The United States' export controls on advanced AI chips have yielded mixed results in constraining China's AI progress, despite recent enforcement actions. Taiwan's Keelung District Prosecutors’ Office executed search warrants on May 21, 2026, detaining three individuals for allegedly smuggling Super Micro servers with restricted NVIDIA chips to China, following a US Department of Justice indictment charging Super Micro's co-founder, Wally Liaw, with diverting approximately US$2.5 billion worth of NVIDIA-powered servers.
Concurrently, China's DeepSeek V4, a 1.6-trillion-parameter AI model, was released, designed for domestic inference hardware like Huawei Ascend 950 and Cambricon silicon. DeepSeek V4 outperformed GPT-5.4 on competitive coding benchmarks, narrowing the US-China AI model performance gap to 2.7% by March 2026. While the US maintains a hardware lead, China's AI chip self-sufficiency reached 41% by 2025, with Beijing blocking NVIDIA H200 imports. However, China faces challenges in mass-producing homegrown chips (SMIC's 5nm yield is ~20%), cluster efficiency, and interconnect bandwidth, maintaining a significant aggregate compute capacity asymmetry with the US.
No comments:
Post a Comment