The AI power & grid bottleneck — transformers, turbines, interconnection, fuel

Web-verified 2026-06-11. Structured + sources: macro-ai-power-grid-bottleneck.json. Overlay — evidence-graded, excluded from the proofs, but mirrors the machine-proven power_adequacy P1 (AI power demand > supply through 2028). Connects to energy_web, fin-ai-depreciation-debttrap, geopolitics-russia-energy-arctic (HALEU), fin-coreweave-oracle.

The AI build is power-constrained, not just capital-constrained. ~$650B+ of hyperscaler 2026 capex (Alphabet/Amazon/Meta/Microsoft) collides with multi-year lead times for the physical grid — and nearly half of US data-center projects may be delayed or cancelled due to grid + component bottlenecks. The money isn't the binding constraint; the grid is.

The bottlenecks

Transformers. High-voltage transformer lead times went from ~50 weeks to ~127 weeks (some critical units 3–4 years); HV substations 3–5 years; the global shortage is expected to last to at least 2029. Domestic supply is short, so developers import — including from China, adding a geopolitical-supply risk to the grid itself.
Gas turbines. The "big three" (GE Vernova, Siemens Energy, Mitsubishi) are effectively sold out for years; new firm gas capacity pushes toward ~2028–2029 — the same window as the demand.
Interconnection. US interconnection queues delay projects for years; utilities warn of regional capacity shortages as early as 2026.
Fuel. Both "clean firm power" answers have chokepoints: nuclear/SMR fuel (HALEU) is ~Russia-gated (geopolitics-russia-energy-arctic; power_adequacy P2 UNSAT to ~2029), and gas needs turbines + pipelines that are themselves backlogged.

The workaround — "Bring Your Own Power"

Because the grid can't connect them in time, hyperscalers go BYOP — behind-the-meter on-site generation (gas turbines, fuel cells, nuclear restarts like Three Mile Island, SMRs). This bypasses the HV-transformer + interconnection bottleneck, but (a) still needs turbines/fuel that are backlogged or gated, and (b) moves the AI build off the regulated grid into self-supplied power — privatizing the energy chokepoint.

Why it matters to the map

Physical mirror of the financial thesis. The $650B+ capex and the $1.4T of OpenAI compute commitments (fin-microsoft-openai) assume power on a 3–5-year delivery clock the 2026–2027 buildout can't meet — so a large share of announced capacity is, like the marks, a promise ahead of deliverable substance.
It compounds the depreciation trap (fin-ai-depreciation-debttrap): GPUs that economically age in ~2–3 years, sitting idle waiting for power they can't get, are stranded, fast-depreciating capital — debt-financed assets dying before they're energized.
It hands leverage to whoever controls turbines (a few firms), transformers (a strained global supply incl. China), and fuel (Russia/HALEU, gas) — the energy chokepoint behind the compute chokepoint. The AI build can't buy past the grid any faster than it can buy past Taiwan.

← Research index · structured data: macro-ai-power-grid-bottleneck.json · macro-ai-power-grid-bottleneck.md