HANGZHOU, 27 AUG 2026 — DeepSeek has restarted a funding round seeking close to US$8bn at a valuation of around 500 billion yuan, roughly US$74bn, with signing expected in late August. It is preparing for a possible listing on Shanghai's STAR Market, with a debut targeted for 2027.
The company took its first external investment in June. This is its second round in about two months.
Twice in a quarter, after years of none
DeepSeek was unusual among frontier laboratories in operating for years without outside capital, funded from the quantitative trading firm its founder also built. That is the context for what is happening now.
The June round raised roughly US$7.4bn at a post-money valuation near 450 billion yuan. The round now closing seeks a similar amount at around 500 billion yuan, an increase of about eleven per cent in the mark over roughly two months.
You can read this two ways. One is simple momentum: results improved, the market repriced, and the company took more money at a higher figure. The other is pre-listing appetite: a company heading for a flotation wants a larger balance sheet and a broader shareholder register, and a second round delivers both.
The two are not mutually exclusive. That eleven per cent is the number to watch — a material increase, and not the kind of dramatic repricing that signals a sudden breakthrough.
A battery maker in an AI round
Alongside existing backers, the round includes Contemporary Amperex Technology, the world's largest manufacturer of electric vehicle batteries. That is not an obvious name on a frontier AI cap table.
The parties have not confirmed the logic, but a few readings are plausible. Energy storage and artificial intelligence meet at the data centre, where grid-scale batteries are increasingly used to buffer the enormous, spiky loads that training and inference impose. A battery manufacturer might reasonably want a position in the industry driving that demand.
The simpler reading is that large Chinese industrial companies with substantial cash are seeking exposure to the country's most prominent AI developer, and that strategic rationale follows the investment rather than preceding it.
The name reliably signals the character of the capital: domestic industrial and venture money rather than international growth investors. That shapes who the company has to convince, and how.
The pause is worth a sentence
The round was reportedly halted last month after the founder was unhappy about leaked remarks becoming public, then resumed.
Companies do not usually halt a funding round over a press leak. That this one did suggests an unusually founder-driven process and a company that treats publicity as a cost rather than an asset. It is consistent with a laboratory that released capable open-weight models while its founder gave almost no interviews.
It is a small detail about temperament, and temperament matters when a company is about to acquire public shareholders who expect regular disclosure.
What eight billion dollars buys under export controls
The money at this scale goes overwhelmingly to compute, and that is where a Chinese laboratory's problem differs from an American one raising the same amount.
A lab in California converts capital into accelerators by placing an order. A lab in Hangzhou converts it into whatever it can lawfully obtain, which is a narrower and more expensive set. Beijing has permitted limited H200 shipments to its major technology groups, a previous-generation part, while current-generation hardware remains outside the legitimate channel — and the consequences of trying the other channel were on display when Taiwan indicted nine people over diverted servers this week.
So the same sum buys less capability, and the gap is not closed by raising more. What it can be partly closed by is engineering: DeepSeek's reputation rests substantially on getting competitive results from less hardware, through training efficiency rather than brute scale.
This is an advantage born of constraint. A laboratory that must be efficient will be, and it would still rather have the chips.
Three laboratories, three venues
The choice of listing venue is the most informative decision in this story, and it is clearer when set beside the alternatives being taken elsewhere.
DeepSeek is heading for Shanghai's STAR Market, where Unitree listed this month to enormous retail demand. Moonshot AI is reportedly preparing for Hong Kong. Anthropic has filed confidentially for a listing in the United States, where the questions it faces are about data centre opposition and compute constraints.
Each venue offers a different investor base, a different disclosure regime and a different tolerance for a company that burns cash while it builds. A STAR Market listing draws heavily on domestic retail participation and on a policy environment that actively favours strategic technology sectors. It is a friendlier venue for this kind of company than New York, and a narrower one.
The valuation gap is larger than the capability gap
US$74bn sits against a most recent private mark for Anthropic near US$965bn, and against float expectations discussed in the trillions. DeepSeek's models are generally held to be competitive, so the gap is not principally about capability.
The gap comes down to revenue, market access and currency. DeepSeek competes on price in a bruising domestic market, its access to Western enterprise customers is politically constrained, and it will be valued in yuan against local comparables rather than in dollars against American software multiples.
The interesting implication runs the other way from the usual framing. If the models really are close and the valuations are an order of magnitude apart, then most of what separates them is distribution and the willingness of large customers to buy — which is precisely what Moonshot's hyperscaler talks are an attempt to fix.
What it means from here
For businesses in the region, a well-capitalised DeepSeek preparing to list means continued downward pressure on what capable models cost. A company raising to spend, in a market that competes on price, does not then raise prices.
The thing to plan for is not today's price but the dependency it creates. These cheap, capable models are funded by investment rounds rather than by operating revenue, so the price reflects a growth phase rather than the sustainable cost of service. Anyone building a business on current inference pricing should know which of those two they are relying on.
The filing date matters more than the funding date. A STAR Market prospectus would force the company to publish its first audited figures for revenue, compute costs and customer concentration, and those numbers will settle several arguments that have so far rested entirely on model benchmarks.
