DeepSeek is building its own AI chip to cut reliance on Nvidia and Huawei

Chinese AI company DeepSeek is designing its own artificial intelligence chip, according to a Reuters report published on 7 July that cited three people familiar with the plans. The chip would reduce the company’s dependence on Nvidia and Huawei, the two suppliers it has used to train and run its models. DeepSeek has not confirmed the project, so the specifics are reported rather than official. Nvidia shares fell about 1.6% in premarket trading after the news.

The chip is built for inference, not training. Those are the two halves of how a modern AI model works. Training is the expensive, one-off process of teaching a model by feeding it huge amounts of data. Inference is what happens every time you actually use the model: it takes your question and generates an answer. Training needs the most powerful hardware money can buy. Inference can run on cheaper, less power-hungry chips, and it is the workload that grows as more people use an AI product. DeepSeek runs inference at scale for real users, which makes it the sensible place to start.

There is a strategic reason to avoid training silicon too. Nvidia’s biggest lead is in training, both in raw performance and in CUDA, the software layer most AI code is written for. China’s factories also struggle most at the cutting edge, because US export controls block the advanced tools needed to make the best chips. Inference is more forgiving on all of those fronts.

We explained DeepSeek’s rise last year: its R1 model matched top Western systems at a fraction of the cost and was given away as open weights, which knocked a reported trillion dollars off US tech stocks in January 2025. R1 was trained on Nvidia’s H800, a China-specific chip that Washington later banned. Since then DeepSeek has leaned on Huawei. Its V4 model, released in preview in April, was adapted to run on Huawei’s Ascend chips, and Huawei said its processors handled part of the training of a lighter version.

Building an in-house chip means DeepSeek joining a race already under way. Nvidia’s share of China’s AI market has collapsed to roughly zero after Beijing told firms to avoid even the H200 model the Trump administration approved for export. Huawei has filled much of that gap and now holds around half of China’s $50 billion AI chip market. But Alibaba and Baidu are designing their own chips, and DeepSeek’s move would add another domestic rival. We wrote before about how DeepSeek’s open approach was already reshaping who gets to build advanced AI.

The chip also lands in the middle of a global memory squeeze, and DeepSeek is talking to chip-design, foundry and memory companies. Memory matters more for inference than people assume, because generating answers quickly depends on fast access to data. Training chips use a premium type called high-bandwidth memory, or HBM, which AI data centres are buying faster than factories can make it. That has drained supply of ordinary memory and roughly doubled consumer RAM prices over the past year. A 32GB DDR5 kit that cost around $80 to $120 a year ago now sells near $375. Inference chips increasingly use cheaper consumer-grade memory instead of HBM. Nvidia is shifting its own inference GPUs to LPDDR5, the low-power memory in phones and laptops, which puts it in a bidding war with device makers. A Chinese chip would mostly draw on Chinese memory suppliers, so the direct effect on global prices is limited. Even so, inference is becoming another claimant on a strained market.

Whether DeepSeek can pull this off is unsettled. The effort is early, with no foundry, process node or timeline named. One analyst, Richard Windsor of Radio Free Mobile, said DeepSeek has almost no chance of selling chips outside China without access to leading-edge manufacturing, which export controls deny it. Even inside China, Huawei’s best chips reach only about 60% of an Nvidia H100 on inference, and the domestic software tools trail Nvidia’s by years. DeepSeek does have money and momentum. It recently raised about $7.4 billion at a valuation above $50 billion, its first outside funding.

For now this is a statement of intent, not a finished product. The detail that will show whether it is real is which factory agrees to manufacture the chip, because that is where export controls actually bite. Watch for a named foundry and a target process node. Until one appears, DeepSeek is joining China’s home-grown chip push rather than leading it.