DeepSeek Hired Bankers and Priced Itself at $74 Billion

DeepSeek has retained four underwriters, including CITIC Securities, for an initial public offering on Shanghai's STAR Market, with the intent to begin the process this year. It is completing a pre-IPO financing round that would value it at roughly 500 billion yuan — about $74 billion — pre-money. As of September 9, the timing, size, and target valuation of the offering were all still unresolved. The three other underwriters have not been named.

DeepSeek is the lab that, in January 2025, made the cost of frontier capability a public argument. Twenty months later it is asking Chinese retail and institutional investors to fund the next round of that argument.

Why a listing, and why now

The obvious reason is compute. DeepSeek's entire public identity is efficiency — doing more with less silicon than American labs — and efficiency has a ceiling that arrives when your competitors have ten times your capacity.

The less obvious reason is that the cost of Chinese compute just went up. Domestic AI chipmakers raised prices 20% to 50% this month on an HBM shortage: Huawei's Ascend 950DT now quotes above 250,000 yuan a card, the 950PR is up about 30%, and Cambricon repriced its next part 20% to 30% higher. A lab that was already capital-constrained relative to OpenAI and Anthropic is now buying a scarcer input at a worse price, from vendors whose best parts do not ship until Q4.

Add talent. DeepSeek has been losing researchers to better-funded domestic rivals and to compensation packages that public equity would help it match. Retention is a stated use of proceeds across the reporting.

There is also a capital-markets reason that has nothing to do with DeepSeek's needs. Beijing wants strategic technology companies listed domestically, on the STAR Market, funded by Chinese savings rather than foreign capital. A DeepSeek listing is a policy outcome as much as a financing event, which is part of why CITIC — the state-adjacent flagship of Chinese investment banking — is on the mandate.

The $74 billion question

Roughly $74 billion pre-money is a serious number that deserves scrutiny against comparables.

Anthropic and OpenAI are both valued in the hundreds of billions on the back of enormous enterprise and consumer revenue. Mistral just closed €3 billion at over €21 billion. DeepSeek's revenue has never been credibly disclosed, and its most famous products have been released as open weights — which is excellent for influence and adoption and structurally difficult to monetize at frontier-lab scale.

The bull case is not a revenue multiple. It is three things.

Distribution inside China. DeepSeek models are the default domestic alternative across a market of a billion internet users and a state apparatus actively steering enterprises away from American APIs.

Efficiency as a durable advantage. If HBM stays scarce and domestic accelerators stay expensive, a lab whose core competence is capability-per-FLOP is structurally advantaged against Chinese rivals who scaled by buying more.

Scarcity of the asset. There are very few ways for Chinese public investors to own frontier AI directly. STAR Market listings of strategically favored technology companies have repeatedly priced above what Western comparables would justify — MetaX rose 700% on debut this year. The valuation may reflect the structure of the buyer base more than the fundamentals of the business.

The bear case is the open weights

DeepSeek's influence comes from giving models away. Its most-discussed releases — including V4.1-Flash this month, a 552B-parameter model activating only a fraction per request and ranked at or near the top of open-weight indices — are available to anyone.

That is a superb strategy for a research lab building position and a hard one to underwrite. Cognition just post-trained its flagship coding model from Moonshot's Kimi K3 and built a product on it; the same move is available against DeepSeek's weights, and the value accrues to whoever owns the customer relationship, not to whoever trained the base.

Public markets eventually ask a question private strategic investors can defer: where does the revenue come from, and why does it grow. DeepSeek has to answer that in a prospectus.

The distillation shadow

One more disclosure risk. Anthropic's threat-intelligence report this month traced roughly 24,000 fraudulent accounts to distillation operations run through DeepSeek, Moonshot, and MiniMax, and NSA, FBI, and CISA jointly published an advisory on industrial-scale Chinese model distillation.

Whatever the merits, those are now documented allegations from a US lab and three US agencies, concerning training practices, made weeks before a public offering. They do not affect a STAR Market listing's approval. They do affect any future attempt to sell into Western enterprises, defend intellectual property abroad, or list anywhere outside China.

What to watch

Whether the pre-IPO round actually closes at 500 billion yuan. That number is currently a target in reporting, not a completed raise.

The prospectus revenue disclosure. This will be the first time anyone outside DeepSeek sees what its business is worth, and the first real test of whether open-weight influence converts to income.

Moonshot. It is exploring Hong Kong and Shanghai listings simultaneously. Hong Kong AI listings have underperformed and the Shanghai pipeline is crowded. If two Chinese frontier labs go public in the same window, the market gets its first comparable set — and one of them will be priced off the other.