DeepSeek builds its own AI chip to cut Nvidia reliance

TL;DR:

  • DeepSeek is developing its own AI inference chip, three sources tell Reuters.
  • The move would cut its reliance on Nvidia and Huawei silicon amid US export controls.
  • Success is far from assured, given manufacturing curbs and limited access to advanced foundries.

DeepSeek, the Chinese startup whose efficient models rattled Silicon Valley last year, is now designing its own AI chip. According to three people familiar with the effort, the chip targets inference, the stage where a trained model answers queries, rather than training, and would reduce the firm’s dependence on Nvidia and Huawei processors. Nvidia shares slipped about 1.6% in pre-market trading on the report.

Following a global custom-silicon trend

The push would mark a strategic shift for a company long focused on model breakthroughs over commercialisation, and it places DeepSeek alongside a broader move by AI developers to control their own hardware. OpenAI recently unveiled its first custom inference chip, Jalapeno, with Broadcom, and Anthropic has been in talks with Samsung over bespoke silicon. For DeepSeek the logic is sharper still: US export controls bar it from Nvidia’s most advanced chips, and Beijing has pressed domestic champions to build alternatives. The company has leaned increasingly on Huawei, releasing a V4 model adapted for Ascend chips, but an in-house design would target the fastest-growing slice of AI compute demand as workloads shift from training to running models.

Looking forward

The obstacles are formidable. Designing a competitive chip takes years and heavy capital, while US curbs cut Chinese designers off from the most advanced foreign foundries and from high-bandwidth memory critical to inference. The effort coincides with DeepSeek’s first embrace of outside funding, a reported $7bn round valuing it between $52bn and $59bn, reversing years of self-reliance. For UK firms, the story is another marker of a fracturing AI hardware market, one where Chinese models are already gaining ground even as the chips beneath them splinter along geopolitical lines.

Share this article