DeepSeek to build its own AI chips as US export controls tighten

DeepSeek announced it will develop a proprietary AI accelerator to replace Nvidia GPUs that are now unavailable under U.S. export controls, a move that underscores a rapid shift in China’s AI hardware strategy toward custom ASICs [1].

From GPUs to ASICs

U.S. restrictions on the export of high‑end Nvidia GPUs have forced Chinese AI firms to abandon the “Nvidia clone” model and accelerate the adoption of application‑specific integrated circuits (ASICs). ASICs trade the flexibility of general‑purpose GPUs for higher efficiency on fixed AI workloads, a trade‑off that becomes attractive when the most powerful GPUs cannot be sourced [1]. DeepSeek’s effort to rewrite its core software for Huawei’s CANN framework illustrates how firms are re‑engineering their stacks to run on domestic silicon rather than Nvidia’s CUDA ecosystem [1].

Market momentum behind custom silicon

Huawei is projected to capture 62% of China’s domestic AI accelerator market by 2026, with Cambricon at 14% and Baidu and Alibaba each around 5% [1]. The performance gap is already narrowing: Huawei’s Ascend 950 and Cambricon’s Siyuan 690 reportedly outpace Nvidia’s H20 (the most powerful chip Nvidia is still allowed to sell to China) by 50‑150% in tokens‑per‑second benchmarks [1]. This competitive edge, combined with the loss of Nvidia’s GPUs, is driving firms like DeepSeek to invest in home‑grown chips rather than wait for foreign supply.

Software stack challenges

Switching hardware is only half the battle; developers must also migrate away from CUDA, which dominates AI software worldwide. Huawei’s CANN and Moore Threads’ MUSA are emerging alternatives, but the ecosystem fragmentation means Chinese developers may need to support multiple proprietary stacks [1]. Analysts note that as AI models grow more complex, the line between ASIC efficiency and GPU flexibility is blurring, suggesting future architectures could blend both approaches [1].

What to watch

• DeepSeek chip rollout – the first prototype timeline and any announced performance claims.
• Huawei’s Ascend 950 shipments – volume and adoption rates as a benchmark for domestic ASIC success.
• U.S. policy updates – any further tightening of export controls that could accelerate the shift to custom silicon.

The push to build indigenous AI chips marks a structural redesign of China’s AI compute landscape, raising the question of whether a fragmented, ASIC‑centric ecosystem can keep pace with the rapid innovation driven by Nvidia’s CUDA‑based platform in the West.