The price of admission: what DeepSeek's shadow market says about Chinese AI

Everyone wants into DeepSeek; almost nobody can buy it directly. The Chinese lab, worth about $52bn when it closed its first-ever outside funding round in late May, is raising again at a reported valuation near $71bn before the new money — a jump of more than a third in under two months, with an IPO on Shanghai's STAR Market targeted for 2027. The arithmetic was already unusual: annualised revenue of roughly $500m against a private price tag 140 to 175 times that figure. What makes the deal stranger is how it is being distributed. Because the official syndicate is effectively closed, brokers are stitching together "shadow" access through special-purpose vehicles stacked as many as eight layers deep, charging up to 18% upfront and taking a cut of any eventual profit. The people paying most for DeepSeek hold the worst claim on it.

That gap — between a headline valuation and the mechanics of owning it — is the real story. DeepSeek's round is less a discovered market price than a manufactured one, set by scarcity, benchmarked against a single listed peer, and subsidised at the top by the least informed participants. What it signals about Chinese AI is not that the technology works, which it does. It is that a great deal of money is chasing very few ways to own a national champion.

A champion, valued to order

DeepSeek's trajectory compresses two years of hype into two quarters. The "DeepSeek moment" of January 2025 — a cheap open-source model that made American labs' capital spending look profligate — turned the lab from obscure to indispensable to Beijing's ambitions. The commercial machine followed. It closed an outside round in late May at a $52bn post-money valuation, raising about $7bn from a coalition that ran from Tencent and JD.com to CATL, a battery maker, and China's state-backed national AI fund. Founder Liang Wenfeng reportedly put in about $3bn of his own money. Within weeks the company was in talks for another round at roughly $71bn.

The revenue behind that is real but thin. The Information reported annualised revenue approaching $500m, built on a newer model, agent workloads sold by the seat at $20–99 a month and tens of thousands of enterprises on its application programming interface. One estimate puts the figure twice as high, which is itself a warning about how little is reliably known. Against that, DeepSeek is projected to spend about $1.6bn on infrastructure this year, roughly three times its revenue and ten times last year's outlay. At the current price, an investor is paying 140 to 175 times revenue for the expectation that inference, licensing and agents compound far faster than the enormous capex that supports them.

A shadow market, and who pays the top

The scarcity that marks the price up is partly manufactured by the deal's own terms. The shares carry no voting rights and a five-year lock-up, preserving Mr Liang's control with minimal shareholder pressure. Direct allocation is so hard to obtain that access has migrated to a secondary market built by intermediaries. One Hong Kong family office reported being offered positions routed through eight separate vehicles. Management fees run to 6–8%, against a customary 2%; in lower tiers, upfront fees exceed 15% and profit-sharing (carried interest) reaches 40%, against a typical 20%. The structure is opaque by design: each layer thins the information before the underlying asset is reached and thickens the cost.

The precedents are instructive. Ant Group's 2020 grey market collapsed when Beijing pulled its IPO; WeWork's pre-IPO secondary market let hype outrun its balance sheet. DeepSeek is not arranging these vehicles, and there is evidence the issuer itself is uneasy: Mr Liang is reportedly reviewing limited partners to eject opaque ones, a sign he worries the tangle could complicate his listing. The round was even paused in late July after a leak of his private remarks conceding that China trailed America by 12 to 18 months in funding and computing power and remained dependent on Nvidia chips — sensitive admissions for a "national AI champion." The pause did not last; the demand was too strong.

What it says about the sector

The natural reference point is Zhipu, a rival that listed in Hong Kong as Knowledge Atlas Technology and whose market capitalisation topped HK$1 trillion (about $128bn) in June, off the back of its own retail wave. DeepSeek, private at roughly $71bn, is pricing itself as a discount to that live quote. Together the top independent Chinese AI houses are said to be worth more than $140bn. The mechanism is self-reinforcing: with the state and big tech embedded on the cap table, an IPO that came in below the last private round — a "down round" — would be embarrassing for everyone involved. So the incentives push each new valuation higher, anchored to whichever peer is listed, rather than to audited profit.

The deeper point is structural, not sentimental. Chinese AI has abundant domestic capital — state funds, deep-pocketed platforms, a $125bn investment landscape — and very few liquid ways to own it. That combination inflates every pre-IPO price against its neighbours and hands the top of the stack to brokers, not to sophisticated buyers. The shadow market is a symptom of money pressing against scarcity, not of fundamentals. It rewards the issuer and the intermediary; the investor paying 18% to get in is doing the funding end of the deal's work.

The real test will come at the IPO, when audited revenue must be disclosed and the private price must be met by public buyers at something near Zhipu's cap. If the audit confirms $400m rather than $1.1bn, the distance between $71bn private and public absorption becomes suddenly visible. For an investor watching from outside China, the lesson transfers beyond this one stock: when a private valuation is surrounded by a market that inflates it and an intermediary stack that taxes it, the price is a statement about scarcity and exit, not about the business. The steepest fees in the shadow market are the surest sign that the marginal buyer is being separated from their money by structure, not by conviction about the model.