The terafab ceiling

On stage in San Jose in March, Jensen Huang put a number on NVIDIA's combined order book for its Blackwell and forthcoming Vera Rubin chips: $1 trillion in purchase orders through 2027. Twelve months earlier he had floated half that. The doubling was the headline, but the more telling line was what he said about it: "I am certain computing demand will be much higher than that."

The bottleneck isn't whether the chips will sell. It's whether enough fabs exist to make them.

Read carefully, that is a reversal of the usual worry. For two years the open question about the AI build-out was whether the demand was real or a bubble. Huang's pitch is that demand is effectively unbounded and the binding constraint has moved upstream — to silicon. NVIDIA is the dominant buyer of TSMC's most advanced wafers, TSMC is expanding that line at NVIDIA's request, and its leading-edge fabs are reportedly booked through 2028. The ceiling is no longer customers; it is foundry capacity.

Two things deserve a footnote. The $1 trillion is cumulative bookings across two chip generations and two years, not annual revenue — a number that doubled in a year reads more alarmingly than it should. And the constraint bottoms out further up still: the world builds only a few hundred of the EUV lithography machines these fabs run on each year. A trillion-dollar order book, in the end, rests on the output of one Dutch toolmaker.