Appetite comes when you get up. Barely a month after having completed the largest round of financing ever achieved by an artificial intelligence company in China, DeepSeek is already studying a new fundraising. The Hangzhou laboratory, propelled to nearly $59 billion valuation in mid-June, is looking for additional capital to finance a titanic project: in-house chips, giant data centers and rapid recruitment.
Key Points
DeepSeek is exploring a new fundraising round a month after completing a record round of $7.4 billion, according to the Financial Times. The funds target computing power: in-house inference chip, data centers in Inner Mongolia and doubling the workforce. The gap remains massive with Anthropic ($965 billion) and OpenAI ($852 billion)
7.4 billion dollars, already digested
DeepSeek therefore began preliminary discussions this week with new investors with a view to launching a new funding round which would value the company around $71 billion before the operation, according to two sources familiar with the matter. In mid-June, DeepSeek closed its very first round of external financing: 51 billion yuan, or approximately $7.4 billion, an all-time record for a Chinese AI company. The post-investment valuation is close to 400 billion yuan, up to $59 billion, making the laboratory the most valuable AI startup in the country. The casting of the previous round says a lot about the importance Beijing attaches to it. If founder Liang Wenfeng signed the largest check, approximately 20 billion yuan (nearly 3 billion dollars) out of its own pocket, Tencent would also have injected some 10 billion yuan and the battery giant CATL around 5 billion. But above all, the national AI investment fund, the real armed financial arm of Beijing in its AI offensive, was also involved. As is often the case in China, everything was done in a locked structure: private capital passed through a limited partnership led by Liang, without voting rights and with a five-year blocking of shares. Only the state fund obtained direct capital participation. The founder keeps his hands free but Beijing keeps watch.
Chips, servers and brains to finance
Why return to the investor table so quickly? Because the race for computing power devours everything on his way. At the beginning of July, Reuters revealed that DeepSeek is developing its own AI chip dedicated to inference in order to reduce its dependence on Huawei but especially on the American Nvidia. A challenge for independence which requires years of development but above all considerable capital. In fact, the most discerning eyes will have noticed that DeepSeek has been publishing job offers for build data centers in Ulanqabin Inner Mongolia. Bloomberg reported at the end of June a target of doubling of staff in all departments with 33 positions open at the end of June. Because there is urgency: the V4 model, presented in April as the new open source benchmark, although impressive, remains behind the best American models according to independent evaluations. But not by much: some sources on the American side put this delay at about eight months.
The race for scale against the American giants
This race for scale should not mask a more nuanced reality for Beijing: the valuation gap and financing with American laboratories remains immense: Anthropic weighs 965 billion dollars after raising 65 billion, OpenAI posted 852 billion in March. Even valued at 59 billion, DeepSeek still weighs ssixteen times less than Anthropic. But comparing these amounts as if they corresponded to the same unit of production would be misleading: in China, one dollar invested does not necessarily finance the same quantity of capacity. Labor and operating costs are lower, integration with local ecosystems is often tighter, and architectures designed to get the most out of less capable or rarer hardware can reduce the marginal cost of training and inference. DeepSeek has notably focused on Mixture of Experts type architectures and on engineering intended to limit the computing power mobilized per request. As well as on slightly less acceptable techniques… More generally, the choice of open source (or more precisely open weight models) also significantly changes the equation. Where American players largely seek to protect their models and directly monetize access via API or subscriptions, DeepSeek rather aims for the organic adoption of its solution via diffusion by developers as well as the rapid improvement of the ecosystem around its models. These strategies lower user pricespromote local deployments and aim to transform part of the technical community into innovation relays. The other side of the coin is that they also limit the direct capture of income and do not eliminate the growing needs for calculation. The search for external capital therefore signals less the abandonment of frugality than a change of scale: Chinese efficiency can reduce costs up to a certain level, but raw power issues still persist. Today more than ever in the race for AI, money is power.