Analysis · Reuters / US News ·

DeepSeek is quietly building its own AI inference chip to cut reliance on Nvidia and Huawei

Chinese AI lab DeepSeek is developing its own custom chip designed for inference rather than training, according to Reuters sources, in an effort to reduce dependence on both Nvidia and Huawei processors. The roughly year-old project remains early-stage, with DeepSeek in talks with chip-design, foundry, and memory partners while facing US and Dutch export controls that block access to the most advanced overseas fabs.

Based on reporting by Reuters / US News — analysis by dalili

The chip targets inference — the stage where a trained model generates responses for users — rather than training new models, positioning DeepSeek alongside OpenAI, which unveiled its first custom inference chip Jalapeno with Broadcom last month, and Anthropic, which has been in talks with Samsung over bespoke silicon. For DeepSeek specifically, the logic carries extra weight: US export controls bar it from Nvidia's most advanced chips, and Beijing has been pressing its technology champions to build domestic alternatives, while Huawei alone supplies roughly half of China's $50 billion domestic AI-chip market.

The obstacles are formidable. US and Dutch sanctions block Chinese chip designers from using cutting-edge overseas fabs like TSMC or importing advanced ASML lithography equipment, meaning DeepSeek would almost certainly have to rely on domestic foundries like SMIC, operating on mature process nodes with clear performance ceilings compared to Western alternatives. Separate trade restrictions have also heavily constrained China's access to high-bandwidth memory, a core component for feeding data into high-speed inference processors — DeepSeek's ultimate success will depend on out-engineering these hardware limits using the same architectural efficiency that made its models notable in the first place.

The push coincides with DeepSeek's first embrace of outside funding, reportedly targeting roughly $7 billion at a valuation range of $52-59 billion, and with a broader recalibration inside China: executives at Chinese companies plan to route 46% of their AI-accelerator budgets to domestic products over the next year, up from about 30% today — a shift that would give any competitive homegrown inference chip a ready market regardless of whether it matches Nvidia on raw performance.

Key takeaways

  • DeepSeek is developing a custom inference chip, joining OpenAI and Anthropic in building bespoke AI silicon
  • US/Dutch export controls will likely force reliance on domestic foundries like SMIC, with clear performance ceilings
  • The push coincides with DeepSeek's first external funding round, targeting ~$7B at a $52-59B valuation

Why it matters

When the lab whose efficiency breakthroughs already rattled Nvidia's valuation starts designing its own silicon, it signals how far down the stack AI companies now feel they must operate to control cost and supply — not just for China, but as a preview of where every major AI lab is heading.