AI infrastructure capacity crisis: even cloud giants are running out of runway

The week of June 16, 2026 produced a cluster of stories that, taken together, describe a structural AI infrastructure capacity crisis that goes far beyond any individual company problem.

The signs are everywhere

• Microsoft is routing GitHub traffic through AWS
• Google is paying SpaceX $920 million per month for compute capacity
• Anthropic is paying SpaceX $1.25 billion per month for the Colossus 1 facility
• NVIDIA has Trainium chips at Amazon sold out through 2028
• Oracle has a $638 billion backlog of committed AI infrastructure contracts it is racing to build

The common thread: AI demand is consuming infrastructure faster than it can be built.

Why is it so hard?

The crisis is multi-layered:

1. Data center construction: takes 18 to 36 months from site selection to operation
2. Chips: NVIDIA Blackwell and Vera Rubin supply is constrained by TSMC advanced node manufacturing capacity
3. Power: AI racks draw 40 to 100 kW each (vs. 10-15 kW for conventional computing), requiring utility-scale power upgrades that take years
4. Investment: Goldman Sachs estimates $7.6 trillion in cumulative AI capex from 2026 to 2031 — equivalent to ~25 percent of annual US GDP

What this means for developers

Two practical implications:

1. AI pricing is unlikely to fall significantly over the next 12-18 months — the infrastructure cost structure is under rising pressure
2. API reliability is under more stress than public incident reports typically acknowledge. Every model provider is managing capacity constraints that can produce elevated latency and service degradation under peak load

Plan for continued constraints into 2027.