The US vs China: Who Will Own AI and Our Future?
Can China overtake OpenAI, Anthropic and Google if its models learn from the answers of American ones — and who will own AI in the end: a few American companies, countries with their own models or billions of ordinary users?
Chinese models, in Max Grigoriev's assessment, are catch-up models: because of export sanctions Chinese labs cannot buy NVIDIA GPUs directly, there isn't enough compute for hundreds of experiments or for expensive synthetic data, and they close the gap through distillation — training on the answers of American models. That way you cannot overtake the model you distill, so the lag holds at three, six, nine months. The exception is Kimi with its Kimi Delta Attention, but how much of its success comes from the architecture and how much from distillation, the public cannot check. Ilnar Shafigullin plays for the other side: Kimi already beats GPT-5.6 Sol in benchmarks, Chinese models are many times cheaper through the API, and OpenAI's reduced limits push average users toward open code and the Chinese "zoo." Why so few countries are in the race, the hosts answer in one word — money: Max describes the AI business as a sandwich in which compute is the bread, the model is the butter and the winner takes the whole market; China, though, will hold on thanks to 1.4 billion people cut off from Western models. Access to AI is already split by country, there are almost no American open models, and Alexander Volchek is convinced that the main money lies not in the enterprise but in billions of ordinary people.
What to watch for
Key takeaways
The head of Google posted an analysis presenting the new Gemini 4 Carbon as a significantly stronger system than Anthropic's Fable and Claude Opus 5.5 and GPT-6 Astra, and Elon Musk, who for some reason keeps reposting the head of Google in particular, reacted that it was very cool. How true this is, Alexander doesn't know yet: the model is still to be tested.
Alexander has long said that neither Europe, nor the Arabs, nor Russia will catch up with the American market — while China seems to be right there. When Kimi K3 came out two months ago, Moonshot itself wrote to him that the model was on the level of Fable and Astra; several months have passed, and enthusiastic comparisons with Fable, which has been working beautifully since summer, have still not appeared.
Max doesn't say China can't do it — nothing is impossible. But Asia, as he observes, rarely leads in research, and he explains this by culture: from the first day of school children are taught to copy a character or a poem correctly, down to the strokes, and the one who copies well is the one who does well. Max himself calls this philosophy.
Because of export sanctions the Chinese can't simply buy NVIDIA GPUs: OpenAI or Anthropic come and ask for five hundred thousand cards, while Chinese labs buy through third countries — slower and more expensive. That limits experiments, training and synthetic data alike: the internet hasn't been enough for code and mathematics for a long time, and back in 2022–2023 OpenAI was hunting for data, including through the Microsoft deal for GitHub.
Chinese models, in Max's view, work well above all thanks to distillation — training on another model's answers; Anthropic recently shut down tens of thousands of accounts that apparently belonged to Chinese botnets. But you can't overtake the model you distill: the target remains a model that is already out, not the one being trained right now in an American lab — hence the lag of three, six, nine months.
Ilnar agrees that China has nothing to do at the frontier for now, but he plays for the middle class of users. The latest Kimi benchmarks compare it with GPT-5.6 Sol, and it slightly beats that model, while OpenAI is cutting limits: the $200 subscription may give ten x instead of twenty. Ilnar has already installed open code to carry out tasks on DeepSeek and Kimi — through the API it is substantially cheaper.
China, Alexander admits, has more electricity and more people: in 2011, he recounts, the government put sixty thousand people per city to work drawing 3D maps. But China, as he senses it, is not heading toward AGI and ASI, and it won't have billions of users: Google has reported a billion Gemini users, while Chinese models, by his estimate, have about half a billion registrations across all of them.
No one comes to Anna with requests for Chinese infrastructure: in her practice Chinese models are used for training tasks and simulations, where it is much cheaper. Ilnar sees a lot of Qwen in business, deployed locally so that data doesn't leave the perimeter, and Max explains: on-premise means Chinese models because only they hand over their weights, and they are popular only because they are open.
Frontier labs heavily censor their models' output — for safety reasons and for political ones. Chinese models are censored too, but differently: if you don't care about the Chinese party's politics, you can get answers that are closed off in Western models. For the same reason the political right in America, according to Max, uses Grok, and Ilnar calls Grok unprincipled: his acquaintances admire how lightly it treats restrictions.
Alexander trusts American data security far more than Chinese. If you untick the "don't train on my data" box in a Chinese model, he doesn't believe it works; in ChatGPT he does: if such data surfaced somewhere, the user, he says, would become a millionaire, and it couldn't be used against him. He is promptly asked whether he sees such a checkbox at OpenAI and whether he believes it.
The European personal-data law, in Alexander's view, is not people's privacy but state control, and here he agrees with Elon Musk and Pavel Durov; he sees Russian laws, including the Yarovaya law, the same way. In America you don't have to agree to cookies every time. Anna adds that several European countries asked for DeepSeek to be removed from app stores, though in Poland it can still be downloaded.
Kimi, according to Max, is a serious evolution of architecture: Kimi Delta Attention and other know-how that no open model had before it. Either it was extracted from Western models, or it is the lab's own research — but research takes hundreds of experiments and compute. The claimed ablations, a fifteen percent gain, the public cannot verify, and what share of the benchmarks comes from distillation, Max won't say.
Why other countries haven't built strong models, the hosts answer in one word — money: the Chinese have it, the rest don't. Ilnar describes a second path — selling compute: DeepSeek, by his account, earns $38 an hour from a single graphics card, Arkady Volozh is building Nebius after Yandex, and Elon Musk's cluster sells capacity to Anthropic. The race, by contrast, requires pouring in money all the time — which is why OpenAI and Anthropic are loss-making.
The AI business, Max explains, is selling sandwiches: compute is the bread, the same for everyone, and the model is the butter. 99% of people go to whoever's tastes better, the winner uses that money to invent even better butter, and rivals die out without money for research. That, according to Peter Thiel's book Zero to One, is how a monopoly is built; Alexander concludes that models outside the top three or five will die out.
Why Chinese models will be decent after all: in China you can't open ChatGPT, Claude or Grok, VPN detection works well, and you can't pay for a higher tier with a Chinese card. That leaves 1.4 billion people who need the technology and have money — and a model trained on them, Max says, will certainly become more Chinese.
In Russia, Alexander reminds, you can't simply open Codex, and in Europe ChatGPT works differently; for a friend of his, Dots vanished from the app as soon as he switched on a Singapore SIM card. Location, according to Max, sits in the system prompt, and companies can refuse service for any reason: such systems don't have utility status, the way an electricity seller does.
That, Max believes, is why open models matter so much — and the fact that American ones lag far behind even the Chinese is a big problem. Leaders who profit from the current situation won't support open source: in the 1980s and 1990s, integrators of UNIX mainframes died out after Linux appeared, and the industry learned the lesson — anyone can buy NVIDIA GPUs, but there is no way to get the weights of frontier labs' models.
Completely closing China's access to GPUs, Ilnar believes, will be prevented by the tension around Taiwan and smuggling through Malaysia, Indonesia and the Saudis. Max sees a calculation in these loopholes: a full blockade would force China to build its own, whereas buying through third countries is a bit more expensive and slower and makes the competitor less competitive — which is also how Russia today buys even chips for washing machines.
While the episode was being recorded, Tibo from OpenAI posted on X: on the principle of "first make it worse, then put it back," limits were doubled on all subscriptions — the $200 plan will again have twenty x, and the $500 one apparently fifty. Ilnar breathes out, while Alexander warns: half a year of fighting for permanence lies ahead, and OpenAI is said to be planning to merge ChatGPT and Work.
The companies' main interest, Max believes, will shift to the enterprise — that is where the biggest money is. Alexander objects: the money is in billions of people, not in Google's hundred thousand salespeople, and if OpenAI and Anthropic gave up making AGI for ordinary people, he would write them off. Anna adds that enterprise quickly becomes too expensive and that no one has yet learned to assess the return on AI in companies.
In the past, a candidate at a remote interview was swapped for another person; now, Anna says, an assistant can sit alongside and help get through the interview in real time, and her team is changing the questions and the very approach to assessment. Alexander asks would-be speakers to record in one take, and at the end he retells a story in which two AI assistants were left alone at an interview and couldn't end the meeting.
What this episode is about
A Sunday ToTheMoon episode in which Alexander Volchek, Ilnar Shafigullin and Anna Volchek are back to a foursome — with Max Grigoriev, a specialist in AI and machine-learning systems who worked at Google and Airbnb. The occasion is Gemini 4 Carbon: the head of Google posted an analysis in which the new model looks significantly stronger than Fable, Claude Opus 5.5 and GPT-6 Astra, and Elon Musk reacted that it was very cool. But the main topic is an opinion Max voiced during a recent recording. For three years Alexander has been saying that no one will catch up with the American market, while China keeps releasing strong models: Moonshot assured him that Kimi K3 is on the level of Fable and Astra, though months later there are still no enthusiastic comparisons.
Max explains why Chinese models remain catch-up models. Asia rarely leads in research — he links this to a culture in which, from the first day of school, children are taught to copy characters exactly. The main thing, though, is compute: because of export sanctions the Chinese cannot go to NVIDIA for half a million GPUs the way OpenAI or Anthropic can, and buying through third countries is more expensive and slower. Without compute you can neither run hundreds of experiments for a new architecture nor generate the expensive synthetic data on which frontier labs move forward today. So the Chinese, in his view, above all distill American models — Anthropic recently shut down tens of thousands of accounts that apparently belonged to Chinese botnets — and you cannot overtake the model you distill: hence the lag of three, six, nine months.
Ilnar plays for the other side — for the middle class of users. The latest Kimi beats GPT-5.6 Sol in benchmarks, Chinese models are substantially cheaper through the API, and OpenAI is cutting limits, so Ilnar has already installed open code to hand tasks to DeepSeek and Kimi. Alexander considers this a story for advanced users: China has more electricity and more people, but it won't have billions of users the way Gemini does, and he would not trust his data to a Chinese model. Anna says that in her practice Chinese models are used for training and simulations, where they are much cheaper; Ilnar says that businesses often deploy Qwen inside their own perimeter; and Max explains the popularity of Chinese models on-premise by the fact that only they hand over their weights: openness is their way to attract attention and funding.
Another reason is censorship: Western labs censor their answers heavily, Chinese models do so in a different sense, and the political right in America, according to Max, chooses the less censored Grok, which Ilnar calls unprincipled. Alexander trusts American data security the most: he believes in ChatGPT's "don't train on my data" checkbox but not in a Chinese model's, and he regards the European personal-data law, like the Russian one, as state control. Anna recalls that several European countries asked for DeepSeek to be removed from app stores. Separately, Max takes apart Kimi: Kimi Delta Attention and other know-how that no open model had before it were either extracted from Western models or came from its own research, but the public cannot check the claimed gain or the share of distillation.
Why so few countries are in the race, the hosts boil down to money: the Chinese have it, the rest don't, and a model released once is of no use to anyone two days later. Ilnar describes a second business — selling compute: DeepSeek, by his account, earns $38 an hour from a single graphics card, Arkady Volozh is building Nebius, and Elon Musk sells capacity to Anthropic. Max explains the market as a sandwich: compute is the bread, the model is the butter, 99% of people go where it tastes better, and that, according to Peter Thiel's book Zero to One, is how a monopoly is built. China, however, will hold on: 1.4 billion people are cut off from ChatGPT, Claude and Gemini by VPN detection and Chinese cards. Access is already split by country — in Russia you can't simply open Codex, Dots disappears with a Singapore SIM card, and location, Max says, sits in the system prompt; that is why open models matter so much, and America has almost none.
Cutting China off from GPUs completely, Ilnar believes, will be prevented by the tension around Taiwan, and the cards get smuggled in through Malaysia and Indonesia anyway; Max sees a calculation in this — a little friction weakens a competitor without forcing it to build its own, which is also how Russia's purchases of electronics through third countries work. By shutting itself off completely China built its own search engines, social networks and TikTok, while the Russian market, according to Max, is ten times too small for that. Toward the end Ilnar reads with relief that OpenAI has doubled limits again, Alexander expects a fight for permanence among a billion users and argues with Max: the money, in his view, is not in the enterprise but in billions of people, and Anna points out that no one has yet learned to measure the return on AI in companies. The episode closes with a conversation about job interviews at which a candidate is swapped for another person or an AI assistant.
The episode doesn't say who will win, but it clearly separates two positions: at the frontier, in Max's and Alexander's view, China, short of compute and its own research, will keep catching up with a lag of months, while for the average user, as Ilnar shows, cheap Chinese models are already good enough — especially when American companies cut limits. These are the hosts' positions, not a review of research: Max himself calls his cultural explanation of Asia's lag philosophy. The accounts Anthropic shut down, DeepSeek's $38 an hour per graphics card, OpenAI's doubled limits, the gain from Kimi Delta Attention, Gemini's billion users and Gemini 4 Carbon's superiority over Fable and Astra are given as the hosts state them and need checking.
Episode transcript
The episode is in Russian; below is an English reading guide to the transcript (the full EN transcript is a machine translation). Voice matching applied to 230 segments: 0 identified, 0 mixed, 230 probable, and 0 unresolved.
Read transcript on a separate page
Loading…