Liang Wenfeng takes $7.4 billion from the Chinese state to keep the AGI story alive.

It is true that DeepSeek’s R1, in January 2025, was a real piece of engineering. It is true that a small Hangzhou lab, on a budget that was a rounding error for the U.S. frontier-model shops, produced a model that did things comparable to OpenAI’s o1, and shipped the weights openly. Engineer-to-engineer, the achievement was genuine, and the open-weights release did some good in a market that badly needed a counterweight to closed-API concentration. The trouble is that eighteen months later the same company is selling a tenth of itself to the Chinese state, on a $50 billion valuation, and using the words “the dawn of AGI” to describe the next chapter. The first part of that sentence is a credit to Liang Wenfeng and his team. The second part is a fundraising document.

Read what the firm is doing, not what it is saying. The WeChat statement late last week declared humanity to be at the dawn of AGI and announced plans to at least double headcount across twenty-seven types of technical and functional roles — a hiring spree in which, notably, every single position is open to interns. The investors in the round include the founder himself (about $3 billion, the largest single ticket), Tencent, Contemporary Amperex Technology, and China’s National Artificial Intelligence Industry Investment Fund. The last of those is not a private actor. It is an industrial-policy vehicle of the Chinese state, set up to direct capital toward companies the state has decided are strategic. A national champion that is also a state portfolio company, valued at fifty billion dollars, cannot be described in the vocabulary of a venture-funded startup. It is, in the old-fashioned phrase, a state-backed enterprise. The only question is which state, and to what end.

Doctorow, in his October 2025 book Enshittification, named the structural pattern by which a useful product acquires monopoly rents and is then operated, with diminishing pretense, for the rent. The story has four stages: good to the users, then good to the business customers, then good for the shareholders, then dies. R1 lived at the beginning of the curve — open weights, cheap inference, no lock-in. The question a $50 billion valuation forces you to ask is which stage of the curve DeepSeek is on now, and whose money the answer is being purchased for.

I want to be precise about the technical layer, because the temptation in pieces like this is to wave the hand at “AI” as if the technology were a single thing whose promise or failure could be settled in a paragraph. The R1 release was a genuine engineering contribution: a reinforcement-learning-trained reasoning model produced at a fraction of the cost most observers had assumed necessary, with weights anyone could download. The cost numbers — the training run reportedly in the low millions of dollars, the inference economics that put pressure on U.S. labs’ per-token pricing — were the basis for the “DeepSeek shock” that wiped a few hundred billion dollars off U.S. AI-adjacent equities for a day or two. That shock was a real piece of information about cost structures in the LLM business, and Liang Wenfeng and his team deserved the credit. But the technical achievement of January 2025 and the financial structure of June 2026 are two different stories, and the second story is the one the $7.4 billion is buying.

What is the second story? The second story is a Hangzhou company that has been at this eighteen months, that has not, in the eighteen months since R1, produced a comparable shock, and that is now raising money on a valuation more than thirty times what a comparable Western lab at the same stage of model release would have commanded. (OpenAI was valued at roughly $157 billion at the end of 2024 — the figure from its October 2024 employee tender offer — having shipped GPT-4 in March 2023 and a series of subsequent improvements; Anthropic, with the Claude line, was at roughly $60 billion in early 2025. DeepSeek, with one major open-weights release and a more limited closed-API offering, is now at $50 billion. The numbers do not, on the face of them, line up with a pure engineering-merit accounting — and the comparison tightens rather than loosens when you note that the Western valuations were set on the open market while the DeepSeek number is being set on a state-aligned strategic fund.) The second story is also a hiring plan — “27 types of technical and functional roles, open to interns” — to “at least double the scale of all departments,” which is a growth-stage corporate plan, not the staffing chart of a lab that has solved the alignment problem and is now on the verge of superintelligence. If you can grow headcount to fill the AGI, the AGI is not imminent. If the AGI is imminent, you do not need to double headcount to fill it. Pick one.

Then the second story is the language. “Humanity is currently at the dawn of AGI.” The line is in the WeChat statement, attributed to DeepSeek, used to frame the hiring push. Adam Becker, in More Everything Forever (2025), has written carefully about the rhetorical operation this kind of language performs — the “TESCREAL bundle,” the longtermist sleight of hand by which a small group of actors in the AI industry channels attention and capital toward hypothetical future capabilities, and away from the documented present harms of the systems already deployed. The “dawn of AGI” framing is the same operation. It does three pieces of work at once. It relocates the moral and analytical center of gravity from the present (where the systems are biased, opaque, electricity-hungry, and displacing call-center workers in the Philippines) to a future (where the systems are godlike and someone, possibly these companies, will control them). It recruits urgency in the audience — the dawn language, like the race language, is the urgency of “if we do not act now, we will lose something we cannot get back.” And it flatters the capital that is being asked for. The investor is not buying a $50 billion share of a company that ships open-weights models at low cost; the investor is buying a position at the dawn of AGI. The two are not the same thing, and the difference between them is the bezzle.

The bezzle, in Galbraith’s 1929 sense, is the interval between the commission of a confidence trick and its discovery — the gravity-defying interval during which Wile E. Coyote is running on air and has not yet begun to fall. Crypto was a bezzle. The streaming-royalty accounting is a bezzle. SPACs were a bezzle.

Doctorow took the title for his 2024 novel and has been using the concept across the Pluralistic corpus as a general theory of tech grift. The question worth asking about this funding round is whether it is also a bezzle — whether the price being paid is a price for an asset that will, in five years’ time, look as obviously overvalued as the SPACs of 2021 look now. I do not know. The columns that pretend to know are doing something other than analysis. But the asymmetry between the technology actually shipping today and the rhetoric the technology is being shipped under is large enough that the question is honest.

Cui bono, then, on the round? The biggest single beneficiary is the founder himself. Liang Wenfeng held close to 90 percent of DeepSeek before this financing, on a valuation that has been variously reported in the low single-digit billions. He has just contributed $3 billion of his own money to a round that values his remaining 80 percent at roughly $40 billion, give or take, on paper. Whether the paper is real depends on whether somebody else, in three years’ time, is willing to pay a multiple of $50 billion for the same asset, or whether the next round is a down round, or whether the company is acquired at a discount. I am not in a position to know, and neither is anyone outside the cap table. The second-biggest beneficiary is the Chinese state, which now has a financial position in a national AI champion and a strategic influence over its roadmap. The third is Tencent — which is in the middle of its own repositioning of Weixin as the AI gateway of choice for the China market — and which now gets a piece of a non-U.S. AI platform and a toe-hold in the inference layer. The fourth is Contemporary Amperex Technology, the battery company, which I expect is in the round for the same reason battery companies were in solar-panel joint ventures a decade ago. The last beneficiary is the open-weights community, in a thin way, because the existence of a state-backed open-weights champion is a strategic counterweight to closed-API concentration, and that is a real benefit, even if it is not the benefit the founder’s statement emphasizes.

Kate Crawford, in Atlas of AI (2021), wrote that “AI is made from vast amounts of natural resources, fuel, and human labor.” The labor in the DeepSeek round is the 27 new technical-and-functional roles, open to interns. The interns are the cost. The cost is being socialized onto the young engineers of Hangzhou, who will be paid at Chinese-tech-internship rates, in a labor market that is now flush with state capital, and will be told they are working on the dawn of AGI. Some of them will be. Most of them, if past rounds at frontier labs are any guide, will be doing the platform-engineering work of running inference at scale, keeping the API up, building the chat product, writing the documentation, doing the security review, doing the data-labeling pipelines that are still, even in 2026, mostly done by underpaid contractors in lower-wage jurisdictions. The gap between “we are at the dawn of AGI” and “we are hiring interns to run the platform” is the actual product. It is also the gap the rhetorical operation is designed to obscure.

The discipline of cryptographic-protocol verification teaches a hard rule: you cannot verify a system whose specification you cannot write down. AGI is not a specification. It is a marketing claim designed to justify the consumption of entire regional power grids and the drilling of deep aquifers for data-center cooling. There is an old tradesman’s observation that the work does not care how you feel about it. The interns they are hiring will be doing the data-labeling and the grunt work of tuning weights, while the valuation gets marked to market on the strength of the AGI claim.

The architecture of extraction does not change when the flag changes; the monopsony just gets a new state backer. Tencent, which recently moved to position Weixin as an AI gateway, is now underwriting DeepSeek’s labor expansion. Harold Innis would recognize the political economy here as a staples-extraction model updated for the compute age: the staple is no longer fur or timber but compute capacity, and the metropole is the state-capital nexus. The “AI race” framing is a manufactured urgency that allows states and hyperscalers to socialize the infrastructure costs — land, water, subsidized electricity — while privatizing the monopoly rents on the inference layer. The public subsuming the infrastructure cost deserves to see the actual thermodynamic and financial ledger, not a WeChat press release declaring the dawn of a new era. Capital raises of this magnitude, pitched on the promise of general intelligence, should be regulated as securities offerings for unproven technology, with mandatory compute-capacity, water-usage, and energy-consumption reporting; the EU AI Act’s general-purpose AI obligations and the SEC’s tightening posture on AI-revenue disclosure point at the right direction.

I am not, to be clear, saying the Chinese state is wrong to make this bet. Industrial policy in support of an open-weights AI champion is, in fact, the kind of industrial policy a state can be forgiven for making, and the existence of a non-U.S. open-weights frontier model is a genuine public good in a market that is, on the U.S. side, increasingly concentrated in three or four closed-API shops. The problem is not that the Chinese state is putting money in. The problem is the framing. The problem is that the announcement language — the dawn of AGI, the doubling of headcount as a strategic necessity — is a fundraising document written in the vocabulary of a religion, and the religion is being used to direct capital that will be measured, in the end, against something much more ordinary. The AGI, if it ever arrives, will not announce itself. It will, like every other piece of infrastructure, arrive as a partial capability, owned by someone, deployed somewhere, distributed unevenly, and then enshittified. Doctorow’s four stages do not pause for national champions. The rent will be extracted in the same way, by the same mechanisms, regardless of which state’s flag is over the headquarters.

The Gerdau acquisition of the Manitoba Rolling Mills in 1995 is the bit of Canadian content in this piece, and it is the reason the round makes me uneasy. A small, technically capable firm — in that case a steel mill, in this case an AI lab — acquires paper wealth in a moment of outside capital arriving, the headcount doubles, the language gets grand, and the firm is no longer the firm it was. The pattern is older than the mechanism, and the mechanism is the same. Capital, finding a thing that works, rides it until the thing no longer works, and then leaves. The Chinese state is making a different bet than the venture capitalists are. The venture capitalists are betting on the bezzle lasting. The Chinese state is betting on the AGI lasting. Both bets are on a future that has not been specified. The deadline that actually matters is the next round. The next round is when the AGI rhetoric meets the cap table. Deadlines are the only part of fundraising processes that the funded actually respect.