DeepSeek to Build Data-Center Chips as US Export Controls Bite

DeepSeek plans to enter the silicon business and design data-center chips for inference, not training. Reuters, citing three people familiar with the matter, says DeepSeek has been working on the effort for about a year, and has been meeting potential partners and hiring engineers.

What is DeepSeek planning?

DeepSeek is aiming to build its own data-center inference chips to reduce dependence on existing suppliers. The company has focused the effort on inference rather than training hardware, and Reuters says the initiative has been under way for about a year with meetings with potential partners and active hiring of engineers.

The stated goal, as presented in the reporting, is likely to lower reliance on both Huawei and Nvidia in China. The move targets the inference segment of data-center compute, a different technical profile than training chips and one that aligns with companies deploying models in production rather than building the largest models.

How do US export controls and incumbent vendors shape the move?

A United States export ban has limited Nvidia from achieving a presence in China comparable to its share in North America and Europe. Nvidia remains the chipmaker for most AI companies in North America and Europe, while Huawei controls about half of the data-center chip market in China, creating a market split that DeepSeeks plan seeks to address.

The export controls are an explicit driver for urgency, the reporting says. Chinese technology firms are not alone in aiming for in-house silicon. OpenAI and Broadcom jointly announced Jalapeo, OpenAIs first chip designed for inference at scale, just a couple of weeks ago. Anthropic has been exploring custom chip design as well, though the reporting notes there have not been publicly visible milestones from Anthropic yet.

OpenAIs public rationale includes reducing reliance on Nvidia and gaining tighter control over its product stack, akin to Apples vertical integration. The reporting also points out a strategic advantage to working at the silicon and data-center levels in a market where data-center access is likely to remain constrained and multiple companies will compete for compute as they scale their models and services.

Why it matters

If DeepSeek succeeds, it would be another sign that AI firms are moving from purely software-first competition toward hardware and infrastructure control. That could change supplier dynamics inside China, where Huawei currently holds about half of the data-center chip market, and reduce the leverage of external vendors blocked by export rules. The shift also echoes parallel moves by US-based firms like OpenAI, suggesting both Chinese and American companies see benefits to designing inference hardware in-house.

What to watch

Look for formal announcements of partnerships, public hiring drives, or a chip reveal from DeepSeek that confirm the Reuters reporting. Also watch how Jalapeo progresses in deployments and whether Anthropic publishes any milestones on its chip work; those developments would clarify whether firms across markets can ease reliance on incumbent chip suppliers.