Allowing NVIDIA’s B30A exports could cut U.S. AI compute edge, says The Institute for Progress

The Institute for Progress said the U.S. advantage in AI computing power would shrink from about 31× to under 4× in 2026 if Washington permits exports of NVIDIA’s downgraded B30A chip to China, in an analysis published this week.

In the worst scenario, where B30A and comparable chips from other U.S. firms are exported, China could overtake the U.S. in compute gained during 2026, the report found.

Based on reported specifications, IFP assumes roughly half the B300’s performance at roughly half the price for B30A, but notes that a B30A cluster achieving the same peak FLOP/s and memory bandwidth as a B300 cluster would cost about 20% more overall due to networking, energy, and server mix.

The paper assesses nine export-control scenarios and argues that restricting powerful AI chips remains the most effective near-term lever to preserve the U.S. compute lead.