DeepSeek develops proprietary AI inference chip

DeepSeek is developing its own AI inference chip to reduce dependence on Nvidia and Huawei, according to a Reuters report citing three sources. The Chinese lab has been in talks with chip-design and foundry partners for about a year and has hired engineers privately. A working chip could reshape China's $50 billion AI hardware market and pressure Nvidia.

DeepSeek develops proprietary AI inference chip Reuters reported on July 7, 2026 that DeepSeek is quietly building its own AI inference chip, citing three people familiar with the effort. The Chinese lab has reportedly been in talks with chip-design, foundry, and memory partners for about a year and has hired chip-design engineers privately, aiming to cut its dependence on Nvidia and Huawei silicon. DeepSeek did not comment on the report. Because Huawei alone supplies roughly half of China's $50 billion domestic AI-chip market, a working DeepSeek chip could reshape competitive dynamics in Chinese AI hardware and add to pressure already building on Nvidia, whose shares slipped after the report. A proprietary inference chip would let DeepSeek fold hardware into the same optimization loop as its model architecture and serving stack, a step so far attempted by only a handful of labs, including OpenAI and Anthropic. That matters beyond DeepSeek: it signals how far down the stack AI companies now feel they need to operate to control cost and supply. What happened Reuters reported on July 7, 2026, citing three people familiar with the matter, that DeepSeek is developing an AI chip built for inference rather than training. The effort is roughly a year old and still early: DeepSeek has been discussing chip design, manufacturing, and memory with outside partners and has been hiring chip-design engineers without public job postings. DeepSeek did not respond to Reuters' request for comment. Industry context DeepSeek has trained and served its models on a mix of Nvidia and Huawei chips; Huawei alone supplies roughly half of China's $50 billion domestic AI chip market, partly because U.S. export rules bar Chinese firms from Nvidia's most advanced GPUs. DeepSeek used Huawei's Ascend chips for parts of its V4 and V4-Flash models. Rivals Alibaba and Baidu are also building their own AI chips and gaining share, and Reuters reported that Nvidia shares fell after the story published, reflecting investor concern about further competition in China's AI-chip market. For practitioners Inference and training workloads have different hardware profiles: inference needs to sustain high utilization, tight latency, and low cost per token under live traffic rather than raw training throughput. A chip tuned to DeepSeek's own model architecture and serving software would let the company manage capacity planning and serving cost more directly, following the same logic behind OpenAI's Jalapeno inference chip built with Broadcom, and Anthropic's own reported chip exploration. What to watch The project is still at the discussion-and-hiring stage, with no named foundry partner, prototype, or benchmark disclosed. Watch for confirmed manufacturing partners, compiler and runtime support, and whether future DeepSeek model releases are tuned for proprietary or Huawei-compatible inference hardware; competitive AI chip programs typically take years and heavy capital, and U.S. curbs on advanced foundry access and high-bandwidth memory remain a manufacturing constraint for any China-based effort. Key Points - 1Reuters' sources say DeepSeek has spent about a year developing its own inference chip with outside chip-design and foundry partners. - 2DeepSeek currently relies on Nvidia and Huawei silicon; Huawei alone holds roughly half of China's $50 billion AI chip market. - 3The project remains early-stage with no named partner or prototype, but adds to pressure on Nvidia amid China's chip push. Scoring Rationale Same underlying Reuters exclusive as the sibling DeepSeek chip event: a major Chinese AI lab reportedly moving into custom inference silicon amid export controls, now corroborated by Bloomberg and cross-verified full wire text. Held at notable rather than major because the effort is early-stage, unconfirmed by DeepSeek, and has no product or named foundry partner yet. Sources Public references used for this report. Practice interview problems based on real data 1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with. Try 250 free problems /problems