DeepSeek Moves Into Chip Design, Targeting Inference Silicon to Cut Nvidia and Huawei Dependence

DeepSeek is building its own AI chip. According to a Reuters exclusive published July 7 and citing three people familiar with the matter, the Hangzhou-based lab has been working on a custom inference processor for roughly a year and has held discussions with outside design, foundry, and memory partners without making any of it public.

The target is inference, not training. That's a meaningful distinction. Training is where you build the model; inference is where the model does work for users, generating responses token by token at scale. It's also, increasingly, where the money goes. As AI deployments grow, inference compute becomes the dominant ongoing cost, and a chip tuned specifically for that workload can deliver better performance per watt than a general-purpose GPU pressed into the same job.

The architecture decision here is deliberate. DeepSeek has already demonstrated an unusual ability to squeeze performance out of constrained hardware. Its R1 reasoning model drew attention in early 2025 precisely because it held its own against frontier American models at a fraction of the compute cost. A purpose-built inference chip would let the company close the loop between model architecture and silicon, optimizing both layers together rather than adapting models to whatever hardware is available.

Availability is the problem. As Reuters reported, U.S. export controls bar Chinese companies from Nvidia's most advanced chips, and Washington's restrictions have also cut off access to the high-bandwidth memory that inference chips depend on. DeepSeek has leaned on Huawei's Ascend line for newer workloads, but as the Reuters report makes clear, the company doesn't want a single domestic supplier either.

According to a Reuters report reviewed by Data Center Dynamics, DeepSeek has hired chip-design engineers in recent months without posting public job listings, and is still in early-stage talks with foundry and memory partners. No named manufacturing partner has been announced, no prototype is confirmed, and DeepSeek did not respond to Reuters' request for comment. That combination, a year of quiet groundwork but nothing shippable in sight, is the honest measure of where this project sits.

The geopolitical layer matters here in ways it doesn't for OpenAI or Anthropic. When OpenAI unveiled its Jalapeño inference chip, co-designed with Broadcom, per reporting by Data Center Dynamics, it was primarily a vertical integration play to reduce Nvidia dependence and tighten control over its serving stack. DeepSeek faces the same commercial logic but with an additional hard constraint: it can't simply choose the best foundry. U.S. export rules effectively limit Chinese chip designers to domestic fabs, and no Chinese fab currently runs at the process nodes that competitive inference silicon requires.

That's the missing component the press release doesn't name. Building the chip is one problem. Finding someone to manufacture it at scale, with access to advanced packaging and high-bandwidth memory, is a separate and harder one. Huawei has spent years navigating exactly that wall, and it's still not shipping inference silicon that benchmarks cleanly against H100-class hardware.

The competitive pressure is real regardless. According to Let's Data Science's analysis of the Reuters report, Huawei alone supplies roughly half of China's fifty-billion-dollar domestic AI chip market. A working DeepSeek chip would add a credible alternative inside that market and put pressure on Huawei's pricing and roadmap, independent of anything that happens across the Pacific.

DeepSeek's chip push also coincides with its first-ever external fundraise. Reuters reported in June that the company was set to raise seven billion dollars at a valuation of between fifty-two and fifty-nine billion dollars, reversing a long-held policy against outside investment. Building silicon is expensive before you've shipped a single unit, and that capital context is worth keeping in mind when reading the timing of this disclosure.

Sources cited:
- Reuters (via U.S. News & World Report) (https://www.usnews.com/news/top-news/articles/2026-07-07/exclusive-chinas-deepseek-developing-its-own-ai-chip-sources-say)
- Data Center Dynamics (https://www.datacenterdynamics.com/en/news/deepseek-to-develop-its-own-custom-ai-chip-report/)
- SiliconANGLE (https://siliconangle.com/2026/07/07/report-chinas-deepseek-follows-openai-developing-custom-inference-chips/)
- Let's Data Science (https://letsdatascience.com/news/deepseek-develops-proprietary-ai-inference-chip-435f0f9e)
- Bloomberg (https://www.bloomberg.com/news/articles/2026-07-07/chinese-ai-startup-deepseek-developing-own-ai-chip-reuters-says)