Skip to content

AI Supply Chain

The physical bottleneck stack behind frontier AI compute. jensen-huang says these are what keep him up at night — not algorithms.

The three pinch points

  1. ASML — extreme ultraviolet (EUV) lithography machines. Monopoly supplier.
  2. TSMC — advanced packaging (CoWoS) required for modern GPU + HBM assemblies. Capacity is gated and takes years to add.
  3. SK Hynix / Samsung / Micron — high-bandwidth memory (HBM). Quality and yield gate model serving economics.

Why it's hard

NVIDIA is growing faster than any company in history while accelerating. Aligning ASML → TSMC → HBM → NVIDIA → hyperscaler capex on multi-year lead times, without over- or under-building, is a coordination problem that "informing CEOs" (his verb) is a meaningful part of the job.

Power delivery is the adjacent constraint: if utilities could offer segmented power delivery guarantees ("I can give you X MW by month Y at price Z"), the industry could plan around it the way it plans around fab capacity.