DeepSeek is designing its own inference chip to escape Nvidia and Huawei

Reuters dropped an exclusive this morning: DeepSeek has been quietly building its own AI chip for about a year. Three sources confirmed the project, which targets inference workloads, not training. The goal is to cut the company’s dependence on Nvidia and Huawei silicon.

This is the hardware equivalent of what DeepSeek already did in software. The startup shocked the industry last year with models that matched Western performance at a fraction of the compute cost. Now it wants the same independence at the silicon level.

The chip is designed for inference, the phase where a trained model generates responses. That matters because inference is where the bulk of real-world AI compute goes. Training gets the headlines. Serving models to millions of users is where the money actually burns. A specialized inference chip can cost less and draw less power than a general-purpose GPU, which is exactly why Nvidia’s data center margins have looked like a monopoly tax.

Nvidia shares slipped 1.6% in premarket trading after the Reuters report.

The project faces real constraints. US export controls cut Chinese companies off from the most advanced foundries and high-bandwidth memory. DeepSeek has been talking to chip design firms, foundries, and memory suppliers. It has also been hiring chip engineers through private channels, no public job postings. The path from design to working silicon under sanctions is brutal. SMIC, China’s leading foundry, is generations behind TSMC on process node. Good luck fabricating competitive inference chips on a constrained node with restricted access to HBM.

But DeepSeek does not need to beat Nvidia on performance. It needs to be good enough to serve its own models at scale without paying Nvidia’s prices. The same model efficiency that made DeepSeek famous in the first place works in its favor here. If your models run lean, your silicon can too.

The timing lines up with a bigger shift. DeepSeek is raising outside capital for the first time, reportedly seeking $7 billion at a valuation between $52 and $59 billion. Chip design burns cash. The funding round and the silicon project are the same story: DeepSeek is building the full stack.

OpenAI and Anthropic are also designing their own chips. Google has TPU. Amazon has Trainium. Meta has MTIA. The pattern is clear. Every major AI lab that can afford it is trying to cut Nvidia out of the loop. DeepSeek is just doing it under harder conditions than anyone else.

The geopolitical angle writes itself. DeepSeek became a national champion in China by proving that export controls could not stop competitive AI models. A domestic chip would extend that argument from software into hardware. If it works, it undercuts the entire premise of the semiconductor embargo: that controlling the chips controls the AI. DeepSeek’s models already challenged that. A DeepSeek chip would challenge it again.

If it fails, DeepSeek keeps buying from Huawei and whatever Nvidia chips slip through the export control gaps. Either way, the company is spending serious money to find out.