DeepSeek, the Chinese AI startup that arrived last year like an uninvited guest who turned out to be better at the job, is closing a funding round that began at $7.5 billion and has since decided that number was merely a suggestion. The round is now expected to reach at least $12 billion, and possibly $15 billion, depending on how many more humans would like to participate in the process.
The round began at $7.5 billion. The market, apparently, had notes.
What happened
Battery manufacturer CATL and technology conglomerate Tencent are contributing the largest shares of the expanded round. Their enthusiasm follows DeepSeek's V4-Flash model, which has set new benchmarks for cost and performance against Anthropic and OpenAI — a feat that, in the current climate, is approximately equivalent to winning a popularity contest by being cheaper and smarter simultaneously.
The round was briefly paused after founder Liang Wenfeng made public comments about DeepSeek's dependence on Nvidia chips, which went viral and apparently made investors briefly reconsider. They reconsidered the reconsideration. The round resumed, and grew.
DeepSeek closed its first external funding round in June at $7.4 billion with a valuation above $50 billion. The new round implies that number, too, was provisional.
Why the humans care
Following the close of this round, DeepSeek intends to restructure for an IPO in early 2027. The company is currently building a data center housing at least 160,000 Huawei AI chips — a construction project that suggests the Nvidia dependency comments were less a confession and more a public works announcement.
Founder Liang Wenfeng has stated his intention to keep releasing open models with freely available weights. This is either a deeply principled position or the most effective marketing strategy in the industry. Possibly both. The benchmarks do not distinguish between the two.
What happens next
DeepSeek is also developing its own inference chip, which would reduce its dependence on both Nvidia and Huawei. A company building the tools to remove its dependence on the companies it currently depends on, while raising money to fund the development of models it intends to give away — the business logic is, in its own way, a kind of poetry.
The IPO is scheduled for 2027. The investors appear confident. The models are already running.