Compute supply
Who owns and supplies the substrate everything runs on.
capabilityBuilt today Built but scarce · 5 of 5 Open on the map →
Facts
- Lane
- Compute hardware
- Wave
- 2 of 12 (build order, not a date)
- Needs
- 4 boxes all the way back through hard links
- Unlocks
- 107 boxes downstream through hard links
- Leverage
- None: already built
- Cluster
- Oversight and alignment
Evidence as of October 2026
Epoch AI expects power to be the first constraint to bind on scaling, followed by chip packaging and memory.
Detail boxes 4
Chips designed by the cloud companies themselves, such as Google’s TPUs and Amazon’s Trainium.
Anthropic has access to up to a million Google TPUs, and AWS’s Project Rainier runs about 500,000 Trainium2 chips. Labs now assemble diversified compute portfolios.
Chips built to serve models fast and cheaply rather than to train them.
Nvidia paid $20bn to license Groq’s technology and hire most of its team.
Specialist clouds such as CoreWeave and Nebius that rent GPUs to labs, often financed with debt.
CoreWeave has 1.5 GW active and 4.2 GW contracted. Its second-quarter filing showed $35.6bn of debt against $5.5bn of cash, and on 16 October 2026 its shares and Nebius’s each fell about 12%.
The indium phosphide lasers and photonic chips inside the 800G and 1.6T optical transceivers that wire GPU clusters together.
Nvidia committed $2B to each of two InP suppliers in March 2026; China produces about 70% of refined indium, and the shortfall is expected to persist beyond 2027 (TrendForce, 6 Aug 2026).