Reuters reported on Monday that DeepSeek is developing its own artificial intelligence chip, designed for inference rather than training, in a project that has been underway for approximately a year. The Hangzhou-based company has been meeting with chip-design firms, semiconductor foundries, and memory suppliers. It has been hiring chip-design engineers privately, without posting openings on public platforms. The goal, according to three people familiar with the effort, is to reduce DeepSeek's dependence on both Nvidia and Huawei.
The second name in that list is the one that matters. Nvidia's share of the Chinese AI accelerator market has already collapsed to zero percent, a fact that CEO Jensen Huang has confirmed publicly. U.S. export controls accomplished what no competitor could. The company DeepSeek is actually trying to escape is Huawei, which now projects it will capture sixty percent of China's domestic AI chip market by the end of 2026. Huawei expects $12 billion in AI chip revenue this year, up from $7.5 billion in 2025, driven by the Ascend 950PR processor, which began mass production in March. ByteDance, Tencent, and Alibaba have all placed orders. The domestic market is supply-constrained. Huawei is the bottleneck.
DeepSeek became the most celebrated AI company in the world by proving that software efficiency could overcome hardware constraints. Its models trained on a fraction of the compute that American labs required. V3 ran on Nvidia's export-compliant H800, a chip deliberately weakened to satisfy U.S. restrictions. V4 now runs an estimated twenty-three percent of all open-weight production AI applications worldwide, more than Llama 4 Scout or Mistral Large 3. The company closed its first external funding round in June, raising approximately $7.4 billion at a valuation exceeding $50 billion.
Founder Liang Wenfeng personally committed roughly twenty billion yuan, the round's controlling share. Tencent invested about ten billion yuan and battery maker CATL about five billion. China's National Artificial Intelligence Industry Investment Fund received direct equity with voting rights and no lockup period. It was the only investor granted governance rights. The state bought a seat at the table of the company the state considers its AI champion.
The chip DeepSeek is designing is for inference: the workload where a trained model generates responses to user queries. This is not the workload that made DeepSeek famous. Training is what made DeepSeek famous. But inference is what costs money every day. Every query to every user runs through inference hardware. The training run is a one-time capital expenditure. The inference fleet is the operating expense that scales with every customer added and every token served.
DeepSeek monetizes through API access at prices that make the inference cost existential. DeepSeek V4 Flash charges fourteen cents per million input tokens. OpenAI's GPT-5.4 charges $2.50. The gap is eighteen to one. At that ratio, the hardware cost of inference is not a line item. It is the business model. A company that prices its intelligence at one-eighteenth of its closest competitor needs inference to be as cheap as physically possible. The custom chip is not an ambition. It is arithmetic.
DeepSeek is not the first AI lab to reach this conclusion. OpenAI unveiled Jalapeño in June, a custom inference chip built with Broadcom and fabricated by TSMC, taped out in nine months. Google has been running its own Tensor Processing Units for over a decade. Amazon designed Trainium and Inferentia. Microsoft built Maia. Meta developed MTIA. The convergence is total: every major AI company now either has a custom inference chip or is building one. The inference layer is becoming as proprietary as the model weights.
But DeepSeek faces constraints that no other lab faces. OpenAI called Broadcom and TSMC. Google fabricates at TSMC and Samsung. Amazon uses TSMC. DeepSeek cannot access TSMC. U.S. export controls prohibit Chinese AI companies from using the world's most advanced foundries. Separate restrictions have cut China's access to high-bandwidth memory, a component critical to inference chip performance. DeepSeek's chip must be fabricated domestically, almost certainly by SMIC, using process nodes at least two generations behind the leading edge, and must function without the memory architecture that every other lab's custom silicon relies on.
The irony is architectural. The export controls that created DeepSeek's efficiency advantage, by forcing the company to do more with less capable hardware, now constrain its attempt to build more capable hardware of its own. The software optimization was an adaptation to scarcity. The chip project is an attempt to own the scarcity. Both are responses to the same set of export controls, but they point in opposite directions. One accepted the constraint. The other is trying to remove it.
There is a deeper strategic tension. Huawei is not just a chip supplier to DeepSeek. Huawei is a national champion, backed by the same state apparatus that invested in DeepSeek's funding round through the National AI Industry Investment Fund. DeepSeek's decision to design around Huawei is a statement that vertical integration matters more than national solidarity. The company that China celebrates as proof that its AI can compete with America's does not trust China's national champion chipmaker with its most critical workload.
Nvidia shares fell on the report, declining in premarket trading. Memory chipmakers also dropped, with Micron, Western Digital, and SanDisk each falling approximately seven percent, though those declines were driven primarily by Samsung's record-earnings selloff earlier in the day rather than the DeepSeek disclosure specifically. The market processed the news as another chip demand signal turning negative. It may be something more specific: the company that proved Nvidia's best chips were unnecessary is now trying to prove that Nvidia's chips are unnecessary at the hardware level too.
The project remains at an early stage. Designing a competitive AI accelerator typically requires years and billions of dollars, even with access to the world's best foundries and supply chains. Without that access, the timeline extends and the difficulty compounds. DeepSeek's $7.4 billion funding round provides the capital. Its $50 billion valuation prices in the software efficiency. It does not yet price in the hardware ambition.
Every AI lab on both sides of the Pacific has now arrived at the same conclusion: the model is not enough. The model needs its own silicon. The company that proved you could do frontier AI with inferior chips has inferred that inferior chips are the ceiling. The efficiency that made DeepSeek famous was never the destination. It was the adaptation that funded the journey to the destination. And the destination, for every lab that has done the math, is the chip.