DeepSeek has retained four underwriters including CITIC Securities for a STAR Market listing, alongside a pre-IPO round valuing it at about 500 billion yuan (roughly $74bn) pre-money. Timing, size and valuation are all reported as undecided, and sources disagree on the year. The analytical problem is the prospectus: open weights foreclose the licensing line a listed AI company would normally disclose, forcing inference, enterprise deployment and strategic position to carry the revenue section.
DeepSeek has retained four underwriters, CITIC Securities among them, to prepare a listing on Shanghai's STAR Market, according to reports that emerged on 9 and 10 September 2026. The Hangzhou company is separately completing a pre-IPO financing round that would value it at about 500 billion yuan, roughly 74 billion US dollars, before the new money goes in.
Nothing has been filed. Timing, fundraising target and final valuation are all reported as undecided, and the sources disagree about the year: several say the process begins during 2026, while the Wall Street Journal has been reported as pointing to a 2027 listing. Every figure below is reported rather than disclosed, and that distinction is load-bearing.
The interesting problem is not the valuation. It is the prospectus. DeepSeek's distribution advantage comes from releasing open weights, which means anyone can download the model and run it, and that choice forecloses by construction the licensing revenue line a listed AI company would ordinarily point an exchange at.
The STAR Market, launched in 2019, exists precisely because Shanghai's main board could not list China's technology companies. Its admission criteria include alternatives keyed to market capitalisation combined with revenue or research spending, rather than to accumulated profit, which is what makes an unprofitable hard-technology issuer admissible at all. On that dimension DeepSeek has no obvious problem. Loss-making frontier research is the category the board was designed to accommodate.
The registration-based regime that governs it, however, is a disclosure regime. The exchange's review is not primarily a judgement about whether the business is good; it is an interrogation of whether the issuer has explained the business fully and consistently. Reviewers put follow-up questions in writing, and those exchanges are published. An issuer whose commercial model is unusual does not get to gesture at it and move on.
So the question becomes concrete. What goes in the revenue section for a company whose flagship product is free to download?
There are answers, and DeepSeek has been visibly assembling them.
The first is inference. Open weights mean anyone can run the model; they do not mean anyone can run it cheaply at scale. DeepSeek sells API access to its own hosted deployment, and in August 2026 it raised those prices substantially and introduced peak and off-peak tiers, a move this publication covered at the time as a repricing of the industry's assumption that inference costs only ever fall. Read against a prospective listing, that repricing looks less like a pricing experiment and more like the deliberate construction of a revenue line a filing can describe.
The second is enterprise and state deployment: private installations behind a customer's own firewall, support contracts, tuning, and the integration work that a downloaded weights file does not perform by itself. This is the standard open-source commercial answer, and it has the standard open-source problem, which is that it scales with headcount rather than with copies.
The third, harder to write into a financial statement, is strategic position. In a domestic listing that is not nothing, because the buyer base includes institutions for which national AI capability is an explicit mandate rather than an externality. A prospectus cannot put that in the revenue table, but the underwriters can put it in the room.
What none of these produce is the clean, high-margin, rapidly compounding licensing line that makes a software prospectus easy to write. Open weights convert what would have been a licensing business into a distribution business, and distribution has to be monetised somewhere downstream. The filing has to say where.
The 500 billion yuan figure is better read as a mechanism than as a valuation.
There are no public comparables for this company. No listed Chinese AI laboratory trades on the mainland at a scale that would anchor DeepSeek's price, and the obvious international reference points are private as well. Where comparables are absent, a pre-IPO round with named strategic investors functions as the price-setting instrument. It produces a number with real money behind it that underwriters can then defend in a filing, in front of a regulator, to a domestic institutional base.
The composition of that round supports the reading. Tencent and CATL have been reported as the largest external shareholders, at 10 billion and 5 billion yuan respectively. A platform company and a battery manufacturer are not financial investors looking for a quick markup on exit. Strategic capital at a set price is exactly what an issuer wants on the register when it needs the price to hold through a listing.
For scale, the reported 74 billion dollar figure sits against the roughly 2 trillion dollar target attached to Anthropic's prospective listing in reporting this year. That is a gap of something like 27 times between two companies shipping into the same category of product. Neither number is a market price, and the comparison is not an argument that one is wrong. The size of the disagreement is the useful fact. It says the market has no settled method for valuing a frontier AI laboratory, and that a listing in Shanghai and a listing in New York are being priced by populations of investors who are not arbitraging each other.
The Financial Times reported on 9 September 2026 that founder Liang Wenfeng has been personally screening prospective investors, partly in order to preserve control ahead of the listing.
That is a more consequential detail than it reads as. Founder control at a US listing is a solved problem. Dual-class share structures are ordinary, and a founder can take a company public while retaining a decisive voting block more or less as a matter of course. On the mainland the settlement is tighter. The STAR Market is one of the few Chinese boards that permits differential voting rights at all, and it permits them subject to conditions rather than as a default, alongside lock-up obligations on controlling shareholders that constrain what they can do after listing.
Screening the register before the IPO is the workaround. If you cannot rely on a share structure to guarantee control after listing, you can select who holds the shares beforehand, favouring strategic holders with reasons to be passive over financial investors with reasons not to be. Which constraint a founder chooses to engineer around is a reasonable signal of which one he expects to bind.
No prospectus exists. The valuation, the timing, the investor screening and the shareholder composition are all reported by third parties and none of it has been filed with or confirmed by an exchange.
It is tempting to treat this year versus 2027 as a rounding error in early reporting. It is not.
A mandate to underwriters is cheap and reversible. Chinese IPO processes involve a tutoring and counselling period with the regulator before a filing is even accepted, and the gap between retaining banks and actually listing routinely runs to years, or ends in nothing at all. If the process genuinely begins this year, the market gets a filing, audited financials and, for the first time anywhere, a public revenue breakdown for a major open-weights AI laboratory. That document would be the most informative thing published about the economics of open-weight models to date, regardless of whether the listing completes.
If the real date is 2027, then what happened this week is a company joining a queue.
The single document that would settle most of what is speculated here is the prospectus. Until one exists, the valuation, the timing and the revenue model are reported intentions, and the most defensible sentence anyone can write about DeepSeek's listing is that four banks have been hired.
Open weights are a distribution strategy rather than a revenue strategy. The money has to come from somewhere adjacent: hosted inference sold by API, where the vendor's scale and hardware access make it cheaper than self-hosting for most customers; enterprise deployment, support and tuning contracts; and in some cases hardware, cloud or platform arrangements downstream. The trade is reach for margin. A licensing business earns per copy, while an open-weights business gives away the copies and has to capture value from whatever remains scarce, which is usually operations rather than software.
In August, DeepSeek raised API prices by up to 1,100% and introduced peak and off-peak tiers. Read against a listing, that looks like a revenue line being built.
Read the DeepSeek pricing analysisOracle reports Q1 FY2027 today carrying $638 billion in remaining performance obligations. That is roughly eight years of revenue at the current run rate, and it is not cash.
ISS and Glass Lewis both say reject it. Neither recommendation can change the result. At a company where one holder controls 86.4% of the stock, a say-on-pay vote stops being a decision and becomes a disclosure instrument.
The headline was the 94% drawdown from a $3.44 billion peak. The more useful arithmetic is the cash sitting inside the thing being bought — and the fact that everyone except the angels was paid in another private company's paper.