What's Actually Happening With DeepSeek's Chip Project

DeepSeek is developing a proprietary AI chip built for inference — the stage where a trained model actually generates the responses you see when you use it. That's according to a Reuters exclusive citing three people familiar with the matter. It's the latest move from one of China's most closely watched AI companies to shrink its reliance on foreign-made hardware, especially as U.S. export controls keep tightening.

A Year in the Making, Still Early Days

The project reportedly kicked off about a year ago, and it's still in its early stages. DeepSeek has been talking to chip-design firms, foundry partners, and memory suppliers, while quietly bringing on semiconductor design engineers. The company hasn't said anything publicly about any of it.

This isn't coming out of nowhere, either. Back in early 2025, DigiTimes Asia reported that DeepSeek had kicked off a major recruitment push for semiconductor design talent — a sign the company was already thinking about building its own processors long before this report surfaced.

An inference-focused chip makes a lot of sense given where the market is headed. As more AI applications go mainstream, a bigger share of computing power gets spent running models rather than training them. That makes custom inference hardware a genuinely valuable asset, not just a nice-to-have side project.

Why DeepSeek Wants to Cut Its Nvidia Dependence

DeepSeek has leaned on Nvidia chips to train its models for years. Earlier this year, a Trump administration official confirmed that DeepSeek trained its V4 model on Nvidia's most advanced Blackwell chip — a move that appears to have crossed the line on U.S. export regulations. In April, the company released a version of V4 optimized to run on Huawei's Ascend chips instead, underscoring China's broader push toward tech self-sufficiency.

The Huawei Workaround Isn't a Long-Term Fix

Here's the thing, though: swapping Nvidia for Huawei doesn't really solve the underlying problem. It just trades one dependency for another. DeepSeek would still be tied to someone else's chip roadmap, timelines, and pricing. A chip the company designs itself hands it real control over performance, cost, and supply — three things that matter a lot more once you're serving a fast-growing base of users instead of just training a model in a lab.

The Money Behind the Ambition

This chip project isn't happening in isolation. It's unfolding alongside DeepSeek's very first round of outside fundraising. In June, both Reuters and Bloomberg reported the company was closing in on roughly 50 billion yuan, or about $7.4 billion. Tencent and CATL are named among the biggest investors in the round.

Tencent and CATL Are Betting Big

  • The round could value DeepSeek somewhere between $52 billion and $59 billion.
  • Founder Liang Wenfeng is reportedly contributing 20 billion yuan of his own personal funds.

Building a chip from scratch is expensive, slow, and the payoff isn't guaranteed for years. But DeepSeek has already built a reputation for hitting frontier-level AI performance without frontier-level spending. This chip bet suggests that ambition isn't staying confined to software.