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...