> 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.

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.

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 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.

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?

I don't think i'm paying more than Deepseek price, ever

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?

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.

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.

The railroads went bust, twice, not because there wasn't profit available, but because investment outpaced profit and wasn't sustainable.

> 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?

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?

>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.

Those arguments will go away when spending on AI infrastructure is no longer some multiple of gross annual software sales revenue, yeah?

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.

> "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.

revenue != profit?

> 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.

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.

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.

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.

[dead]