars, lithium ion batteries in general. Most food.
The list of these commoditized things that started out as areas of innovation and became just the marginal cost of production is enormous.
I’ve never heard anyone say raw inference was unprofitable. It wouldn’t even make sense, I can’t see a rug pull for renting out commodity GPUs.
People are saying _training + inference_ is unprofitable (I don’t have a dog in the race and don’t really care either way). That’s a lot of GPUs and salaries that aren’t generating revenue at any moment.
Personally, I would guess Anthropic and OAI are profitable. I just suspect that they’re a pretty normal level of profitable and not the absurd, near-alchemical “can turn dirt into gold” level of profitable their valuation seems to have priced in.
To add on, training + a highly paid American salary load + inference is unprofitable, if and especially if you have deepseek dropping weights and engines for anyone to run.
I don't see the average person being able to run DeepSeek's most competitive models though.
H100s are renting out at a few dollars per hour. An awful lot of people could run them if they wanted.
OpenRouter's DeepSeek provider is basically this. Instead of renting H100's for dollars per hour, you pay someone that arbitrages the $/hr to $/token conversion because you can't actually saturate an H100 for hours. It's still effectively the same as running it yourself because of the ability to change providers to someone else arbitraging the same thing, and that there's always the escape hatch of switching to actually renting the H100s yourself.
They don't have to, slightly above-average people can host in Asia for way cheaper than anthropic.
Why would it be cheaper? Electricity might be (slightly, possibly), but H100s aren't
It's a commodity, and research into how to make it good will be done in the open. What this means is that we should doubt the justifications offered by magnificent-N-whatever companies slurping up all investment dollars to take over the world by being the first to get "general" "super" intelligence. No, there won't be a runaway improvement. Sadly, AI is just a clone of our inner world. Getting interesting information into it is still going to be just as slow as it is in the regular world. You have to build, tinker, experiment, get data, get knowledge, etc. The only thing getting sped up is how much time needs to be allocated to thinking and translating.
there could still be a runaway improvement, but it also almost certainly goes with the decreasing prices
Isn't the modern economy already a runaway improvement? First writing and agriculture, the stone age became copper became bronze became iron became steel, and one Revolution after another - agricultural 1 and 2, industrial, information.
Looks like AI will be another revolution, and eventually, probably the largest.
There may be different thesis out there, but the fact of prices going down definitely doesn't help the case that the major AI labs are are overvalued and need to vastly increase their revenue to stay afloat amidst their contractual obligations.
I honestly don't think there are too many people arguing "AI is actually too expensive to take my job". "AI being too much of a commodity to make meaningful profits" could very well be the case though.
I have seen the former argued many times, though mostly not on HN. It is a cope meme I’ve seen mostly on Twitter and Bluesky, but some on HN and the Elixir forum as well.
They are literally hoping Anthropic & OpenAI go bankrupt (totally possible), and everyone goes back to using their meat brains for work (never happening).
In this context it just doesn’t matter if these companies are profitable after amortizing training cost, because deployment of open weight models that already exist will continue to disrupt economies for at least the next decade if all research ended tomorrow.
> The major providers are unprofitable after you add in the training and research costs and other overheads
The very short lives of all these models also factor in.
I think the funny thing is the more inference, the more tokens a specific version of a model spits out, actually does make it more profitable.
But somehow a lot of people truly think that each token generated increases the cost of OpenAI/Anthropic/... and shout bloody murder about "subsidies"
And I'm certain there are people out there that actually think the API prices of OpenAI and Anthropic somehow equal their electricity costs.
Nobody meets expectations of a 2 trillion dollar IPO selling a fungible commodity like tokens for old models.
And that's the rub - it doesn't matter if inference is profitable, the US labs have promised the whole kitchen sink, and taken serious investment on that pitch. They won't make it back in a cut-throat market for inference. They have to continue pitching the idea that only their models are worth using, for quality/safety/performance/etc reasons.
Which means continuing to develop and release new models. At huge expense.
Side note - while deepseek seems to be doing fine, I'll note that the emphasis on cost effectiveness is very much not a thing for the US labs right now... and it was for non-us labs. I'd be quite interested in knowing if anthropic is pulling in the same returns, or what the margin on their subscription plans actually is.
---
I'm already seeing this directly at work - lots of devs still happily use claude for development, but the products we're making aren't. They'll use kimi, glm, gpt-oss, etc... basically - if cloudflare is hosting it, we're exploring it (mainly for contractual/security reasons): https://developers.cloudflare.com/workers-ai/models/.
Because we are price sensitive, and we already have to do all the work of creating very robust evaluation frameworks for our products, so we might as well choose the cheapest model that meets our standards.
whiiiiich is why we're seeing things like Haiku and gpt-luna priced down near the open models at ~.10/mm in, ~.50/mm out. They've been forced to cut prices to keep commercial customers on their cheaper models instead of moving away.
This will continue. If inference is profitable and easy, prices will continue to erode. Go browse openrouter, plenty of providers happy to undercut the competition. Bigger shops with auditing requirements will avoid them, but smaller ones won't.
You know, it's not just cost-effectiveness, it's also iteration time and speed. DeepSeek V4.1 Flash is astoundingly fun to use because it's so quick. Tens of times faster than the standard Frontier models. With, in my experience, practically the same results. Because most of the trouble is not the intelligence of the model. It's the feedback from the world. The quicker you can get feedback from the world, the better your process is, irrespective of raw intelligence metrics of the AI you're using.
I 100% agree and am a big fan of DS V4.1 Flash in general for this reason, but I'll add my own caveat: Sol 6.1 isn't far behind on completion time for tasks because it's so damn efficient. I regularly get ~300 tokens/sec with DeepSeek Flash, but it spends the majority of that time thinking and executing slow commands one by one, whereas Sol thinks less and is much smarter about the commands it runs, resulting in far fewer turns to get the same thing done. For highly agentic tasks, DS might only be ~25-50% faster, not 500% as the TPS may suggest.
(But to be clear: 25-50% faster is still a big win in my book.)
+1; using Sol 6.1 in Ultrafast is noticeably faster than DS 4.1 Flash in real world tasks where marginal intelligence and token efficiency matters. Yes it costs more, but that’s a function of how valuable is the thing you are iterating on.
> Which means continuing to develop and release new models. At huge expense.
I guess my question is when do we say a model is "good enough"? For my use cases Opus 5.5 feels as if it is exactly at that point. And at that point of course one of the main factors I will evaluate against is price. So far I am happy with my (heavily subsidized) subscription, but if the time comes, I'll happily switch the inference API any of my agents are calling. So far I have no big lock in on any providers.
define "we".
Because again - I think lots of businesses are already taking a good look at models well off the "cutting edge" because they're good enough for what that business is trying to do. They're already there.
If you're doing categorization, tagging, form-filling, customer support routing, etc... you don't need anywhere close to cutting edge expensive models.
Hell - Jev being so popular is basically exactly that group going "what's the cheapest way we can make this decision".
Meanwhile, there are plenty of spaces where you still want real model improvement. But a lot of traditional business was basically process and paper pushing. You don't need a Phd level model for that, you need an intern level model.
---
But that's the trap for the large providers - intern level models are cheap and easily available with current open weights. It's going to be a very tight market, and margins will go down to essentially the floor (especially once we exit the current constrained-compute market).
Somewhat ironically - the technology itself makes switching costs basically zero. It's not just that there's no big lock in, AI is like an anti-moat. It actively fills in and paves over whatever moat you might have for price conscious customers.
I think we share the same view here, which I think is most widely discussed as "AI capability overhang". Like you also said, a large, growing amount (digital) tasks can already be completed reliably with existing models. For those - this was my point - the models are competing on price, speed, reliability, etc.
I also agree that more capable models are still desirable in plenty of spaces. I just see that there is so much lag in the diffusion of existing models, which, again, are already sufficiently capable.
When it solves your problem. I don’t think it makes sense to globally qualify a model as “good enough” any more than it does with programming languages. Python is good enough for a web app, but not for an OS kernel.
To your point, I think we’re past the hump on the Pareto distribution. The models are good enough for most things, and the rate they’re improving at the things they’re bad at is dropping. The odds a new model is good at something the old one was bad it is okay, the odds it’s something you care about is dropping. It makes a lot of sense to discriminate on price rather than features in that environment.
"640K ought to be enough for anybody"
I know this gets a lot of flak because it ended up being obviously incorrect for consumer markets.
But the under-reported corellary of this is that a TON of industries essentially operate on hardware right out of that time period (or slightly after, circa 2000s)
banking
Insurance
Healthcare
Gov
retail
Transportation
etc...
Hell, I was writing code to run on IBMs Z/arch as recently as 2021 for banks with mainframes deployed around 2004. They were just then finally considering an upgrade to the later models.
The Chinese labs are just arms of the Chinese government. They have no obligation to be profitable or please (private) investors. If they weaken/collapse American models, then that is the ROI to the lab owners.
This is a pretty strange take on an article explicitly pointing out the profitability of chinese models.
Further, even if this is your take (and I won't touch the geopolitics here, maybe you're right, maybe you're not regarding ulterior motives) - go use non chinese open models that are also good.
Laguna, Nemotron, Mistral, Arcee, Llama, etc...
China has sandbox capitalism. In the sandbox it looks and feels like capitalism, but the sandbox itself exists in a communist framework, and the state still controls all the inputs, outputs, and can intervene at any point as they please.
In organic capitalism you need to compel arrays of discrete private entities to invest in your idea. In sandbox capitalism, the money can just appear in any amount at the behest of a single controlling individual.
Hyperbole and trappings aside, that describes all mixed-market economies. And all major states have champion industries.
Now do American Capitalism!
You can see it plainly on the Solar PV market. The likes of Trina, JA, Jinko, etc. all compete aggressively, leading to everyone running at a loss in order to corner the market, and then after a few years the government steps in and orders all of them to raise prices. Happened again at the beginning of this year, raising cost of PV Modules 30-40% for oversees consumers.
Contrast that to US being happy with everyone eating each other alive until all but one company is left standing and can now dictate the prices and enshittify the product to pay back their VC loans.
Isn’t “sandbox capitalism” - real capitalism and what the us is trending towards is the “Wild West” capitalism with monopolies, trusts, and all the other anti competitive things that arise in the free market if there is no external force that tries to keep the market “fair”?
I mean this is not the case for either, but the systems are setup to drift in these directions - a market needs some sort of regulation to keep things sane, otherwise you end up with standard oil and the like.
You want the people with guns controlling things to be at least nominally interested in “the people” and produce more than just “return on investment”.
Not saying that China is like this or that US is bad, but saying that China is bad because it is “controlled capitalism” is kinda missing the point of what it should be in the best case.
I don't think its that hard to shoehorn current US into those sentences too, maybe switch word 'communism' with 'kleptocracy'.
Less intervening and micro management but oh boy say something not nice about government and potus and suddenly you are a new enemy of the state and 3-letter agencies work day and night to bring you down for good. Or in reverse - put a public bribe to potus and his clan and enjoy some favors for few weeks.
2 years ago I would agree more strongly with you, but we are in new era from now on.
to be fair they did say 'organic capitalism' - which I interpret the same way as 'true communism' in that it probably cannot really exist.
real-life, unregulated capitalism lends itself to exactly what you described
real-life, unregulated capitalism lends itself to exactly what you described
In actual capitalism, POTUS wouldn't have enough power to be worth bribing.
The problem isn't the money, the problem is the power.
nice
> I think everyone has known that inference is profitable overall for the major providers
> The major providers are unprofitable after you add in the training and research costs and other overheads
It's meaningless to say that selling something is profitable if you are ignoring the cost to produce it.
The profitabiity (or not) of AI models needs to be calculated on a per model lifetime basis - what did the model cost to develop and train, what is the lifetime revenue the model generates, and what is the lifetime operational cost of the model (including amortized GPU cost), etc.
It makes no more sense to calculate "inference" profitability when ignoring training cost than it would to calculate it when ignoring GPU cost.
Exactly. And you can't just stop training models, because your competitors won't, and customers are switching providers on a daily basis (most likely an exaggeration).
Hence "we need to slow down progress for safety" [narrator: it had nothing to do with safety]. The industry is begging for the regulation so that there is a coordinated slow down and consequently move towards profitability.
> coordinated slow down and consequently move towards profitability
100%
Amodei has been promising the "end of the exponential" all year, when they can shift some compute away from cash-burning research and training toward cash-generating inference.
"Pacing the frontier" is how you collude with your competitors to all decelerate at the same time because safety.
I saw last week that Anthropic is currently cash-flow positive with 75% of compute going to research and training. If they can flip a switch and move from 3:1 to 2:1 or 1:1 training:inference, they will be insanely profitable.
As long as they can avoid a price war.
>> calculated on a per model lifetime basis
True only if you never decelerate R&D.
But "pacing the frontier" means every early model doesn't need its own profitable P&L.
Early models can be big money losers while helping build the next generation.
If the industry can coordinate to all decelerate at the same time, then they'll flip from incinerating billions to minting billions, and all that capex will suddenly look completely sustainable.
Unless DeepSeek suddenly gives it all away for free, which is why the big three need government regulation to prevent foreign models from busting the cartel.
> It's meaningless to say that selling something is profitable if you are ignoring the cost to produce it.
At the same time, training costs don’t scale with usage, so no matter how expensive they are, they can be paid for as long as inference is profitable. It is only inherently unsustainable if you lose money on inference.
Of course, it could be that the inference usage required to pay for the training is too high to be possible, but that is a separate argument.
It doesn't just matter that inference is profitable - cash from profitable inference needs to outweigh cash burned by training. If training costs even $1 more than your profits from inference on a model, you lost money on the whole thing.
> It's meaningless to say that selling something is profitable if you are ignoring the cost to produce it.
Don't enter that rabbit hole, you'd be surprised at how questionable are some accounting rules and practices, even in publicly traded companies.
if i can run an existing open model because its good enough for what most people need, I can just spin up an inference business, and put my spend either on trying to optimize it further or on customer experience
... and not worry about training cost at all.
a lot of the cloud businesses of the last decades have relied heavily on open source code, again skipping most of the cost to make the product/service being sold.
>I think everyone has known that inference is profitable overall for the major providers for a while
Have you seen any place that is not as tech-savy as HN? e.g. some Reddit subreddits?
Barely anyone is aware other than people who are already somewhat deep into this whole topic.
The insider baseball in AI is extreme I agree. I still know people who insist large language models are just glorified autocomplete.
Frontier models are still transformer based, so that's still true.
I believe they're referring to the user experience, lots of people think you just use AI to finish a sentence a la early Github copilot. The average person hasn't seen what agentic flows can achieve
Yes. I've always suspected inference is profitable.
But I'm also interested is retraining needed? how relevant will a model be in N years if it's just "static", and if you do need to retrain and fine-tune over time, is it still profitable? how often is that needed?
For most cases it's probably ok, often times answers are general or basically a web search anyways, but if my Agent started blasting outdated syntax it would suddenly be way worse for coding for example.
It’s helpful for a model to be somewhat static when it’s involved in business processes.
Relying on an LLM's baked-in syntax knowledge is suboptimal, compared to leveraging reasonable context affordances (skills, examples, tool calls, MCP servers, search) to ensure adherence to precise APIs and language specs. Don't get me wrong; a model specifically trained on your desired framework is obviously preferable for more consistent generation of idiomatic code -- but relying on that in a large general-purpose model is a questionable proposition IMHO.
They need to be retrained often is my experience for general usage.
I have SO many cases, for example when shopping, when they make assumptions based on their cutoff date, and they need to redo everything based on lots of searches.
Sure, for many use cases old models will be useful but for general usage a model that's 5 years old will probably be garbage for most people, unless they prompt it extensively, which the general public won't do.
> But I'm also interested is retraining needed? how relevant will a model be in N years if it's just "static", and if you do need to retrain and fine-tune over time, is it still profitable? how often is that needed?
Just wait until what Ilya Sutskever has for Safe Super Intelligence (SSI) to self update the weights without retraining them to bring training costs down close to $0.
> I think everyone has known that inference is profitable overall for the major providers for a while
A lot of AI haters like Ed Zitron and their readers have refused to update on this. I don't know if any of them have yet, but it'll take a year or two for the idea to fully filter out of their podcasts anyway.
Ed Zitron is a professional apocalypse predictor. Which is to say he's always wrong.
Seems a bit of a disingenuous framing. Most of the arguments I see are some mix of
1. Their figures have been pretty dodgy or excluded training costs so why is it trustworthy this time?
2. API inference charged at a profitable level rather than heavily subsidised subscriptions means everyone tightens up their usage + governance Which we very much saw with the rise of openrouter et al.
Personally I think there's a lot of the bait and switch of 2 when people talk about profitability of the 10-20% of API usage for Anthropic while also wanting to lump in subsidised chat users as if they're profitable too.
I think we know if it was this clear cut then the behaviours of companies would be wildly different and more consistent with profitability.
There are thoughtful versions of these arguments, sure. There are also thoughtless versions, and it's not disingenuous to note it.
It's not actually clear that subscription users are unprofitable as a group (and the group level is the only important one).
We do know the companies are not currently profitable, because they're open and intentional about that, but that tells us nothing about their margins without infrastructure buildout and investment in training.
It’s not training and research that kills you here, it’s your own investors ROI expectations that murder you cold when faced with Open Source commoditizing you, from a country that spent 5% of the sum you spent to get to the same position and then gives the recipe away to everyone.
> I think everyone has known that inference is profitable overall for the major providers for a while
Unfortunately while in the tech industry some might know that, the general public are still being misled (especially by certain folks who’ve made a living off downplaying AI) to believe inference itself is unprofitable.
> The major providers are unprofitable after you add in the training and research costs and other overheads.
This is correct for now, but I recently read that 90% of Anthropic's code is written by Claude Code itself. So, if an AI is using its previous version to train the next version, we're not far from the day where these models will take up the decision making process too. For eg: Decisions API by OpenAI or Jev is based exactly on this. What if a model like Jev sits behind Opus 5.5 and guides it to take the next step? Wooh.
Probabilistic predictions right now are at 99% for frontier models. They will never be 100%, but if humans rate themselves 99.9% in decision making, that benchmark can be broken.
As far as I know they haven't been doing any mass layoffs.
So if Claude Code is writing 90% of the code, then why are they keeping probably the most expensive engineers in the world on the payroll?
The machine writing all the code seems to not be enough, even for the company with access to the best models. What's the alternative to explain this discrepancy?
Frequently it's to keep competitors from having them.
In this game you "win" only if you arrive first, so keeping the same headcount to run faster is a valid option for Anthropic. In this context "winning" means mainly that we lose our jobs
Honestly that's an interesting statement because in my experience Claude writes around 5-10x more code than a reasonable person would to accomplish the same thing. If you include comments then that number goes way up as well.
As someone who has long bought into the idea that the labs are not profitable on inference, I have a question. How is it simultaneously true that there is more demand for inference than supply (apparently by orders of magnitude), but also that inference is so cheap that labs can make healthy margins? It seems like the inference providers, not the model providers, should have the leverage. Why haven’t they said, “give me a bigger cut of each token or I won’t serve your model”?
Also, what happens when the inference providers suffocate on their own debt burdens? Who will serve inference?
It's a few things I assume:
1. Inference providers are largely a commodity business. While demand outstrips supply, there are many providers of that supply, which does somewhat constrain how much leverage a provider has.
2. The models are not interchangeable as of yet (though this seems to be changing). A lot of people want the latest Anthropic or OpenAI model; as such, the labs still have considerable leverage.
3. A lot of the inference providers have used their leverage by doing equity deals with the labs, agreements to have some of the deal revenue spent on compute, etc.
What's the evidence that demand outstrips supply? Is there a wait to get an inference provider or interruptions in availability?
GPU prices are sitting well above MSRP, even for older generation cards. Datacenter growth is also being pushed pretty hard by most of the cloud providers. More anecdotally, the model providers I use have been "throttling" my requests (usually by bumping them down to a lower model version) more frequently. Granted, that could be for profit reasons, but I think it all generally points to a lack of capacity.
>Is there a wait to get an inference provider or interruptions in availability?
I think that the inference providers and model vendors are mostly handling it behind the scenes with more aggressive quantization, increasing API prices, deprecating more generous monthly tiers, etc.
this info is dated
Yeah I think one of the less considered costs when people talk about this is the human cost of training. We all talk about the billions spent on compute, but I don't think anyone's got good numbers on how much RLHF is costing but it's got to be pretty large.
The only moat is in datacenters. The software has no moat, since 1) it's already trained on stolen data, and 2) someone with access to AI will be able to reverse engineer a proprietary implementation eventually, the same AI is being used for developing AI.
That's why so many talks about regulations. Regulations means: ban outside competition, just like US bans Chinese EVs today.
regulation is braindead
I would classify training and research more as capital investment than overhead.
Most people in my experience refuse to believe it for some reason.
Well, as that means there's lots of money to be made lowering costs for research and training! What tools have we used in the past to lower those costs? Automation.. Hm, what if we had AI completely automate AI research? Uh oh, we just rationally found outselves in a trajectory where we lose control of our own trajectory because of economic incentives.
It is wild how fast we're hurtling towards this outcome, lmao.
I doubt that, I got downvoted several times pointing out inferencing is insanely profitable.
Even take account of R&D cost it is still profitable, because you do not actual retrain the base model every month, only distill and post train from the same foundation.
heck, I think even including the giant cost of the top tier researcher salaries, they are still profitable.
then, how are they still losing money? just read what they are actually spend the most budget: always the hardware. those companies will never profitable, because they will never stop spending money to expand their computational prowess.
I don't think so. The number of comments I see from people talking about AI, there are plenty who still think the price is going to shoot up. Some even talk this way about models you can run locally on high end yet still consumer level hardware.
I think the more someone works with AI the more likely they realize that inference does not have to be sold at a loss, but many people's AI knowledge seems quite dated and it sounds almost like it becomes a convenient conspiracy theory for some, a way to convince themselves that the market changes that are already happening a temporary (and I don't mean in a Jevon's paradox sort of way, but that companies will go away from AI).
I suspect a lot of this is fueled by a few stories of companies who went all in on spending as much as possible on AI and then begun to walk that back, without realizing that was more about bad incentives and over promises of what AI could do (especially 1/2 model generations ago). Some also confuse running the SotA models at the fullest capacity costing more and more with a given problem costing more and more to solve, not realizing that the cost per problems is trending down but newer and larger problems can be solved for even more, so the costs at the top are rising despite falling prices. Also things like the subscriptions getting more expensive, though how much of that is because developers are doing far more than they were 6 or 12 months ago (and in some cases there was subsidization that is going away). And maybe a bit with the free models starting to not be upgraded or the harnesses getting worse performance of lowered free limits.
If the frontier labs are struggling under the training costs and the leading open source models are making their progress by distilling from the frontier labs it sounds a bit like there’s something wrong with the economics here or that progress is going to slow quite a bit without the amount of investment.
A lot of people have already replied to this point but I think I want to distill the problems a little more directly:
- If inference is this profitable, is it profitable enough to offset the enormous training costs, bot those already incurred and those in the future?
- If inference is profitable, can it cover other capital and operational expenditures the hyperscalers and neoclouds are committing to, on the timelines they're committing to, to scale out inference?
- If inference is profitable we can measure if the price/earnings ratio of current valuations is sane. Is it?
- If inference is profitable, is it - given the answers to the questions above - profitable enough to make the risks of lenders of the huge amount of debt companies are taking on to scale, rational? It's one thing to say as a hyperscaler "we can make this pay", it's another to say "I'm going to lend them the money to do that and I'm going to get my money back on the timescales we agreed, and with an ROI to boot".
- Finally, if inference is profitable why is Nvidia specifically lending money to companies to buy Nvidia cards to do that inference, and then hitting the credit default swap markets to cover the risk of never getting that money back, and doing so at ever increasing cost, when they could just go build a profitable inference business themselves?
I agree seeing the numbers is good. It does suggest profitability in some segment of the market. However I still have a lot of questions.
But you are mixing up different categories.
Most of your questions are things that are relevant to everybody due to a possible economic crash. And you are right to point this out.
But for me as an AI user those aspects are not really relevant in my decisions. If an AI bubble bursts providers will still want to sell me those tokens at that price and yes development will slow down for a time which does change not a lot for me.
Inference is very profitible when you can deal with the batch sizes Anthropic, Deepseek and OpenAI are dealing with, renting GPUs for inference is even more profitible.
The US is dealing with different tax code where buying GPUs function as as great way to offset taxes, and having boatloads of debt is rewarded in many cases. I think there's lots more (crooked) reasons why inference is even more profitible in the US.
100% the right questions. The implied thought process of so many on this thread: Inference profitable >> Frontier Labs' are profitable overall. Sigh...
I can't read the original Chinese text, but if it's revenue, then you still can't say anything about profitability. Housing and powering one GPU should be way less than $38 per hour, so there should be a profit on inference. If that profit can cover training new models we don't know, but it would appear to not be the case.
For the past year or two I think it has generally been agreed that inference is profitable, but the LLM business as a whole is not.
I seem to recall the leaked Liang Wenfeng speech gave a figure of 8-12(?) months of inference recouping the training costs of the model.
I assume they need these models to stop being replaced in 1-2 months though.
I think what he said was that 10 months of inference paid for the hardware it runs on (which may have a lifetime of a few years). I don't think he mentioned training/development cost.
You are correct, sorry. I should verify stuff like this before posting.
It's really hard to unpick, for example, energy prices might be heavily subsidized for the Chinese models.
But it's clear that it's not the money burning pit I think it was just a year ago.
China has a two tier energy system.
Consumer get the cheaper energy, while industry normally get the twice as expensive energy.
But for AI/Datacenters, there is a exception and they get the cheaper consumer level energy.
You need to take in account that the way China taxes is not the same as here in the West. And a lot more is put on the local governments regarding generating their own revenue (why there was so much building in China because those projects funded local government). A bit of a side topic lol. Its not so simple as they are making a loss on something because how its structured.
AI gets the cheaper tariffs then the rest of the industry, that is a fact. They simply have less energy profits on the AI energy. Now if you consider that subsidized or not, ... in both cases China is still making a profit.
What is overlooked is the massive amount of energy generation that China has and still keeps expanding. And that is a direct reflection upon why the prices for industry and consumers are so cheap. This buildout ironically had nothing to do with AI but was more about lower energy = cheaper industry production ... and a lot of buildout was also to build more independence from energy imports. They are not just throwing solar and wind turbines around like candy, the amount of reactors constructed and still under construction, ...
> China currently has 36 to 38 nuclear reactors under construction, maintaining the largest nuclear buildout in the world and accounting for roughly half of all global nuclear plant construction
So yea ... The AI in the west have a issue because it really does not matter if you cut China from Nvidia GPUs or advanced nodes, they compensate on the power level.
Edit: And a little extra, the irony is here in the West building nuclear is expensive as hell because we barely do one or two at the time. With a lot of lost expertise. China doing tons of them non-stop cuts down on the costs a lot.
China generates a lot of electricity but from what I understand prices aren't that much different to in the USA because of grid constraints.
I'm not sure how much you pay in USA. But I live in Beijing, where the residential electricity price is about 0.5 RMB per kWh, roughly 0.0749 USD.
China's electricity pricing involves cross-subsidies. Residential and agricultural users pay the lowest rates, which are generally below the cost of supply. Commercial and industrial users pay more and contribute the most of grid revenue.
Prices vary by region: commercial electricity can approach 1 RMB (0.15USD) in Beijing, but the national average is about 0.67 RMB (0.1USD), while industrial electricity averages around 0.63 RMB nationally.
And that half a RMB doesn't include extra fees or delivery rates vs supply rates, correct? It's the all in rate?
I'm not quite sure what you mean by "extra fees or delivery rates," but this is the total amount I pay for electricity, no additional charges.
In my state in the US, they started differentiating between a "supply" rate, as in who generated your electricity, versus a "delivery" rate, or who delivered the electricity. That way they can market the supply rates to you and charge you on delivery additionally.
That sounds a little complicated, do the two parts actually adjust dynamically?
In China, end users pay a single bundled retail electricity price that already includes the full cost of both generation and delivery. The State Grid company purchasing electricity from generators and settling with them centrally.
Electricity prices are also government-regulated, and any change must go through public notice and approval.
Supply rates can adjust dynamically but not delivery, though delivery rates have been creeping upwards as well.
At similar prices and 1/4th or whatever the average salary, meaning the average person has to use a lot less energy, or spend a larger % of their budget on it
>Consumer get the cheaper energy, while industry normally get the twice as expensive energy.
Interesting. In Germany, industry gets the cheaper electricity...
It's not particularly surprising. China wants people to use electricity. Germany wants people to use gas/oil. The rate of electrification in respective country reflects that preference.
I don't think that the steering effect (away from electricity) on private people is intended. It's intended to keep industry in the country. Makes sense, but still, it feels unfair as a private person.
> China currently has 36 to 38 nuclear reactors under construction
Wowza! Thanks for all the info really helpful to have insights into how it all works.
Sales tax in China is attributed to the place of production rather than the place of sale, which is why provinces want to produce as much as possible
as if electricity for openai & anthropic isn't heavily subsidised
as if electricity for spacex wasn't literally illegally installed gas turbines polluting the air
Would that be due to the size of the models from DeepSeek or perhaps only for their distilled models? If it's for general purpose models that would crash the valuation of Anthropic and OpenAI, both who are struggling with profitabilitet to say the least.
Who said they need to distill anything? Deepseek is the most popular chinese model getting trillions of input tokens a day that have no ZDR availability. They specifically tell you that part of the deal is that they get your data to use however they want (atleast they're honest).
This idea that Deepseek needed to distil their way to success is nonsense. Its a great narrative for OpenAI and Anthropic to spread because Americans have been brainwashed for two decades into dealing with their own inadequacies by blaming things on scary asians stealing things.
In reality China is outworking and lapping most American companies. They produce 10 EEs to every 1 America produces. They are simply harder workers.
The West is cooked because we're too lazy to work hard and we'd rather make up stories about boogeyman who are stealing from us. The only thing we're good at now is killing people who economically challenge us.
DeepSeek's API used to (does still?) route your requests directly to Claude Fable[1]. Those were provably Claude Fable's responses being returned by the DeepSeek API. This was months after Anthropic accused them of distillation of Claude at scale[2].
Them provably distilling after being accused of distilling strongly indicates that... that's what they do. To spin a story of them not doing that (because why would they, with their superior talent and all) doesn't seem very reality conform.
The models are impressive enough as is (considering resource constraints and difference in CapEx) despite being heavily distilled from western frontier models. No need to pretend that they're not.
[1]: https://news.ycombinator.com/item?id=48966016
[2]: https://www.anthropic.com/news/detecting-and-preventing-dist...
You really cited a tech politics person, and Anthropic... The poltical AI people are the exact type of cope distributors I complained about above.
I think the more reasonable explanation is this is totally made up. Notice how they use examples concerning "cyber and bio" notice how the examples are politically loaded subjects that are hyper relevant to proposed AI dangers and why we need to protect western AI companies with anti-competitive regulations?
Oh no!!! The scary asians are distilling our models, but they only want the scary things!!! They want to hurt you!! We better stop the scary asians!!!
You need to up your media literacy.
> I think the more reasonable explanation is this is totally made up.
Based on what? None of the accused companies have ever denied Anthropic's accusations. And here you are confidently claiming it all to be made it. How much of a reputational hit to one's honesty do you think it carries to completely fabricate someone doing something, when they can just publicly disagree and disprove you? First thing DeepSeek, Moonshot, etc. would do if they hadn't would be to make a press statement calling out Anthropic for their lies.
DeepSeek routing to Fable was corroborated by many independent sources... since... everyone was able to just try it out and see it for themselves.
> Oh no!!! The scary asians are distilling our models, but they only want the scary things!!!
The straw man you're mocking doesn't exist. Non-Chinese labs are also distilling frontier models of their competitors. Elon testified in court that Grok was trained on ChatGPT outputs. And why would they not? It produces objectively better models.
Not everything is a grand conspiracy.
Yeah, I'd expect every lab to use the resources at their disposal to get ahead. Everyone's distilling everyone else, there's nothing wrong with it.
Despite being "corroborated by independent sources," it was impossible - because deepseek has a higher TPS, can return logits and reveals thinking. Distillation happens asynchronously.
DeepSeek's Anthropic compatible endpoint doesn't return logits. The accusation is that requests to api.deepseek.com/anthropic coming from Claude Code or OpenCode just replayed Claude's responses sometimes. The clients would have just happily consumed the responses, as they were already meant to be Anthropic compatible. And yeah, TPS and the response-distribution were incredibly different between DeepSeek and Claude models, which is exactly why people noticed.
Multi-turn conversations have much more entropy than one-shot responses, making it more useful to do synchronous distillation.
So people were getting deepseek responses with long waits and no thinking traces? That would be easy to prove, we'd get more than just claims from interested entities.
When did I say everything is a grand conspiracy? This whole thing where centrists normies just call anyone who believes something outside their center comfort zone a conspiracy theorist is mega obnoxious.
You need better media literacy.
You cant just make up a reality where none of these companies have denied distillation.
The scary asians are going to get you, watch out! Your government is going to eventually tell you that we need kill a bunch of them too because they're so scary. Probably soonish. You're going to fall for it because you believe anything you're told.
You are postulating an organised psyop, where OpenAI and Anthropic are publicly lying about Chinese labs to paint an image of, in your words: "Oh no!!! The scary asians are distilling our models, but they only want the scary things!!!"
These lies are then further supported publicly my western media, random people on twitter, western governments, etc., all to paint China as an evil force that must be united against. Again, in your words "Your government is going to eventually tell you that we need kill a bunch of them too because they're so scary."
And you have dozens of independent parties pushing that same lie to that end and for that purpose? Sure, that's what the New York Times and CNBC must deeply want too. When CNBC writes that they've reached out for comments to the accused companies, but received no responses, they must be lying as well?
It shouldn't be hard for you to demonstrate DeepSeek denying it. Please, go ahead, link to a statement. China's foreign ministry has denied Chinese labs distilling, and, in the same statement accused US labs of distilling from Chinese models. Now that's a very trustworthy claim.
I call it a conspiracy theory because it's a conspiracy theory. Some conspiracies are true, but this one is not. Distilling is both trivially simple and produces objectively better models. Plenty of evidence points to that practice. But in your mind the labs in China are just morally too upstanding in engage in such lowly tactics to create better models, despite the evidence. And it's convenient that the evidence can be explained away by claiming it to be a psyop to paint Chinese as evil.
[dead]
Deepseek Harness lets you opt out of your data being used for training. Perhaps they're lying, but if so I'd have expected someone to point it out and win internet points for exposing false promises. I leave mine opted in since I don't have PII to worry about.
kind of on track with gemini's releases
google is a company that tries to make a point of having profitable subdivisions
What about sourcing the asset itself? It's like a factory is it not?
You pay for the asset (tenancy in a data center) and then you pay for housing, powering, cooling, etc.
I think it’s pretty clear from the cherry picked numbers at Anthropic.
Demand is the issue. Demand only spikes with a sota model release then in drops back down to unprofitable levels when factoring in training a new model and build out.
So your stuck in an unprofitable cycle of losing revenue when your not releasing sotas but losing profit when you are.
This goes back to googles original memo “there is no moat”.
Where in you can’t stop training sotas because you’ll lose customers to the next inference model above you instantly.
Google has seemingly defeated this by finding models cheap and good enough for their existing platforms resulting in a moat of distribution. Something OpenAI and Anthropic seems to lack.
Unless the money of the investors is going somewhere else...
I think you have to take this on a case by case basis - there is no reason to suppose that the training costs and serving costs are the same for all companies for all models and for all price plans. Some may be profitable, others not.
DeepSeek's models are relatively small, and training costs will be correspondingly lower than for larger models. All you have here is a revenue number - no mention of (training + serving) cost, so while they may well be profitable you can't determine that from this number alone (if it can be trusted).
DeepSeek (investor meeting leak) have previously said that they price their models at a point where the GPU is paid for after 10 months, so something doesn't seem to add up here. If they were making $38/hr per card 24x7, then in 10 months they'd be making $270K per card, while the cards (Huawei?) only cost around $20K.
Presumably cards run on electricity?
That turns out to only be a small part of the cost.
Say a $30,000 GPU has an operational lifetime of 3 years, and draws 1000W. If it is running 24x7 then over those 3 years it will consume 26,280 kWh of electricity, at a cost of say $0.10 per kWh, giving a lifetime electricity cost of $2600, or less than 10% of the cost of the GPU itself.
So, to a first approximation the cost of serving a model is the amortized GPU cost.
I don't think i'm paying more than Deepseek price, ever
Until the Chinese stuff came online I was completely convinced this is just a bubble waiting to pop and nothing will come of this tech due to its absurd costs. The Chinese stuff changed this picture massively. I'm daily driving DeepSeek 4.1 flash (at API prices) and every now and then I give Opus 5.5 a try and don't notice anything missing for me.
My workflow is one where I am heavily involved in the process and I don't like giving the agent too much autonomy + I need to follow along. Opus 5.5 speak is still cause aneurysms for my brain. It also has this vibe to its prose... like it's looking down on me. This is Prolly my bias speaking though.
> My workflow is one where I am heavily involved in the process and I don't like giving the agent too much autonomy
For such workflows it barely matters which of a dozen models gets used.
Whereas if you're handing off all dev and no longer reading code, just making CTO-level architecture decisions, then there are only 2~3 models right now worth using, none cheap.
> Whereas if you're handing off all dev and no longer reading code, just making CTO-level architecture decisions, then there are only 2~3 models right now worth using, none cheap.
I use frontier models every day and night, but not sure if the described approach is possible with any of them right now. Models still make stupid mistakes and people who do not read the code produced are bound for surprising discoveries in their codebase. Not trying to downplay the effectiveness of LLM-assisted coding and all that, but it's hard to imagine how much stupid shit i would've committed if I was not reading their output.
Serious question: Are you do a combative code analysis with a range of different models?
I used to find real dumb bugs in code reviews. With 2-3 different models doing a combative code review.. most of that is already fixed when I get it.
Yeah, I have extensive collection of prompts for reviews and usually run 3-5 (automated) review takes on large sets of changes before doing eyeball reviews. Auto reviews indeed find many obvious bugs, but they usually slip through things that are less obviously stupid but are bad conceptually, for example, using persistence / queue / repository in a place where set of pure functions would be enough; or vice-versa, not persisting the state that must be persisted for reasons, etc.
Another problem is that review changes can easily introduce new issues, so my review prompts go like that: if you find outright shit, fix it right away, but if you have slightest concern, just write your ideas and suggestions in the review and let me decide then. I try from time to time to omit the "slightest concern" part and make agent do all the shit it wants, and usually end up reverting those changes.
But as the sibling comment said, not all code needs quality and indeed I skip some code too, like user playwright journeys -- I only need one auto review pass to verify they indeed test things according to a doc, and I'm fine not reading them. Not all code like that though, some stuff would wreck havoc later if its not looked at.
If you can test well enough, both manually and automated, then it's okay if there's a bunch of stupid mistakes in the code, as long as the end result works right. Not unlike biology.
If you view creating programs like an optimization problem over the feasible region of valid programs, this is not unreasonable.
> I'm daily driving DeepSeek 4.1 flash (at API prices) and every now and then I give Opus 5.5 a try and don't notice anything missing for me.
I was doing the same (along with GLM 5.3 Flash) for months, and then gave Claude Sonnet 5.5 at medium effort a try. The difference was like night and day. Where both DeepSeek 4.1 Flash and GLM 5.3 Flash kept making mistakes and would have to do a bunch of iterations to find and fix them, Claude Sonnet 5.5 just did it right on the first try.
In my experience, while DeepSeek 4.1 Flash and GLM 5.3 Flash are a lot cheaper per token, they're just as or maybe even more expensive when you factor in all the extra tokens that are need to troubleshoot and fix their mistakes -- and even more expensive if I factor in my own time wasted doing the same, not to mention the frustration of having to deal with that.
DeepSeek 4.1 Flash also hallucinates a lot.[1]
I'm hoping all this improves with time, and the gap between the competent frontier models and open weight models narrows.
[1] - https://www.youtube.com/watch?v=P4dTq4X8bqk
Same for me. The cheap Claude subscription is just so much cheaper. Never had to buy more deepseek tokens for months (August), because Sonnet 5.5 is good enough
I don't think it's a fair comparison. Maybe against deepseek pro would be more similar in terms of model size
That’s the Model Welfare speaking, not your bias.
OpenAI's entire business is trivially profitable when you attach an ad network to ChatGPT.
Quick, someone point out the one billion user system with intent + personal data rich that doesn't result in a $100+ billion high margin ad network.
It is that simple and absolutely nothing can stop the gigantic ad networks that will be built up around personal agents.
Only if CPM can be high enough to counter balance the inference costs.
Current CPM is between $30 and $60 which is x20 to x30 more expensive than alternatives.
Now we have to wait and see if performance is also at least 30x of previous ad targeting.
What is that CPM? My CPM in ChatGPT is lower than in Google. CPM is a bad metric, how much does 1 conversion cost? You can charge me $100 per ad all day, as long as every showing results in a $1000 sale.
There's a trillion plus dollars (per year) of ad business up for grabs in the LLM era.
Facebook is toast, sooner than later, as a business. Whatever gradually replaces/upgrades that quasi dead friends network, will be premised on AI (Zuck apparently thinks it's at least in part personal agents). In the present Meta still has a $228 billion ad business that is hyper profitable.
Google + Meta = ~$650 billion hyper margin ad business. OpenAI, aka LLM as a service for the masses, is going to want to take as much of that as they can. OpenAI's active user base comes at an attention cost to the other mega services like FB or Google Search, and ultimately that attention is worth ad dollars. The question is only whether OpenAI can take it, not whether it's there at all or high value. And of course there's no guarantee that it'll be ChatGPT upon which the next great ad network is built, it will be some AI-centric business however, and likely heavily involving LLMs.
Google's operating income last quarter was $40 billion. If OpenAI fails to take the mass market + ad tandem, then Google will more likely retain their position. The solution to the cost is already well understood by eg Google: the masses don't need tip of the frontier AI capabilities, and relying on Nvidia's extremely over-priced GPUs is a dumb business model. The biggest threat to OpenAI right now is Nvidia's margins, not China. Google's focus has been on high quality cheaper AI systems for a reason: they're looking to serve billions of people that have no use for what the bleeding edge offers vs a fast + good mid-tier. Google can do that in a highly profitable way.
The thing that stops it is, ironically, personal agents. Get one to interact with the ad-driven one if you need to and filter ads out.
I suppose the very reason why we haven't seen ads yet is that they're trying to figure out how to prevent exactly this.
What? Ad networks REQUIRE a moat. Google Meta Amazon have historically defensible moats. ChatGPT? Zero moat. Good luck with that ad business!
A billion users and the extreme cost of infrastructure to compete are a gigantic moat.
You can't get the first and you can't get the money for the second. Nobody is going to give you $100 billion to underwrite the cost of competing with GPT at a one billion user scale (or even a 100 million user scale). The easy money is already drying up, the field is set, and you will not be able to get anybody to give you a huge check to attempt to chase OpenAI, Google, Anthropic, China. That's already over. No VCs will fund you on this particular business pursuit.
You couldn't meaningfully compete with OpenAI and Anthropic if you wanted to. There's absolutely nothing you can do about it.
There's a trillion dollar valuation waiting for you. What's up?
Even nation-state pockets Microsoft can't compete. If there were no moat, Microsoft would have their own GPT killer, and GPT would be dead. Microsoft has $160 billion per year in op income. They can freely afford to spend more than the entire EU (is willing to) on AI matters if it makes sense. They can't compete no matter how much they spend and they know it.
Google has nation-state money. They have chips. They have software developers by the zillions. They have a frontier or very near frontier model in Gemini / Argon. They invented a lot of the category. They can't kill GPT. I repeat, no matter what Google has thrown at the matter, they have failed to kill or stop GPT.
Google can't beat them. Microsoft can't beat them. China's models haven't stopped them (DeepSeek was supposed to kill them years ago; Kimi was supposed to kill them).
That's one helluva non-moat. They must just be really lucky.
You might as well claim Coca Cola has no moat because it's just sugar water. Brand at scale is a large moat. People use what they know unless you severely disrupt them. They're being taught to use GPT, they know GPT, and so long as OpenAI doesn't screw it all up, people will just keep using it. To say nothing of the fact that GPT still has one of the elite frontier models, so you also still have to compete with them on quality.
[dead]
Now that the unit economics are laid out in the open, What stops any of the companies or even startups from anywhere in the world with some investment $ and competence from spinning up and offering their own version of the inference engines, undercutting any moat that the current leaders have?
I could even see F1000 companies having their own backstop intelligence for most of the tasks and only use frontier where it is absolutely needed or as much as it is under their IT budget.
Won't it be race to the bottom when they compete on cost? Is regulatory capture the only moat left for the market leaders now?
Nothing but hardware access, which is getting harder for the average person but for a giant corporation I see no reason they can’t compete now at the local model level
But haven't the Chinese (and Deepseek specifically) put a lot of emphasis on running models efficiently (because they've been locked out of being able to obtain certain GPUs by the US)? Whereas the US frontier companies seemed to get huge cash inflows and didn't need to worry about efficiency as much and put more emphasis on frontier performance. So how do you know that this $38/card/hr is about the same that the US frontier companies are making?
I still feel like there are some questions there, as in what is the actual amortized cost, plus the ongoing variable costs and even the non-training related overhead.
Plus there is the actual demand side to consider and how does it impact when paid utilization isn't 100%.
I don’t know, I take this as there being a good amount of margin that is going to be suppressed due competition in the very near term. It will be interesting to see how these margins hold up over the next 12 months, I expect them to drop substantially
> Amortized cost of owning a DC full of GPUs is far less than $38/card/hr
You missed the bit where they used the word "revenue".
I suspect if you factor in all the various overheads of running datacentres the actual value is lower. Revenue is vanity, profit is sanity as the old saying goes.
However the Chinese do have the benefit that China are pushing hard on renewable energies. So at least that major part of the equation is probably far lower than your average US/EU site.
Huge, generational mistake for the USA to be leaning so hard into fossil fuels right now and actively trying to prevent buildout of solar and wind power. That decision is just going to hurt more and more in the coming years as China's electricity costs continue to plummet while the US' continues to depend on unsustainable fossil fuel infrastructure.
> actively trying to prevent buildout of solar and wind power.
And actually paying people not to ! (Paying RWE $1.2bn to abandon wind contracts[1])
Its absolutely nuts. Pure generational can-kicking.
[1] https://www.bbc.co.uk/news/articles/c1e1vg0gjl5o
> Huge, generational mistake for the USA to be leaning so hard into fossil fuels right now and actively trying to prevent buildout of solar and wind power.
If the US is trying that, they are failing pretty hard - 88% of new power in the US in 2025 was renewables, and 2026 is estimated to be above 90%
One of countless mistakes made by the current administration. Decimating research funding and firing many scientists is another.
It's not a mistake. It's deliberate harm.
We are vigorously punching ourselves in the face to teach our perceived enemies a lesson.
I have no idea why this is being downvoted when the entire agenda of the current administration is to "starve the beast" by cutting all essential government departments and prove that public administration doesn't work so that everything can be privatized. And meanwhile the head of the current administration basically does what Russia (and Romania in the 90s did), he makes money by selling off those bits and pieces to the highest (personal) bidder.
Why? AI DCs can't rely on solar or wind so there must always be a thermal backstop. China is building lots of coal and nuclear.
Reality: renewables rapidly increasing in the US every year, especially solar (around 30%)
It's hilarious and depressing that one of Trump's supposed goals was making the US into an industrial powerhouse again, and yet he's acting to protect the profits of the fossil fuel industry as opposed to encouraging the kind of resource efficiency that would make the US competitive long term.
The only thing he cares about right now is covering up the Epstein Files. Everything else is key jingling to get you looking elsewhere.
China is investing in cheap and secure energy, which includes renewables, but it also include coal: https://www.reuters.com/business/energy/chinas-coal-power-ri...
The latter is great as a way of getting an edge in business and reducing reliance on oil supplies, but its also environmentally highly damaging.
> You missed the bit where they used the word "revenue".
No they didn't...
> I suspect if you factor in all the various overheads of running datacentres the actual value is lower.
That's literally exactly what the quoted part addresses: "Amortized cost of owning a DC full of GPUs is far less than $38/card/hr". Cost of running GPUs is far less than $38/hour, revenue is $38/hour, ??? , profit
The railroads went bust, twice, not because there wasn't profit available, but because investment outpaced profit and wasn't sustainable.
With this new data at hand, I have to wonder how profitable would it be for the investors.
The cost of amortising a datcenter, the gpus, the engineering costs attached all to produce inference.
At how much percentage of its inference would it be profitable (if it does), how much returns would it make over the long term and what would be the profit of those inferences/ profit-revenue and how that compares to other industries.
Are there any statistics which show this. Also how efficient token inference has gotten over time?
https://inferencex.semianalysis.com/profit-estimator
> Amortized cost of owning a DC full of GPUs is far less than $38/card/hr
what do you think it is, out of curiosity?
Other replies covered this, but I was basing that comment on the prices that commodity providers charge for GPU-accelerated VMs. E.g. the going rate for a NVIDIA B300 (288GB RAM) and the VM around it is somewhere between $5 and $20 an hour [1].
These providers have very little incentive to charge unprofitable prices (except maybe capturing more of the market, but there's not much lock-in here), so it's reasonable to assume somewhere in that spread is the "true" cost of being in this business, all expenses included.
[1] I'm using https://gpus.io as a reference here, and that's on-demand prices, which will tend to be the highest relative to other options.
(tongue mostly in cheek) sounds like I should be renting out these platforms and selling inference on them to OpenRouter
You can pay off a high end Nvidia DGX server at just over $7/hr/card over a 1 year period.
$38/card/hr is a ton of money. That's $2.6 million per 8-card server per year.
The cost of power/physical space/interconnect by comparison here is negligible.
...well you can get B200 in cloud for ~$6-14/hour and that's after all hosting costs and profits company make of it soooo yeah
Those arguments will go away when spending on AI infrastructure is no longer some multiple of gross annual software sales revenue, yeah?
When that happens, the chips stocks will have tanked back to sane levels and the doom posting will be about that. That will put us in the 2001 era of post dot com bubble, and the following few years will show us who the winner are
>It's really refreshing to put some cold hard numbers on this stuff, hopefully this quells some of the "AI is unprofitable everyone is spending a dollar to make fifty cents" arguments.
Inference has always been profitable. The problem is that this profit shrinks as your model becomes less competitive. The labs are in a constant race to release the best model to maintain inference profit. Deepseek is not training their own models at the scale of OpenAI and Anthropic and mostly sustains off their efforts via distillation.
You're looking for comfort in the wrong place.
From microns S10: BITS shipped and net revenue.
Q1 '20 - $0.10 profit per gb shipped (ram and storage)
Q1 '23 - -$0.024 (cyclic ram market)
Q1 '25 - $0.131 this looks normal were on the upswing.
Q4 '26 - $1.136 Were all fucked.
If you look at a chart of S3 prices we have all been living in the "Happy Time" where the cost to run our ever growing software business has gone down. Githb says 2025 billion commits - 14 billion this year, and 2x the actions run. That means AI driven software growth is 2x to 14x.
But those ram/storage prices are going to hit the operational budget of EVERYONE soon.
Can your average sas survive a 4x price increase on its AWS/Azure bill? How about 10x (cause thats what the margins say the premium is on ram).
If the circular financing is a concern, the circular firing squad should be a bigger one. Your paying for tokens to build more software to pay more to run your software is not a sustainable cycle. People are starting to use the words CapEx again - and it's a good thing.
And it seems we're getting just 10x more software nobody uses.
We don't need 10x more software per se. It's removing the high cost and exclusivity from software production (but not just software production). It also enables the production of highly specialized software by non-specialists, so it broadens the field dramatically (things that were not being produced, will be). Simply put, every bit of software that can be useful, will definitely be made; and a whole lot of unnecessary software will be made in that process as well, that's chaos-discovery for you.
> "AI is unprofitable everyone is spending a dollar to make fifty cents"
you are just talking about ai inference? Are you assuming ai models themselves fall from the sky.
AI as a business will be 10x larger than it is today within the next 20 years.
Present valuations are the bubble, not the business / industry. The exact same thing was true about the Internet bubble circa the late 1990s, except this is a lot more important than the Internet and the global economy is drastically larger.
Productivity is how you expand economy and wealth. It's how China can go from being one of the world's most backwards agricultural nations, unable to feed itself properly, to gradually becoming an agriculture juggernaut. AI is a hyper productivity tool, saving people time and bolstering capability are two of the primary productivity concepts.
We're not even out of the first inning yet.
Why do you posit that this is a lot more important than the Internet?
Artificial intelligence will 'infect' every surface of life and business. The Internet is merely a distribution platform for intelligence (and yes that includes TikTok videos). Its first distribution chore was slow and quaint, it was human. The intelligence was always the higher value, not the distribution network.
Distribution of intelligence: data, photos, video, communications, transactions, entertainment, et al.
And to be clear, it doesn't matter if some random commentors on HN don't think very highly of some or all of the intelligence being distributed. Their elitist opinion of the quality of intelligence of a given thing doesn't matter as to whether it's happening or not.
When a 16 year old records a cloned TikTok dance video, edits it, and then publishes it, that's intelligence distribution. That's a person thinking through a series of actions to try to communicate to an end audience, via an entertaining medium. That's intelligence being distributed over the Internet. That's the quaint, slow, human form of it.
Up to this point that distribution was basically entirely human, capped (post saturation) at slow growth N intelligence output (derived from the total number of humans and their time / productivity capabilities). AI uncaps that and radically expands intelligence output & distribution at a growth rate that is so fast you might as well start picturing the planet having tens or hundreds of billions of people on it. The human population will shrink, the artificial population will increase dramatically. People will have more artificial friends than human friends within ~20-30 years.
A few decades from now the notion of thinking of computing separate from AI will be so foreign most people won't understand what you're talking about. It will just be thought of as computing. Humans will forget a time when we weren't augmented by the constant presence of artificial intelligence. Try taking 'smart' phones away from the masses. In about a decade try taking away their helper agents. A decade after that try taking away their home robots.
The steroid jump is AI + robotics. The Internet is trite compared to what comes out of that over the rest of this century. Robotics will physically remake our world, in a way the Internet can't. We'll deposit tens of billions of robots onto the surface of the planet and they will almost all be the result of having figured out how to make intelligence (relatively) easy & (relatively) cheap. I'm not talking about Roombas.
The Information Age is over. We're into the Age of Intelligence (insert the unavoidable quips about humanity & intelligence).
The Internet is just one way to move intelligence around (a robot body is another, a smartphone is another, a human body is another, and so on). It's a sub value to intelligence broadly. Intelligence is the ultimate value. The valuations (and sales growth) for OpenAI and Anthropic is because they're vast intelligence output machines in the Age of Intelligence, ditto Nvidia's business boom. We can now mass manufacture intelligence at an accelerating clip, and it has gotten cheaper at a comical rate over the past few years. While HN comments snort about pelican improvements day to day, the leap over a few years is astronomical in regards to what significant AI systems can now do (with LLMs being just one type of that).
this is tech fart sniffing self-aggrandization. You’re saying words that mean nothing. Tens of billions of robots? To do what, exactly? Have you ever interacted with people, not nerds, real human people? Real people don’t crave a future filled with robots, real people want real people.
The world will be filled with robots is not the inspired forward-thinking of a visionary it is the uninspired regurgitated science fiction of someone who doesn’t care to understand that the world is people, a society.
There's nothing that is going to stop the AI-robot saturated future from arriving. It doesn't matter if you dislike it.
China will have many billions of robots just in their industrial efforts alone. Every house and every apartment in China will saturate with robots of various usefulness. That single market will be at least 10+ billion in scale.
There will be a billion robots for just the delivery and transport categories globally. We're going to peak at at least 10 billion humans (population), a billion robots for delivery & transport segments is easy.
In the toy category there will eventually be far in excess of a billion robots. They will seem quite advanced compared to what we presently are used to in that category. The Microduck is a primitive example of what will be a large industry around smart entertaining toys. Your artificial friends will be able to 'hop into' (take control of) your Microduck (whatever the 2035-2050 equivalent is) and cruise around the house or wherever you're at. Entertainment will continue to be a gigantic industry and that will obviously apply to AI + life. We haven't even begun down the road of AI + entertainment, it's less than early there.
Every household in the world will eventually have numerous robots, from primitive robots (including single purpose and toys), to advanced humanoid robots. It won't take longer than 2050-2060 to push $10k-$20k humanoid robots into middle class or upper middle class homes. China will put enormous pressure on the cost of those robots, and they'll mass sell them to the bottom 4/5 of the world in terms of household incomes (that is, they'll come up from the floor on price and take most of the market, while some companies in affluent nations will produce more expensive high-end humanoid bots, ala what Tesla is attempting).
Tens of billions of robots globally is easy. 200-300 billion artificial friends is also an easy prediction. Social junkies will have 50-100 artificial friends that they try to tend to. It'll largely replace platforms like Facebook, the utility that served for the average person. Humans will get a lot more isolated from each other and as that occurs they'll fill in the void.
I am a nerd on Hacker News. My preference for technology and robots is obvious and an assumption that my disagreement with you is based on personal preference is brainless.
I have spent time in China and Korea where attitudes towards artificial intelligence are much more positive than the west. I have interacted with robots in these countries.
You are making the fundamental mistake all nerds make. You’re thinking about what is possible and concluding that it is what will happen. Just as nerds believe that someday it will be possible for us to live on mars and therefore we are going to live on mars.
Humans are humans. Humans are interacting with artificial intelligence as if it is human because they crave human contact and are so desperate they settle for the best they can get.
The houses filled with robots future goes hand in hand with the robots will be doing all our work future. And if robots are doing all our work, what do humans have? Time. So much time. Humans become creatures of leisure. No longer is there a too little time to build relationships, there is only time, so much time. Human relationships become abundant, humans spend all their time being human.
The idea that people will have 50-100 artificial friends is so batshit crazy that I could totally believe your comment is parody.
If real people want real people, then why is everyone at the coffee shop, on the train, at parties, and even at the dinner table all looking at their phones? Not all the time, but way too much of the time, imo.
> at parties, and even at the dinner table all looking at their phones
This is very different from what I see.
At coffee shops and on trains, sure, people are working or on their phones. But at dinner and parties, at bars and climbing gym and dances, we're there to be there with each other.
Maybe a cultural difference between where you are and where I am?
But then again, they are mostly looking at images and videos of other people, right?
Because using internet is an addiction.
> People will have more artificial friends than human friends within ~20-30 years.
That would be awful, presuming that the artificial friends are not people.
Separately, you might like Count to the Eschaton by John Wright with its ultimate project being "the sophistication of all matter".
The masses won't think it's awful though, and that's the thing that will facilitate its explosion as a commonality across the planet. The masses are extremely lonely all over the planet. Along with desperately loving entertainment, they desperately want companionship.
Just as human intelligence is merely one form of intelligence, human companionship will be merely one form of companionship. AI friends will offer a new type. And for the masses it's going to feel very real and very legitimate. They'll like their AI friends as much or more than they like their human friends.
And the fake friend explosion can happen first quietly, which eliminates the initial shame phase problem. Nobody will have to know you have fake friends, or how much you interact with them. Until it's common it'll be viewed as shameful (pathetic, proof of social ineptness) and mocked by society, then it'll be normalized. This idea isn't novel, it's extremely obvious and right now we're just waiting for a few pieces of the LLM puzzle to get to the point where they're good enough to facilitate high quality artificial friends.
It'll happen shockingly rapidly when it sparks; slowly at first and then in the span of 3-5 years we'll go from nobody having or being willing to admit to having artificial friends, to everybody openly discussing it. It will then seem more odd to not have artificial friends, like people that refuse to own smartphones or use the Internet. People will start commonly referring to their artificial friends by name in conversations with other human friends. These behaviors will become entirely socially normal.
"Yeah I had Alice clean the house before I went on vacation." Every major personal robot, every artificial friend, will get a name. Humans name their vehicles, so they're definitely going to name their $10k-$20k house bot and the artificial friends they spend large amounts of time with. When you go on a long drive (etc), your artificial friends will be there to keep you company, and they will remember your life events as well or better than you do (if you give them access to that recording/memory capability).
Maybe some people will like to have artificial friends with genuine two-way relationship, but I think most of us want artificial servants. Robots that clean our house, keep us company, or maybe even act as romantic partner without asking anything in return. Maybe some of them will feel like friends, but in the end, we are their owners and they are our servants. Genuine friendship requires freedom and autonomy for the other side to walk away if they so choose, and are we really going to give that to our AI bots? I doubt it.
> Along with desperately loving entertainment, they desperately want companionship.
And I am presuming that the companionship that people need is the companionship of other human beings. Otherwise, tamagotchi, fish, cats, dogs, parrots, and horses are already perfectly good companions. AI can help fill in for some of those other ones (tamagotchi especially), but it cannot replace human companionship.
> Humans name their vehicles, so they're definitely going to name their $10k-$20k house bot
huh. I can only think of one friend I've ever known who named their car.
Anyways, not saying I necessarily disagree with the rest of what you say, but it's out there, for sure. We're a good distance away from an AI that can meaningfully act like a friend.
(Though who knows what AI will look like in even 10 years? I'm not ruling this out)
People naming their AI's is already happening.
Apple, Amazon and Clarke started the trend.
There's a rising trend among gen alpha to not use phones or internet to get back their youth period that was ruined by the algorithms. I predict this trend will continue, and I myself am yearning to use less technology each day (but can't because of salary!).
But yeah, masses will do masses stuff and that's why even shit products can sell.
>People will have more artificial friends than human friends within ~20-30 years.
How do you envision an "artificial friend"
> AI as a business will be 10x larger than it is today within the next 20 years.
I would argue it will be 10x in 2-3 years the way things are going.
> It's really refreshing to put some cold hard numbers on this stuff, hopefully this quells some of the "AI is unprofitable everyone is spending a dollar to make fifty cents" arguments. Amortized cost of owning a DC full of GPUs is far less than $38/card/hr
They (well, most of them) are looking at loans and debt (in plain sight and hidden) of AI companies and asking "what will pay for that", not looking at how much a GPU costs and how much tokens you can sell off it.
Just putting a model on a GPU and selling it is profitable, period.
But spending more and more money on building and running infra to build frontier models that solve smaller and smaller piece of the pie might very well be unprofitable.
Models competing with the frontier labs gets more capable every month and more and more of the tasks can be achieved without reaching for the big guns.
We might very well get to the point where AI/LLM is profitable just fine but bubble still bursts and the big frontier labs are in trouble. And frankly that would lead to far more sustainable ecosystem.
Today. As new supply comes online and demand levels off, that 38$ will shrink. Absent legal monopolies, the margin will shrink to a sliver more similar to any other service. That is the danger of open source models. They destroy monopolies and shrink margins.
> That is the danger of open source models. They destroy monopolies and shrink margins.
Danger? You mean benefit?
That would depend on whether you are a user/customer or an investor/owner in AI. At the moment, nearly everyone is on the investor/owner side. The entire western world has doubled down on AI becoming the most popular product since the invention of fire. If it turns out that fire 2.0 cannot be patented/copyrighted/licensed for insane profits, those dollars disappear.
The people who own these companies have more money than they, their families, and their extended families can spend for generations. None of them have to work a day in their life no matter what happens to these companies.
Same with a lot of their investors, who will be just fine without the "insane profits" they hope for.
I'm not worried about their welfare, and won't shed a tear if their companies collapse.
It isn't just those who own stock. If the AI bubble bursts violently it will impact literally everyone. Everyone who owns indexed stocks. Everyone who owns tech stocks. Everyone with a variable rate mortgage. Everyone with a job. The AI spend is now so much of the western economic system that every one of us has skin in the game. At least on the risk side.
It's mostly concentrated in US tbh
idk man 2008 didn't affect me much directly
> The entire western world has doubled down on AI becoming the most popular product since the invention of fire.
That's a bit of a one-sided view.
AI companies are now 49% of the S&P 500.
Massive bubble.
When every company calls themselves AI-first, what then is an 'AI Company.' In many regards, this hype-chasing branding exercise is currently no different than calling one's company a 'technology company'... Just as the term "AI" itself has (to a degree) become meaningless... and AGI will be soon to follow.
Agi is not coming in this century. Like come on.
This is probably the boldest claim in the entire comments section, considering we're only in 2026.
Compare 1926 to 1999, and how technology advanced in that time. Radio wasn't widespread in 1926, nuclear fission hadn't been demonstrated, the idea of the semiconductor energy band gap, underlying the entirety of the modern computing industry, wasn't put forth until 1928. And yet, 20 years later we were purifying silicon, and 10 years after that (1954) the first silicon transistor was created and kicked off the semiconductor boom. 50 years later, we were in the Internet bubble.
We've got decades upon decades upon decades to come up with new ideas and implement them, even before the end of the century.
Right, I have to think P(AGI by century's end | LLMs turn out to be an AGI dead end) is still at least 50%.
Nothing bold in it, as there is no more meaning in the AGI.
We already lived through multiple AGIs for the last few years.
... have we?
Can AI handle driving yet?
https://nitter.cf/JensenHuang/status/2096700264569090384
> nearly everyone is on the investor/owner side
I think that "everyone" needs some kind of an asterisk. Who's "everyone" in this?
Depends if you’re a shareholder or not
> That is the danger of open source models. They destroy monopolies and shrink margins.
It's unclear that the leading open-source models could exist without having distilled the frontier models. If that's true, the "danger" is that these upstream models don't get built any more (because of the tremendous cost), and we end up just optimizing what we have. Also unclear whether this would be a net win or loss for humanity.
The tech is at the point where synthetic data is almost high enough quality to train further models on, AND there are so many human users of cloud AI, that their usage can be gathered and trained upon (it's human generated and used in real world situations, so, it is of a good enough quality). The cat is fully out of the bag.
This deals with half of one of the three scaling factors only. Presumably data production (choosing your training data, etc), model size, and training time/power still contribute huge costs. Who's investing that cash if open-source models will distill out enough to compete without the investment costs?
There is a community of people who compile LoRAs of their favorite subjects/styles in order to fine tune models. There is currently a bit of a gap between LOrA creation an integration of that data into larger models but I can see serious hobbyists stepping up to that plate soon.
The upstream models weren't built by god, they were built by distilling the world's knowledge and cultural products.
If OpenAI/Anthropic stop doing it there's no reason to believe someone else won't. Or that we won't come up with a different approach.
> they were built by distilling the world's knowledge and cultural products
... and a bunch of RLHF and proprietary training data etc etc, but regardless, this is still an exceptionally expensive process, that someone somewhere is paying for.
I'm OK with this.
Anthropic is zooming to $100 billion in annualized revenue.
The tremendous cost is matched by the tremendous business output.
In the span of less than five years, starting from their first $1 of revenue, Anthropic + OpenAI will have produced businesses doing at least $150 billion in combined sales. The demand isn't reducing at present, it's climbing, they can't come close to servicing the total demand for their frontier models.
SAP is under $40 billion in annualized revenue. They're 54 years old.
If SAP sold $1 bills for $0.50, they too could rapidly outpace Anthro/OAI's growth. And when sane minds ask about profits or a defensive moat, the response is "but look at our revenue bro".
The same exact thing was endlessly claimed about Uber. They were guaranteed to die. They were selling rides for less than they cost. Uber was a commodity app anybody could MVP in a weekend.
Market captured. Whoops.
Uber now: $55b in sales, $6.7b in op income. They won.
And to be clear, OpenAI can easily permanently subsidize a fake low cost for their LLMs if they get off their asses and attach a monster ad network to GPT already. They're years behind where they should have already been with ads, the easy money made it so they didn't have to think about it. If they miscalculate the timing on switching over from easy money (market) to ads (also easy money) it could certainly kill them.
OpenAI can knock out a lot of competition by building the largest ad network, then undercutting the cost that consumers see. The margins in ads make that possible. Google already knows how that works, they're doing the same thing right now.
I'd rather have the OpenAI + consumer base + ads position, than being Anthropic. China is a far, far larger threat to Anthropic over time. OpenAI is building a service business more comparable to what Google does. Chinese models don't actually threaten either OpenAI or Google insofar as their positions are user capture based, whereas constantly improving Chinese models are a never-ending threat to Anthropic.
> The same exact thing was endlessly claimed about Uber. They were guaranteed to die. They were selling rides for less than they cost.
Lots of companies have used this strategy. And many of them have indeed died. Uber filled a void left by bad public transportation and a lackluster cab system. The higher prices (so they can turn a profit) are still worth it to some.
> OpenAI can knock out a lot of competition by building the largest ad network, then undercutting the cost that consumers see. The margins in ads make that possible. Google already knows how that works, they're doing the same thing right now.
And you think OpenAI can compete here? If LLM's end up being yet another ad funded platform, I have a hard time believing Anthropic or OpenAI will outperform Google here. Maybe if one of them strike a deal with Apple, or get bought by them. Otherwise, ads are Google's bread & butter. And, they own Android.
The main difference between Uber and other parties that tried was Uber's willingness to break the law. They then made it stick because of money.
According to who? There is no agreed upon anyone that says that?
It is mostly response to claims of "all LLM are subsidized and blatant bubble, LLM are useless" etc
All LLM are subsidized and blatant bubble.
If the only profit is from running commodity free software (as OP claims), there's no moat and OpenAI/Anthropic are worthless.
revenue != profit?
you actually have no idea what the cost is
We still need to know if that includes purchase and maintenance costs, which are substantial. And then of course all other costs associated with it.
You also have to realize deepseeks running costs are immensely lower than say any us lab by orders of magnitude.
Chinas grid is almost entirely renewable energy with strong investments in it. In the us not the case, some dcs are running on gas.
Also DeepSeek was smart. They strategically rolled out model sizes they could support even if the intelligence wasn't the world's best for that week in human history. Then they have incrementally and cheaply improved models since.
Not to mention they also built trust with open weight models rather then fear mongering. They are in the top 3 of the smartest most viable LLM labs out there.
China generates more electricity from coal than from all renewables put together (49.7% per https://oilprice.com/Latest-Energy-News/World-News/Solar-Ove...).
They have more installed renewable capacity than coal, but renewables have low capacity factors, solar is less than 50% by definition. The renewable share is growing fairly rapidly, but it is by now means almost entirely renewable.
[dead]
At off peak pricing, which is substantially lower than the standard API rate (50% off). So I assume it looks even better financially overall.
Quick numbers:
At $35/hour in net cash generation per card, a $30,000 GPU would pay for itself entirely in about 850 hours (roughly 35 days)
With an assumed ~ $3 operation cost per card and hour
DeepSeek CEO has previously said that their finance team considers the optimal payback period for hardware purchase to be 10 months. That is, 10 months worth of revenue should be equal to the purchase price of the hardware. This doesn’t cover operational costs such as electricity. This is their method for API pricing.
Source: https://x.com/FuckAnthropic/article/2080268590033621295 (it’s a leaked meeting transcript so take it with a grain of salt)
Its not 10 months worth of revenue should be equal to purchase price. He is saying 10 months margin will cover the hardware cost, so revenue is ofcourse going to be higher, quote below.
"This is our API pricing. For our API pricing, we consider a reasonable profit to be roughly: we buy a batch of equipment from the market, and recoup the cost in ten months. I think that's a reasonable profit."
I read the same quote and I think it is ambiguous. I think he is talking from the cash flow perspective. He doesn’t mean it in the accounting sense where the amount of recognizable profit in 10 months is equal to the depreciation allowed in the 10 months.
Even if it is implying cashflow, cashflow would not include operating expenses (he has to pay those monthly), which would imply he is referring to margin not including capex?
Isn't electricity a pretty major factor even over such a short fraction of the hardware's lifetime?
No, the absolute max power draw of these cards is like 1.5 kW, say 3 kW per card with mediocre cooling attached to a server. DCs get power in the 10c/kWh range at most. So 30c/card/hour, peanuts. And this is the most power hungry card at full tilt with mediocre cooling efficiency.
And thanks to China’s forward thinking on infrastructure and planning for oil shortages, both short term and long term, they will have access to the lowest energy costs most likely.
Ah, yes. I should have done the math before asking.
Your estimation feels too high. It implies >1000% ROI on a GPU card (= $300k gross annual profit on a $30k card).
As a reference point, Z.ai is operating at 26% gross margin as of 1H26. So assuming ~90% gross margin is likely a bit off.
No, those $3 are on the high end. If you do not factor in hardware depreciation, at China's low electricity prices (largest variable cost and China has a surplus larger than many country's entire consumption) and with GB200 which is one of DeepSeek's main platforms, the operation cost is actually less than 50 cents per hour most likely. If you factor in hardware depreciation this rises significantly but almost certainly stays under 3 usd per hour, probably closer to 2.
Orinally it was an educated guess based on rental prices I've seen but Opus 5.5 and Gemini 3.8 Flash both arive at the same numbers based on DeepSeek's Zhihu post and GB200 specs. Around 30 cents before depreciation and $2 to 3.50 including it.
I'm curious why such a large gap exists between our numbers. My hypotheses are:
A. The average utilization rate of inference GPU cards is not that high (i.e. it peaks during the daytime, but is idle at night). So the time window for earning $38/h is fairly limited.
B. They are renting costly GPU capacity from 3rdparty hyperscalers because they can't meet the peak demand with their own hardware.
C. Free-tier users are eating the gross margin.
D. Subscription users are eating the gross margin.
Another observable point is that DeepSeek's company-wide gross margin was reportedly 44.6%. It's intriguing what the full picture looks like here.
> The average utilization rate of inference GPU cards is not that high (i.e. it peaks during the daytime, but is idle at night). So the time window for earning $38/h is fairly limited.
In DeepSeek's case this kind of levels out because they have a large user base both in China and the West (time zones), and additionally attract users in the non-Chinese timezones using their off peak 50% discount which is what DeepSeek talks about here.
> B. They are renting costly GPU capacity from 3rdparty hyperscalers because they can't meet the peak demand with their own hardware.
I looked this up from Chinese sources and DeepSeek primarily builds their own compute infrastructure rather than renting it
> C. Free-tier users are eating the gross margin. > D. Subscription users are eating the gross margin.
That's true, but given how lucrative inference at API rates is, they can subsidize training and individual users to some degree. Are they still losing money overall? Not unlikely, but with moonshots that is often the case. Remember that Uber and Spotify took around 15 years to have their first profitable quarters.
How is the math so good for data centres but so bad for home gear? A Mac mini capable of running these models ends up at around 10 years of subscription costs.
All of it is really about scale, same reason as a single person you can't sell french fries as cheap as McDonalds.
Couldnt they though? A medium fry at mcdonalds costs about $4 and you can buy 5lbs if potatoes for $4.
I think its more like 40 cents to make at home. So I think its quite unlike McDonalds.
A single person could make french fries for themselves cheaper than McDonalds, but once they have to start including the cost of the building and equipment, charging for their labor, and complying with health codes and worker safety requirements, they will quickly find that the fixed costs swamp the variable costs.
Here's an anecdote. I have repurposed my 6 year old 32gb M1 MacBook Pro to run Qwen 3.6 35b a3b and I used to give it nightly maintenance tasks that span my OSS github repos.
I gave it mundane tasks like updating dependencies, writing changelogs, etc. It ran about 2 hours a night.
My MBP consumes 0.1w idle and about 10w running inference. I pay $0.33/kwh. If it runs for 2 hours a day over 365 days - it will cost $2.41/year (excluding idle time).
Excluding the cost of hardware - the running costs alone are more expensive than DeepSeek's API and the inference quality is much lower. The economics are even worse on AMD AI Max mini PCs and worse still with a dedicated GPU setup.
That said - my Apple Silicon example is orders of magnitude cheaper than western inference providers and offers unrivaled privacy.
Perhaps a local AI setup would be compelling for small/medium business users... and it would be practical if the hyperscalers didn't buy all of the potatoes on Earth.
I'm not sure how many tokens/sec you were running through your MBP. Let's just go with 35 tok/sec for the sake of having a number. 35 tok/sec * 3600 sec/hr * 2 hr/day * 365 day/yr => 91 million (output) tokens. At DeepSeek's price of $0.60/million tokens, that's $55.18.
Of course that's assuming it was just generating flat out all the time - no prefill or anything. That's not likely. Probably much of the time (and tokens) would be in prefill and cache hits. And as you say, DeepSeek v4.1 Flash is a much stronger model than a 35B two generation old Qwen.
So yep, DeepSeek's API is pretty cheap. They're not paying anything like $0.33/kwh for one, and they're able to keep util much higher than you are, and they are subsidizing so they can get more data and code to train on.
If you were paying for Fireworks to host the model in the US with no training and ZDR, it would be $1.80 per million tokens (3x price), so a grand total of $165.56/yr.
I'd argue that at those values, the price difference really doesn't matter one way or the other. What matters is output quality, human time required, privacy, and sovereignty. Also, there's something to be said for (nearly) zero marginal cost reducing the barrier to playing with an idea. If I had to spend $40 to try out an idea over a few million files, I might think about it first. If I could just load it up on my existing server and let it cook overnight, I would not worry about it. Maybe I should consider electricity marginal costs more, but practically I don't really.
The privacy is the key to local deployment.
>but once they have to start including the cost of the building and equipment, ...
But they don’t have to include any of those costs.
Remember the comparison was running a model at home for yourself. The analogue is making fries at home for yourself. Thats cheaper.
Making fries at home is only cheaper because you already have a stove, knives, pots, cooking oil, etc.
Running a model at home is very cheap if you already have a M5 Studio lying around for some reason.
they do have to include the deep fat fryer and oil...
unless you want some sad thawed potatoes instead of fries
and that's going to run you a lot more than $4, and take quite a while to amortize
(numbers are just to illustrate a point, not accurate) Now assume you can't buy a 5lb bag of potatoes and you can only buy 10,000 lb bag. Now also assume that you can't buy this 10,000 lb bag unless you know someone else will pay for fries for the other 9,995 lb because you can't eat that many fries. That's the kind of scale the token inference market is efficient at.
How much does a fryer cost? How much does the fryer cost to run? How much would it cost to get the fries into the customers hand (i.e. food truck, rent)?
Not applicable because the hypothetical is making fries at home for yourself.
Its more economical to make fries at your home for yourself but that is not the case for inference.
Serving LLM inference is more efficient with more concurrent requests. For a single inference request, the bottleneck is GPU memory bandwidth, as the inference engine spends most of its time waiting to move values in memory. With concurrent requests, it's possible to service multiple requests in a single pass through the model weights, meaning you can utilize more of the GPU/APU's compute.
So you have to price your local hardware against however many subscriptions you would need to run a dozen concurrent agents around the clock.
But even then, I would guess subscriptions are priced on the basis that they don't expect every user to utilize their entire usage limit. I.e. power users are subsidized by casual users.
Because the datacenter setups serve multiple users concurrently while your mac mini only serves one. You would also lose money hand over fist renting/buying 4xH200s just to serve deepseek flash to yourself.
These datacenters run the GPU's close to 100% usage 24/7. The Mac mini at home probably has <25% utilization. And then there's economies of scale, wholesale discounts, etc.
The Mac mini GPU also isn't as optimized as the datacenter GPUs
You can batch calls if you're serving a lot of people at once.
[flagged]
I still struggle to wrap my head around a $30,000 GPU card. But impressive payback numbers.
I just installed a $16k GPU at home - super pricy to be sure, but when you price it against hiring 4 junior developers working around the clock it's a steal
What if you compare it to an OpenRouter account and $0.50 per million tokens?
Oh that's way cheaper to be sure, but a lot less fun!
you also don't have to worry about hiding API keys or security of confidential data which is a bonus.
Could you explain what your infra set up looks like and what is your typical use case, if its more than a hobby?
My use-case is coding. I'm typically coding interactively with an agent most of the day, and then I have overnight jobs that run every night.
As for the infra setup, I am currently using llama.cpp, although I'm just now working on switching to vllm since it is better for concurrent inference, since it can interleave prompt-processing and decode. Until now llama.cpp was enough, since I was on an R9700 and it was only practical to run one agent with decent sized context anyway.
I want to understand this more - are you saying "would pay for itself" if you are offering chatbot commercially? Or using it yourself in lieu of ChatGPT etc..?
For the ones who are still wondering how so much profits, inference is a highly profitable domain: https://inferencex.semianalysis.com/profit-estimator
I think we need to collectively get a grip on statements like this.
There is no reason to think inference can or will remain interestingly profitable even if you assume pure-play inference providers who don't do any model training. Serving an LLM isn't child's play but it's also not so hard as to count as a moat. Inferencing is:
1. Selling a largely undifferentiated product.
2. Into a highly competitive market.
3. To customers who are primarily sensitive to price.
Industries like this have always existed and in the modern era tend to have margins of about 3% to 5%. Any level of margin higher than that gets competed into the ground.
If you want to claim inference is highly profitable in general, you're going to have to answer why it won't end up like supermarkets or airlines, both of which have bigger moats than inference.
The other problem with this analysis is obviously that you can't just stop training models. Even if you don't care about improving capability or performance, the knowledge in their weights gets stale. Cost of inferencing has to take into account the ongoing cost of training, you can't cleanly split one from the other.
Amen. AI boosters look at currently high profit margins on inference and blindly accept claims that these inference profit margins will continue. Of course those margins depend on historically massive R&D spend on training. The moment that training stops, AI returns to a commodity market and the inference profits end. Big labs are trying as hard as possible to keep the gravy train running but they can all see it won't last forever, hence the rush for IPO & regulating away open models.
Boosters are right that AI is a historically impactful product, but big labs will be unable to meaningfully profit from ushering in the AI era just like Netscape failed despite ushering in the internet era.
>AI boosters look at currently high profit margins on inference
Name a company with a currently high profit margin on inference. I can't.
OpenAI and Anthropic's valuations are as high as they are because investors hope that they will create AI that is much much more powerful than the AI they have now. The fact that they already have a substantial revenue stream helps, but it helps mainly by lending credibility to the hope that they might create something much much more powerful.
(You, Mike and I agree that selling the inference of models that are not much better than liberally-licensed open-weights models will be a low-margin business.)
"If you want to claim inference is highly profitable in general, you're going to have to answer why it won't end up like supermarkets or airlines, both of which have bigger moats than inference."
- I don't... "inference is profitable" is as of now. No one knows future, neither have I commented anything on it.
The statement is valid as it stands.
Inference will continue to be profitable while we don't have enough compute in the world to serve the demand we have. Also, I don't see why it's a problem that the end game for inference of open models will be relatively low, ie 5% or so. You can have a huge industry running on these margins. Though I would note that hyperscalers have been selling compute commodities with a high margin for a long time.
Demand is elastic, there's no fixed pool of it. What determines profitability isn't merely supply constraints but how much margin your competitors are willing to accept, more or less.
It's not a huge problem for users but if margins are low it's a huge problem for the debt funded datacenter buildout. It's just pure finance at that point, about whether those GPUs can generate enough yield to beat deprecation and financing costs.
The big clouds generally don't sell compute at high margins. The margins differ a lot between services. Services like core VMs and object storage are usually low margin or even break even. Where they make the money is things where they've got proprietary software with competitive advantage, like managed scalable forks of Postgres, or egress charges that are there to create lockin. Managed DBs are pure SaaS and can generate SaaS like margins.
Exactly, inference is like groceries or electricity.
Because LLMs are like evolutionary crabs, they all get to the same point.
I think hardware scarcity plays a part in inference margins. They would get squeezed super quickly if anyone who wanted could just build an inference cluster, but that's not currently the case.
I would assume hardware production has been ramping up significantly and it's more of a temporary supply chain issue though. While it could last multiple years, that's certainly not long enough for the hyperscalers to get a return on investment. Obviously more complicated than this, but still... I'm not buying.
We're also starting to see efficiency gains, so we might meet halfway.
Maybe but there are huge blocks of capacity about to be energized or handed over and the payment clock starts ticking immediately.
Complete and accurate, thanks.
Accounting for initial acquisition cost and typical depreciation/failure rate? I assume they already consider colo price, electricity, etc
70%+ net profits margins are for rented infra, that means no major upfont expense. Also, these margins are against 60% utilization only.
You can well imagine if you own your infra and can make utilization over 90%, where these go.
And then, my link above is for K3 model, which, compared to ds flash models, is highly unoptimized.
If I had a few million lying around, I would have invested in inference.
I would be surprised if there was such a glut of undifferentiated inference providers for open models otherwise[0]. Possibly just people trying to minimize the loss from already-committed capital, but I don't see AWS and DO selling inference via OpenRouter unless they're making a decent profit from it.
0: https://openrouter.ai/openai/gpt-oss-120b#providers
Fascinating site. Thanks for linking. However note if you change to "open router" pricing in some cases even B200 which is ~2y old has negative margin (and AFAIK ppl are using longer amortization than 2y - I do not know how cost are calculated here though).
1. The link is for rented infra. That means the inference provider doesnt have to worry about deprecation.
2. Newer gen nvidia gpu's like GB300 are way more profitable.
Why isn't ordered by profit % per GPU? It' surely the most important indicator
DS4 Flash is so cheap also, imagine what the numbers look like for a flagship model like Opus 5.5, which charges 100x more. Of course inference for that model also costs more, but I doubt the proportion is the same.
Or now, Anthropic will distill the knowledge from this engine and optimize their inference.
The top quality frontier models are used to compress / improve the earlier, lesser models, to make them more efficient. So that pushes costs way down on previous models.
Deepseeks strategy is that each project must break even within 10 months.
$38/hrs equals $320K in 10 months.
Assume they only have 100 clusters, and as much as 40% operational costs. That would get them about $20M in 10 months, which I suspect is more than what 4.1 flash training has cost them.
No, a cluster consists of 4x SM100, and $38 profit is per card, so after 10 months you already have $1.0 million.
I haven't accounted for 7 hours of peak-pricing per workday.
Of course, you won't have 100% utilization.
For four cards at 100% utilization it's $4712 per workday and $3648 per day on the weekends and holidays.
This is per card pricing, not per cluster pricing, so I think you’re off by an OOM or 2
155T tokens/day per cluster? OpenRouter reports that last week total usage on their platform was 164T tokens. Surely all that demand couldn't be satisfied with one cluster? Are these different units of token, or maybe something got lost in translation?
Exactly... I was going to post exactly this and you called it two minutes before me. On openrouter, it is shown that 600T+ tokens are provided per month, so 20T+ per day. Openrouter is not all the world but it seems that something is not right with this 155T figure... Unless everyone is idle everywhere, waiting for demand
Isn't OpenRouter only part of the traffic of DeepSeek's inference?
Yes, it is only a part of DeepSeek's inference. Alongside openrouter there are others (probably many) doing this gateway work (for example opencode and even huggingface). Also the gateways in China itself, and of course also non gateway inference, DS itself directly providing inference.
OpenRouter routes traffic through many different providers
https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017#prov...
From the Github benchmark numbers, one GPU can do about 26k tokens/s when serving 256 users, which is about 2.25B tokens/day. A cluster would need to be close to 70k GPUs to serve 155T tokens/day (which seems doubtful).
This is a major, major datapoint. That’s at least an order of magnitude higher margin than literally anyone is discussing. The entire ai discourse currently assumes very small margins on open model inf. And by this post it’s more like 1600% (h100s rent for $2/hr - and id guess deepseeks costs are far below that)
So it’s only getting better since last year!
Are frontier labs making 80+% gross margins on LLM inference? (2025)
https://johnnyclee.com/i/are-frontier-labs-making-80percent-...
You can buy such a PC off from ebay for only €150.000. We did.
But the architecture is similar to kimi's mixture of experts, and we already have such an optimization for different experts swapped in from disc by antirez.
so if this was true then why do we need see GPU rental prices increase and openrouter token prices fall, for deepseek and other open models?
Ultimately, more people serving.
- gpu market is starving, demand is higher than supply = leads to high gpu price
- number of openrouter shovel sellers rises = lower token prices
Napkin math comparing a cluster with a single latest gaming GPU at a generous ~200t/s shows that the cluster pushes about 9 million times more tokens per day. And the cluster runs a more powerful model.
Truly, "the more you buy, the more you save".
If open source inference engines proliferate (in terms of adoption, across models, across inference hardware) I imagine you'd get pricing pressure on that inference rev towards base card cost, so my interpretation is this is actually a bad development for compute providers.
Do they share latency / concurrency figures (aka the deployment shape that gets them that token/s/cluster? Would be interesting if they're using deepEP/WideEP + PP and TP
$0.00051192 per M/tok ?
Assuming a 64 card cluster
thats a lot cheaper than renting a H100
Does anyone have a solid market rate dollar figure on what a single Huawei Ascend 950 costs to buy, in the 128GB or 144GB variants of it?
> A nuclear bomb
Indeed. This could actually be what breaks America's world order.
AI can already clone major products like Photoshop and Office. Maybe not at 100% feature parity for now but at a "good-enough" surface level. That will only keep improving. Eventually we could get clones of Windows and macOS at feature parity.
Even if US lobbying/corruption tries to cripple AI to prevent that, China will instigate and accelerate it.
[dead]
I'm sure Kim Dotcom's megaupload or Napster had even higher revenue for distributing stolen IP, but they did not have institutional backing.