DeepSeek Is Building Its Own Inference Chip to Escape Both Nvidia and Huawei
Reuters reports China's DeepSeek has spent a year exploring custom AI accelerators and is now hiring chip designers and talking to foundries. The goal: silicon independence from every supplier it currently depends on.
DeepSeek is designing its own AI inference chip, Reuters reported this week, citing sources familiar with the project. The Chinese lab behind the V-series models has been exploring custom accelerators for about a year and is now actively recruiting experienced chip designers while holding talks with chip design houses, foundries, and memory suppliers.
The target is inference — the serving side of AI, where trained models answer queries — not training. That’s the pragmatic choice. Inference is where DeepSeek’s costs live: its models are among the most heavily used in China, and its low API pricing (the strategy that made it famous) only works if serving costs keep falling. It’s also the more achievable silicon problem. Inference chips tolerate narrower designs and older process nodes better than training chips do.
The strategic logic cuts in two directions. Everyone expected Chinese labs to design away from Nvidia — export controls made that mandatory. What’s notable is that DeepSeek also wants independence from Huawei, whose Ascend accelerators are Beijing’s officially blessed domestic alternative. DeepSeek reportedly struggled to train on Ascend hardware last year, falling back to Nvidia clusters for training while using Huawei silicon for some inference. Building its own chip is a bet that neither supplier will serve its roadmap.
It’s the same playbook running in the US. OpenAI is co-designing accelerators with Broadcom, Anthropic leans on custom capacity deals, Meta just pushed its Iris inference chip into production. Frontier labs everywhere have concluded that renting someone else’s silicon at margin is incompatible with serving models at scale. DeepSeek following suit was a matter of when.
The obstacles are real. A competitive accelerator takes years and billions to ship. US restrictions bar Chinese designers from the most advanced overseas foundries, and separate curbs have cut China’s access to high-bandwidth memory — the component that matters most for inference throughput. SMIC’s 7nm-class capacity exists but is oversubscribed. The project is described as early-stage, and early-stage chip projects die quietly all the time.
But dismissing DeepSeek on execution has been a losing trade since January 2025. This is the company that trained a frontier-class model for a fraction of Western budgets by squeezing efficiency out of constrained hardware. A team with that profile designing silicon for its own exact workload is precisely the kind of project that ships something unglamorous, cheap, and good enough.
If it works, the export-control regime loses another lever. Chips you design at home can’t be embargoed.