Ban the H20: Competing in the Inference Age

How inference scaling should change US AI strategy

A guest post by Venkat Somala. He previously served at the Special Competitive Studies Project, the House Select Committee on the Chinese Communist Party, and has held roles in product and data science in industry. TLDR: U.S. export controls targeting China’s AI capabilities focus primarily on limiting training hardware but overlook the growing importance of inference compute as a key driver of AI innovation. Current restrictions don’t effectively limit China’s access to inference-capable hardware (such as NVIDIA’s H20) and don’t account for China’s strong inference efficiency. While…

saved by

• David Oks

related reading

• Dario Amodei — On DeepSeek and Export Controlsdarioamodei.com
• 2028: Two scenarios for global AI leadership \ Anthropicanthropic.com
• China's AI Models Are Closing the Gap—but America's Real Advantage Lies Elsewhere | RANDrand.org
• Should the US Sell Hopper Chips to China? | IFPifp.org
• Should the US Sell Blackwell Chips to China? | IFPifp.org
• Fab Whack-A-Mole: Chinese Companies are Evading U.S. Sanctionssemianalysis.com
• How Much Compute Does China Have?chinatalk.media
• The Inference Shift – Stratechery by Ben Thompsonstratechery.com
• Where will China get its compute in 2026?substack.com
• Why does China open source their AI models?firstscattering.com
• Dario Amodei — On DeepSeek and Export Controlsdarioamodei.com
• Who’s Afraid of Chinese Models? – Stratechery by Ben Thompsonstratechery.com