The mechanism: On January 15, 2026, BIS finalized a rule shifting license review for advanced AI chips (Nvidia H200, AMD MI325X-class parts) bound for China from presumption-of-denial to case-by-case — but only if the chips ship direct from the U.S., pass third-party security testing on U.S. soil, cap China-bound volume at 50% of U.S.-customer volume, and clear KYC/remote-monitoring requirements. Blackwell-class silicon (B200, GB200, GB300) stays flatly denied. The effect isn't just "Nvidia sells less to China." It's that every hyperscaler and sovereign buyer now needs two separate supply chains — a full-fat U.S./allied SKU and a compliance-hardened, testable, traceable version — and someone has to design, certify, and assemble both. That someone is the server integrator, not the chipmaker. Nvidia sells silicon; Dell and Supermicro build, test, and ship the boxes that have to prove where they came from and where they're going.
Who cashes in: