Inference Chips Undercut Export Controls

Exporting 'cut‑down' Nvidia H20s to China would still grant a dominant share of inference compute, which is increasingly critical as reasoning agents and long‑task models proliferate. The article argues controls focused only on training‑class chips miss that high‑memory, software‑integrated inference GPUs can erode the U.S. advantage and generate new training data. — It shifts export‑control strategy from a narrow training‑hardware lens to recognizing inference capacity as a strategic lever in the AI race.