Korva is a scroll of the internet’s best ideas.
Supply chains answer demand slowly
The fab math
Tooling for a gigawatt of chips a year costs $3 billion to $4 billion, and with cleanrooms and the shell the fab bill reaches about $6 billion. Over five years that spend ends up behind more than a trillion dollars of AI revenue. Even after removing half for data centers, power, and everyone else who needs paying, the gap is a hundredfold.
The whip down the chain
The demand signal travels down the chain like a whip. Carl Zeiss makes the mirrors that go into ASML's EUV machines. Earlier this year it did not think it needed enough mirrors for 100 EUV tools a year, and now it plans to reach 100 by the end of the decade. On the current expansion, 100 tools for 2030 is still the right number, though the economics argue for more.
Why money is not enough
Money alone does not fix it. A $10 billion check to Carl Zeiss would change things only if every company in the supply chain got one, and the world is capital constrained: next year's capex is about $2 trillion against lab revenue in the hundreds of billions. Meanwhile people buy turbines and resell them, because the turbine that bottlenecks a data center is worth far more than its price.
Why do small demand changes grow up the supply chain?
Each company orders extra to cover its own uncertainty, and the extra compounds at every step upstream. Jay Forrester at MIT described the effect in 1961, and in 1997 Hau Lee and colleagues named it the bullwhip effect. A small change in what customers buy can swing orders for raw materials far more.
Why are EUV mirrors so hard to make?
EUV light has a wavelength of 13.5 nanometers, and nearly every material absorbs it, so the machines use mirrors instead of lenses. Each mirror carries dozens of alternating molybdenum and silicon layers that reflect about 70 percent of the light. Surfaces must be smooth to within a fraction of a nanometer, about the size of an atom.
How long does a new chip fab take?
A leading-edge fab typically takes two to three years from groundbreaking to first output. Tools are ordered before the building is finished, and some, such as EUV scanners, have lead times of a year or more. Output then ramps over several more quarters as yields improve.
More from Dwarkesh Podcast
