Should the US Sell Blackwell Chips to China? (ifp.org)

U.S. policymakers are weighing whether to allow NVIDIA to sell a downgraded Blackwell chip called the B30A to China. Reportedly built with one AI processor die and four HBM stacks (versus the flagship B300’s two dies and eight HBM), the B30A would deliver roughly half the B300’s performance at about half the price. That still makes it far more capable than previously approved chips: estimates put the B30A at ~12–17× the power of the H20 and more than 18× current U.S. export-control thresholds. If exported at scale, models show U.S. aggregate AI compute advantage over China could fall from an estimated 31× (if exports are banned) to under 4×, and in aggressive scenarios even flip to a Chinese lead. Technically and strategically, the B30A matters because large clusters of these cheaper chips can match top-tier B300-based training clusters for roughly 20% more cost, enabling Chinese labs and cloud providers to deploy near-frontier supercomputers despite domestic chip limitations. Allowing exports would therefore undermine current export-control goals, accelerate China’s frontier AI development, and risk diverting scarce global fabrication capacity away from U.S. users. The report argues the better policy is tighter controls on semiconductor manufacturing equipment (especially DUV immersion lithography) and high-bandwidth memory to block China’s ability to produce advanced AI chips at scale.