Tokens, Watts, and Geography - AI infrastructure

The sell-side consensus on AI infrastructure prices capex in absolute terms — six hundred billion here,
seven hundred billion there — and prices inference output in narrative terms — “the price of intelligence is
collapsing.” Both framings miss the same thing. Inference output is a manufactured good. Its marginal cost
is dominated by electricity. And electricity is not priced globally, it is priced by interconnect.
This paper develops a four-tier framework for the geographic pricing of AI inference between 2026 and
2030. Tier 1 (Sovereign Capable) — China, the Gulf, the Nordic countries, France — combines low
marginal electricity cost with state-backed capacity expansion. Tier 2 (Emerging Stretched) — India,
South Korea, Japan — has committed hundreds of billions to sovereign compute on grids that are not yet
stable enough to clear it. Tier 3 (Power Constrained Wealthy) — the United States in the PJM
interconnect, Germany, the United Kingdom — concentrates capex without the physical electricity to
dispatch it; PJM has now cleared its annual capacity auction at or near the regulatory cap three years in a
row, and the 2027/2028 auction came in short of its reliability target for the first time in the RTO’s history.
Tier 4 (Arbitrage Suppliers) — Russian Siberia, Canadian hydro, the Brazilian Norte/Nordeste, Australian
Pilbara — owns the cheapest electrons on earth and is now exporting them, increasingly, to Tier 1
workloads via cross-border power agreements and dual-resident cloud structures.