DeepSeek Moves Into Custom Silicon With Inference Chip Push

DeepSeek is building its own AI inference chip, according to a Reuters exclusive published July 7 that cited three people familiar with the effort. The project is still early-stage: no named foundry partner, no prototype on record. But the architecture decision is already significant, because inference chips are where the next cost war in AI actually gets fought.

The distinction matters. Training chips get most of the press because they're where models are built. Inference is where models earn their keep, processing every user query after the model is already deployed. As Reuters reported, DeepSeek's chip is specifically targeted at that stage, the point where a trained model generates responses for users, rather than at building new models from scratch. That's a deliberate choice. Inference is the fastest-growing segment of AI compute demand, and specialized silicon can run it cheaper and cooler than general-purpose GPUs.

DeepSeek's constraint set is not optional. U.S. export controls block Chinese companies from buying Nvidia's most advanced chips, and as Reuters and SiliconANGLE both noted, that has made Huawei the dominant chip supplier inside China. DeepSeek has leaned on both Nvidia and Huawei silicon to train and serve its models, including Huawei's Ascend accelerators for its V4 series. A working in-house chip would let the lab break out of that two-vendor dependency entirely.

There's also a strategic dimension that goes beyond supply security. As SiliconANGLE reported, DeepSeek would be following a playbook already in motion at OpenAI and Anthropic. OpenAI recently unveiled a custom inference chip called Jalapeño, co-designed with Broadcom. Anthropic has been weighing its own silicon. The pattern is clear: labs that want to control their unit economics at scale need to own more of the stack. A proprietary inference chip lets a lab fold hardware optimization directly into the same loop as model architecture and serving software, something only a handful of organizations have attempted.

According to Reuters, DeepSeek has spent roughly a year on the effort, quietly hiring chip-design engineers without public job postings and holding discussions with external chip-design, foundry, and memory companies. DeepSeek did not respond to Reuters' request for comment.

The memory piece is where the real technical difficulty sits. U.S. export curbs have also cut China's access to high-bandwidth memory, a component that's critical to competitive inference chips. That's not a problem DeepSeek can solve just by hiring good engineers. It'll need manufacturing partners who can supply the memory stack, and those conversations, per Reuters, are ongoing but unresolved.

Nvidia shares fell after the report published, a reflex reaction investors have now practiced a few times when DeepSeek makes news. The more considered read: a chip that can't leave China because of export rules, and that has no named production partner yet, isn't a near-term Nvidia revenue problem. It's a longer-arc signal about where the Chinese AI stack is heading. Huawei, which holds roughly half of China's domestic AI chip market, has more direct exposure if DeepSeek succeeds.

This story also lands as DeepSeek is reversing its longstanding policy against outside capital. Reuters reported in June that the company was preparing a maiden funding round valuing it at between $52 billion and $59 billion. Hardware development at this level costs real money. The timing of the chip disclosure and the funding round aren't coincidental.

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)
- SiliconANGLE (https://siliconangle.com/2026/07/07/report-chinas-deepseek-follows-openai-developing-custom-inference-chips/)
- Bloomberg (https://www.bloomberg.com/news/articles/2026-07-07/chinese-ai-startup-deepseek-developing-own-ai-chip-reuters-says)
- Let's Data Science (https://letsdatascience.com/news/deepseek-develops-proprietary-ai-inference-chip-435f0f9e)