DeepSeek, the Chinese artificial intelligence lab that forced price cuts across the industry with cheap, openly published models, is seeking at least 80B yuan, around $12B, in new funding. Tencent is leading the round alongside the battery maker CATL (Contemporary Amperex Technology Co., Limited), and Bloomberg reported the talks on 6 October, citing people familiar with them. The total could approach 100B yuan against an original target of 50B yuan.
DeepSeek has announced nothing, and the figures carry the caveats that go with any deal described before it closes. Demand reportedly rose after the company's most recent model release, and the same report said DeepSeek is exploring a listing on the STAR Market, Shanghai's board for technology companies with the brokerage CITIC Securities handling preparations, without settled timing or terms.
This would be the second raise of the year. An earlier 2026 round brought in roughly 50B yuan at a post-investment valuation near 400B yuan, about $59B, and the composition of that round was unusual.
Liang Wenfeng, who founded DeepSeek and runs it, put in around 20B yuan himself, close to half the total, which kept direction of the company with him while outside capital came in. Tencent contributed about 10B yuan and CATL about 5B, alongside JD.com, NetEase and China's national AI investment fund.
The Models Explain The Demand
What changed this year is a product line that competes on price rather than on being first. DeepSeek released V4 in April in two sizes: a 1.6T-parameter version with 49B parameters active at any moment, and a 284B version with 13B active, both handling 1M tokens of context. An updated smaller model followed in September. V4 scores 80.6% on SWE-bench Verified, a coding benchmark, which puts it ahead of other openly published models while trailing the frontier systems by an estimated three to six months.
The licensing is the part competitors feel. DeepSeek publishes the weights under an MIT licence on Hugging Face, which means any company can download the model and run it on its own hardware. It can also adapt the model and sell the result, without asking permission or paying anything.
Then there is the price of the hosted service, which fell 75% in May. V4-Pro now costs $0.66 per 1M input tokens off-peak and $1.98 per 1M output, while the smaller model runs at $0.15 and $0.60. On output that works out roughly 34 times cheaper than GPT-5.5.
A company giving its models away and charging a fraction of the going rate for the hosted version is buying market position with capital, not revenue. That is why the size of this round matters more than it would for a company selling at a margin. Cheap open models also push capable systems onto smaller hardware, which is the same shift that has turned the module into a computer at the edge of networks.
An Unusual Investor List
Tencent is a games and social media company, and CATL makes batteries for electric vehicles, with neither an obvious backer of a frontier AI lab by the standards of Silicon Valley. That reflects where money for Chinese AI companies actually comes from. The pools available are industrial groups, internet platforms and state-linked funds rather than the venture firms and cloud providers that finance American labs, and the national AI investment fund sitting in the previous round makes part of this state-aligned capital.
The founder's own stake changes the usual dynamic too. Liang funding nearly half of a round himself is not a structure most companies could attempt. It is also the reason a lab taking large sums from outside investors has not had to concede control of its direction.
Export Controls Decide What It Buys
The harder question is whether $12B converts into the compute it would elsewhere. American rules introduced in May 2026 require an export licence for any advanced AI chip heading to China or Macau, covering Nvidia's current Blackwell and Rubin lines and closing the route through third countries. A limited exception allows H200 sales to about 10 approved Chinese companies, capped at 75,000 units each and carrying a 25% revenue-share condition.
The market has already moved. Nvidia's share of AI chips sold in China fell from 95% in 2023 to around 8%. Huawei now holds 50% to 60%, on about $12.1B of annual Ascend revenue.
Domestic supply has its own ceiling, and it is not the processors. SMIC, China's largest chip foundry, can produce the processor dies for more than 1M Ascend chips a year, but China's high-bandwidth memory production holds actual shipments below 300,000 units without stockpiled foreign supply.
DeepSeek has split the difference in response. It has moved inference and serving, the work of answering users, onto Huawei Ascend and Cambricon hardware. Pre-training for V4 still relied on Nvidia H800s bought before the current restrictions.
So the constraint on a Chinese lab is not capital. It is memory and advanced silicon, which no amount of yuan reliably purchases, and the efficiency work DeepSeek is known for grew directly out of that limit.
The Race Is Mostly Domestic
Set against its neighbours, DeepSeek is well funded rather than dominant. Zhipu AI carries a valuation around $95B, DeepSeek sat near $59B before this round, Moonshot AI at $30B and MiniMax at $17.7B. Chinese AI valuations have risen through 2026 in a market where several labs are publishing competitive open models at once.
DeepSeek's history makes its position odder still, because it took no external money until this year. Founded in July 2023, it ran instead on High-Flyer, the quantitative hedge fund Liang also built, and a lab with no investors produced the model that moved American markets in January 2025.
Not everything has landed. R2, the follow-up to that reasoning model, remains unreleased and is reportedly held back over performance. The company's visible progress this year has come through the V4 line instead.
What Would Confirm It
Nothing here is official, and the distance between a reported round and a closed one is where these stories usually change. The figures come from people described as familiar with private talks, the range between 80B and 100B yuan is wide enough to suggest the terms are unsettled, and DeepSeek has made no statement. A STAR Market filing would make much of this public in a way no funding announcement does.
What is already established is the shape of the business the money would support. Models are given away under an open licence, hosted access is priced at a fraction of American rates, and the compute supply depends on how much memory China can make.
