Energy / power leg — the AI + defense power bottleneck, nuclear/SMR PPAs, the turbine wall, and the Russia-enrichment chokepoint

Built from research/energy-power.json (web-verified: IEA Energy & AI; US DOE; PJM auction releases; Utility Dive/GE Vernova; Centrus 8-Ks; DOE/NRC on the Russian-uranium ban). Models: models/graph/energy_web.py, models/z3/power_adequacy.py. Power demand/grid charted elsewhere.

Thesis. Power is a binding physical constraint on the AI buildout, not a footnote. New firm supply lags AI demand every year through 2028; the hyperscalers are racing into nuclear/SMR PPAs; gas turbines have a multi-year backlog; and the fuel for advanced reactors runs through a Russia-controlled enrichment chokepoint.

1. Demand — AI is ~half of US load growth

• Data centers: ~4.4% of US electricity (2023) → 6.7–12% (2028); ~half of all US load growth to 2030; global DC ~945 TWh by 2030.
• PJM capacity auction hit its cap three auctions running ($329 → $333.44/MW-day); ~$9.3B of the increase is data-center-driven — the grid is already repricing scarcity.

2. The hyperscaler nuclear/SMR scramble

• Microsoft — 835 MW, ~$16B 20-yr PPA (~2028, Constellation/Three Mile Island restart).
• Amazon — 1,920 MW through 2042 + SMR exploration.
• Google — 500 MW, first corporate SMR-fleet deal (50 MW by 2030).
• Meta — nuclear procurement (Datacenter Frontier).

These are the AI core reaching for firm, carbon-free baseload — but most SMR capacity is late-decade, behind the demand curve.

3. The supply walls

• Gas turbines: GE Vernova ~80 GW backlog into 2029; heavy-duty lead time ~3 yr, CCGT 5–7 yr; ~80 GW of orders vs ~30 GW/yr capacity. Crusoe ordered 19 turbines (behind-the-meter gas). The turbine wall is a hard near-term ceiling.
• HALEU / enrichment (the Russia chokepoint): Russia is ~44% of global enrichment, ~35% of US imports, 27% of US enriched-uranium demand (2023); HALEU for SMRs was ~100% Russia. The Prohibiting Russian Uranium Imports Act (May 2024, eff. Aug 11 2024, waivers to 2027) banned imports — and Russia retaliated with export curbs. Domestic Centrus produces ~900 kg/yr HALEU (Piketon, OH) — far short of fleet need. Midstream (conversion + enrichment) is the real chokepoint, not mining.

4. The formal results

• Z3 power_adequacy P1 — UNSAT: new firm power supply < AI demand every year through 2028 (2026: 18 vs 25 GW; 2028: 34 vs 80 GW — the gap widens). Earliest balanced year is >2028.
• Z3 P2 — UNSAT: domestic HALEU is short of SMR-fleet need through 2029 → the US stays dependent on Russia-controlled enrichment — the energy analogue of the rare-earth chokepoint.

5. Why on-thesis

Power is one of the three physical chokepoints (capital, rare earths, power+HALEU), and two of the three are adversary-controlled. Cross-ref macro-ai-power-grid-bottleneck, macro-ai-datacenter-water-siting, geopolitics-russia-energy-arctic, defense-rare-earth.

Sources: DOE — data-center electricity demand; Utility Dive — PJM auction; Utility Dive — Russian uranium ban / HALEU; Centrus — 900 kg HALEU milestone.

← Research index · structured data: energy-power.json · energy-power.md