DeepSeek, the Hangzhou-based artificial intelligence startup, is designing its own chip, three people familiar with the matter say.
The move would reduce its dependence on Nvidia, the dominant American chipmaker, and on Huawei, the Chinese technology group whose Ascend processors it now leans on.
Understanding why matters more than the announcement itself, because the choice reveals how China's AI industry is adapting to being cut off from the best hardware.
Why inference, not training
The first clue is what kind of chip DeepSeek is building.
It is targeting inference, the stage where a trained model answers user queries, rather than training, the far more demanding process of building a model in the first place.
That distinction is the heart of the strategy.
Training is where Nvidia's lead is widest, thanks to its software, its high-speed chip-to-chip connections and its access to the most advanced manufacturing.
It is also where United States export controls bite hardest, because building competitive training chips requires cutting-edge lithography that Chinese factories cannot yet match.
Inference is a gentler target.
It is more forgiving on the manufacturing process, meaning a chip built on an older, more widely available production method can still be competitive.
It is also acutely sensitive to serving cost, the price of answering each individual query, which is exactly the metric a specialised chip can improve.
Every interaction with an AI assistant requires inference, so demand for this kind of computing is growing quickly and rewards whoever can run it cheaply.
A pricing war extended into silicon
Controlling its own inference hardware would extend a strategy DeepSeek has already pursued through pricing.
In May 2026 it cut the price of its V4-Pro model by 75%, dropping the top rate to under $0.85 per million tokens from $3.30.
A purpose-built chip that lowers the cost of running its models would let it push those prices lower still.
Where China holds an edge
This is where Chinese firms may hold an inherent advantage.
DeepSeek designs both the model and, now, the chip that runs it, allowing the two to be tuned together in a way that general-purpose hardware cannot easily match.
The company has built its reputation on squeezing strong performance from constrained resources, a discipline forced on it by limited access to top chips.
Around a recent model release it said it had used a data format well suited to home-grown chips soon to be released, hinting at exactly this co-design approach.
The scale of domestic demand reinforces the case.
Chinese firms plan to shift 46% of their AI-accelerator budgets to domestic suppliers within a year, according to industry estimates.
Nvidia's position in China has collapsed under export restrictions, leaving a large captive market for whoever can supply credible local alternatives.
Beijing has been pressing its technology champions to build those alternatives, adding political tailwinds to the commercial logic.
When DeepSeek optimised its V4 model for Huawei's Ascend chips in April 2026, it set off a buying rush, with ByteDance, Tencent and Alibaba all approaching Huawei for orders.
That episode showed both the appetite for domestic hardware and DeepSeek's power to shape which chips the market adopts.
The hard limits
The obstacles, however, are substantial.
Designing a competitive inference chip demands years of engineering and large amounts of capital, neither of which guarantees success.
US curbs restrict Chinese access to leading-edge foundries, the factories that fabricate chips, and to high-bandwidth memory, the fast memory that AI accelerators need to perform well.
Those constraints cap how advanced DeepSeek's chip can be and how many it can produce.
Some analysts are sceptical it can sell silicon beyond China without access to the most advanced manufacturing, limiting the venture to the home market.
The chip push coincides with DeepSeek's first planned outside funding, a maiden round that Reuters reported would raise $7 billion at a valuation between $52 billion and $59 billion.
That capital would help fund a long and uncertain bet, one that reflects both the pressure of export controls and the particular strengths China is trying to build around them.
