DeepSeek begins developing custom AI inference chips to reduce dual dependency on NVIDIA and Huawei.
The AMW Read
Updates DeepSeek's §4.2 case study with a strategic move from model-maker to chip-designer, introducing a new variable in the US-China AI hardware contest, with cross-segment implications for inference economics and the China AI chip market.
DeepSeek begins developing custom AI inference chips to reduce dual dependency on NVIDIA and Huawei.
DeepSeek has initiated an in-house AI chip development program focused on inference processors, according to a Reuters report citing multiple sources. The Chinese AI firm has been working on custom silicon for roughly a year, and is in discussions with chip design firms, foundries, and memory suppliers. The chips are designed for inference workloads — the compute stage where trained models generate responses to user queries — rather than for training new models. DeepSeek has been quietly expanding its chip design engineering team through private industry networks rather than public job postings.
Why it matters: DeepSeek's move to custom silicon updates two structural forces in the AI substrate — the hyperscaler-distribution pattern and the capital-compression arc playing out in China's AI ecosystem. After US export controls cut off access to NVIDIA's H800, DeepSeek shifted to Huawei's Ascend chips, creating a new dependency on a vendor that is simultaneously a cloud competitor and whose hardware roadmap and capacity constraints caused a reported delay in DeepSeek's next-generation model in 2025. This mirrors the broader vertical-integration play seen at OpenAI (Jalapeño inference chip with Broadcom), Google (TPU), and Amazon (Trainium/Inferentia), but with a geopolitical overlay unique to Chinese firms: the foundry and HBM-access constraints imposed by US restrictions on advanced semiconductor manufacturing and high-bandwidth memory.
Ground truth: The choice to start with inference rather than training silicon is strategically rational — inference is the recurring, user-count-driven cost center for DeepSeek's rapidly scaling API service, and it requires less cutting-edge process technology than training chips. Success is far from guaranteed: competitive AI chip design typically takes years, and even after tape-out, access to advanced foundry nodes and HBM remains constrained by US policy. The program also likely explains DeepSeek's reported shift to external fundraising — its first-ever outside capital round at a $52B-$59B valuation — since chip development is capital-intensive. The more immediate competitive threat from this news is not to NVIDIA but to Huawei, which currently holds roughly half the ~$50B China AI chip market that NVIDIA vacated.
#DeepSeek #AIChips #Inference #HuaweiAscend #USChina #ExportControls #VerticalIntegration
