Fabrication (Foundries)
Where designs become physical chips. The single highest-capex industry in the world.
What this layer does
A leading-edge fab is a $20–30B factory operating at the absolute limit of human engineering. Three companies can fabricate at the 3nm / 2nm node today — and effectively only one (TSMC) can do it at scale, yield, and on schedule. Every Nvidia GPU, every Google TPU, every AWS Trainium chip is fabricated there. This is the most strategically important industrial cluster on earth.
Below the leading edge, mature foundries (GlobalFoundries, UMC, Tower) make the analog, power, and RF chips that surround every accelerator. Memory makers (Samsung, SK Hynix, Micron) operate their own integrated fabs — both designers and manufacturers.
Sub-categories
The chips at the heart of every AI accelerator. Effectively a TSMC monopoly with Samsung distant second and Intel re-entering.
Mature nodes (28nm and above). Analog, power management, RF, MCUs — everything around the GPU.
Memory companies own their own fabs. Massive capex cyclicality.
Vertically-integrated logic. Intel is the canonical example; many auto/industrial chipmakers also fit here.
SiC and GaN fabs — relevant for AI power management (Layer 9.7) and grid (Layer 17).