Breaking up with your only chip supplier is hard to do, especially when there are two of them and both hold most of the cards. But according to Reuters, DeepSeek is quietly trying to write its own exit plan — one silicon wafer at a time.

A Chip Project Nobody Confirmed Yet

Reuters reports, citing three people familiar with the effort, that Hangzhou-based DeepSeek has spent roughly the last year working on a custom AI inference chip. The project is still early-stage — DeepSeek has been quietly hiring chip-design engineers through private channels and holding talks with design, foundry, and memory firms, but no tape-out date or production timeline exists. Officially, DeepSeek has said nothing.

Crucially, this isn't a training chip aimed at building the next model from scratch. It's an inference chip, meant to handle the unglamorous but expensive job of actually running trained models and spitting out answers to millions of users every day — the part of the AI bill that never stops arriving.

Two Suppliers Is One Too Few

DeepSeek currently leans on Nvidia and Huawei for its AI processors, and after U.S. export restrictions squeezed access to Nvidia's H800 chips, it leaned harder into Huawei's Ascend line. Depending on two suppliers, one of whom is subject to geopolitical whiplash and the other of whom is your only domestic backup, is not exactly a comfortable place to run a business from.

This is the same playbook OpenAI and Anthropic are already running with their own custom silicon bets — because once your model is popular, inference costs stop being a rounding error and start being the whole spreadsheet. DeepSeek's recent $7.4 billion funding round gives it the runway to actually try, even if "early stage" here likely means years, not months, before anything ships.

Everyone's building their own chips now. At this rate, the real AI arms race might end up being fought in fabs, not fine-tuning.

Source: SiliconANGLE