Chinese demand for Nvidia’s H200 artificial-intelligence chip is reported at more than two million units for 2026 — against available inventory of roughly 700,000 — setting off an emergency production ramp and a potential order book above $54 billion, according to reports on the consequences of Washington’s December 2025 decision to allow the chip’s sale to China under conditions.
The numbers, if they hold, describe a policy producing the opposite of scarcity. The December reversal replaced prohibition with a taxed, licensed channel — a 25 percent fee and case-by-case approvals feature in the companion rule our Technology desk examines — on the theory that an American chip sold beats a Chinese chip substituted, and that dependence is leverage. The reported order surge is the theory’s first stress test: two million units of ambition against 700,000 units of fact means allocation, queueing, and a foundry asked to find wafers that a year of planning had assigned elsewhere.
That foundry problem lands in Taiwan. An emergency ramp beginning in the second quarter of 2026, as reported, competes for the same advanced capacity that American hyperscalers have reserved for their own next generations. Every H200 wafer allocated to the China channel is, at the margin, a wafer not allocated to a domestic cloud build — which is why the 50 percent volume cap and licensing conditions in the governing rule are not bureaucratic decoration but the mechanism that decides whose data centre waits.
Sceptics of the reversal make the strategic argument plainly: two million H200s, networked well, train frontier-class models, and no fee recovers the capability once transferred. Supporters answer with the substitution clock: restricted demand accelerated China’s domestic accelerators from adequate to competitive, and an American-installed base — with its software ecosystem — is the only durable moat the United States has yet demonstrated in AI hardware. Both arguments cannot be fully right, and the reported order book suggests Beijing’s buyers have voted: they would rather queue for Nvidia than standardise, yet, on the substitute.
Investors should keep the reported distance. Order interest is not shipped revenue; licences, the fee’s mechanics, foundry allocation and Beijing’s own guidance to its champions can each halve the headline. But directionally, the episode has already delivered its lesson: in AI hardware, demand at this scale does not respect the neatness of policy categories. Washington chose a taxed river over a dam. The reported flood is now testing whether the toll booth was built in the right place.
Related reading:
