Compute bottlenecks tightening
Approval times for U.S. exports of Nvidia and AMD AI chips to China are slowing after staffing losses at the Bureau of Industry and Security, creating a bureaucratic choke point for hardware access. (startupnews.fyi). Recent summaries also note hyperscalers control roughly two‑thirds of global AI compute, concentrating capacity in a few providers and changing where frontier work can realistically run. (x.com)
A handful of U.S. officials now sit between Chinese buyers and the Nvidia and Advanced Micro Devices chips they need for artificial intelligence. (ttnews.com) The office is the Bureau of Industry and Security, the Commerce Department unit that reviews exports of dual-use technology, meaning products with civilian and military uses. A June 2025 Government Accountability Office report said the bureau had expanded from 403 funded positions in fiscal year 2013 to 585 in fiscal year 2024, but still lacked a long-term workforce plan. (gao.gov) Bloomberg reported on April 10, 2026 that approvals for chipmakers and other exporters have stretched into months after nearly 20% turnover among rulemaking and licensing staff. The same report said processed licenses across industries fell by about 25% last year, leaving billions of dollars in export backlogs. (bloomberg.com) The current bottleneck sits on top of rules that tightened in April 2025. Nvidia disclosed on April 9, 2025 that Washington had imposed a license requirement on H20 chips for China, Hong Kong, Macau, and other covered destinations, and said the move could trigger up to $5.5 billion in charges. (sec.gov) Reuters reported on April 15, 2025 that the Commerce Department extended the same licensing approach to Advanced Micro Devices’ MI308 chips and equivalent products. That meant the two most prominent U.S. suppliers of China-compliant artificial intelligence accelerators were both pushed into case-by-case review. (usnews.com) That review process matters because these chips are the engines inside modern artificial intelligence systems. Training and serving large models takes thousands of linked processors, so a delay in licenses can idle data-center plans the way a missing turbine can stall a power plant. (oracle.com) The squeeze is sharper because compute is already concentrated in a few cloud companies. Epoch AI said on April 14, 2026 that Amazon, Google, Meta, Microsoft, and Oracle together held 67% of the world’s cumulative artificial intelligence compute as of the fourth quarter of 2025, up from 60% in the first quarter of 2024. (epoch.ai) Epoch AI also said labs including OpenAI and Anthropic largely depend on those hyperscalers for research and inference capacity. That leaves frontier work clustered inside a small set of providers even before export controls or licensing delays cut off additional supply. (epoch.ai) Washington has also changed the rulebook while the backlog grew. The Bureau of Industry and Security published its “Framework for Artificial Intelligence Diffusion” on January 15, 2025, then the Commerce Department said on May 13, 2025 that it was rescinding that Biden-era rule before its compliance date. (federalregister.gov, bis.gov) China is not standing still while U.S. licenses pile up. CNBC reported on April 8, 2026 that Alibaba and China Telecom were launching a data center in southern China using Alibaba’s own chips, part of a wider push to build domestic computing capacity under U.S. restrictions. (cnbc.com) For Nvidia and Advanced Micro Devices, the chokepoint is no longer just what Washington bans on paper. It is how fast the Bureau of Industry and Security can decide which shipments move at all. (ttnews.com, sec.gov)
