Semiconductor research firm SemiAnalysis estimates that over 1.5 gigawatts of restricted Nvidia AI accelerators — potentially up to 1.5 million GPUs — have been diverted to China despite U.S. export controls, excluding compliant models like the H20 and H800. The firm's Accelerator Model suggests this volume is comparable to or exceeds xAI's Colossus One deployment of 220,000–230,000 Hopper GPUs and approaches the 1.2 million Blackwell GPUs targeted for Colossus 2. Power calculations imply 1.5 million B200 SXM units (1,000W) or 2.14 million H200 SXM units (700W), with mixed deployments falling between. The U.S. Department of Commerce reportedly concurs with the scale assessment. Two major smuggling cases support the findings: a March 2026 indictment of Supermicro co-founder Wally Liaw, Steven Chang, and Willy Sun for diverting $510 million of a $2.5 billion Southeast Asia server sale to Alibaba in China using label substitution; and a Taiwanese case involving nine individuals from Nvidia, Albatron, Chief Telecom, Flying Tiger, and Supermicro who allegedly shipped 130 B300-based servers (74 reaching China) via shell companies and forged documents declaring Japan as the destination. Separately, reports allege Jensen Huang's former partner Alice Huang facilitated $7.5 billion (approximately 240,000 GPUs, mostly H100/H200) in smuggling. SemiAnalysis speculates the violations may explain Huang's departure timing, coinciding with the House passing the Remote Access Security Act (RASA) on January 12, which Nvidia and Oracle opposed. Analysts argue Huang's removal accelerates stricter cloud-access restrictions but warn total bans could drive Chinese adoption of domestic accelerators, undermine open-source contributions like DeepEP and DeepGEMM, and erode U.S. monitoring capabilities and revenue.
Read original
hackernews