dyb

Notes on AI Acceleration Bottlenecks: Energy & Grid Constraints Dominating 2026–2028 Digestion Phase

אם ירצה ה׳

Analysis h/t David Shapiro

Core claim → AI progress enters digestion phase 2026–2028; binding constraints are thermodynamic/grid/physical, not safety/ethics/regulation debates (operationalized into risk frameworks, contribute near-zero friction).

Energy Demand Projections (Cross-verified)

  • 2024 baseline (US operating data-center capacity): ≈25 GW
  • 2026 forecast (S&P Global/451 Research): 75.8 GW (IT + cooling + lighting + ancillary)
  • 2028: 108 GW
  • 2030: 134.4 GW
  • BloombergNEF late-2025 update: 106 GW by 2035 (36% uplift vs prior forecast)
  • Nuclear-equivalent: >130 × 1 GW reactors needed in ≈6 years
  • Stargate exemplar (OpenAI/Microsoft/partners): 16 GW (Nvidia/AMD) + 10 GW custom Broadcom; first GW H2 2026; total infra ≈$1 T

Grid Interconnection Queues

  • Large-facility connection: routinely 5–7 years
  • Northern Virginia hub: developers report 7-year delays (stricter siting, cost-sharing disputes)
  • Implication → new capacity applied 2025–2026 online ≈2032–2033; misaligns with AGI roadmaps targeting 2030

Lead times (Q2–Q4 2025 averages, Wood Mackenzie): - Large power transformers: 128 weeks (≈2.5 years) - Generator step-up (GSUs): 144 weeks (≈3 years) - Reported range: 80–210 weeks; some manufacturers quoting 5 years for new orders

Shortfalls projected: 30% power transformers + 10% distribution units through 2025–2026

Demand surge 2019–2025: high-voltage power transformers +116–119%, GSUs +274%, substations +91%

Physical reality → 100–400 ton units; require rare super-heavy railcars (≈10 nationwide); logistics add months

Nuclear & Baseload Limitations

  • Optimistic restarts/new builds: meaningful contribution only 2030s
  • Small modular reactors → post-crunch arrival
  • Direct deals bypassing grid: Microsoft $1.6B Three Mile Island revival; Google/Amazon similar nuclear tie-ins

Near-term Workarounds

  • On-site microgrids → solar + battery storage + natural-gas turbines (no interconnection delay)
  • Iron-air batteries → stationary, long-duration, cost-effective; manage solar intermittency at scale
  • Desert-scale solar campuses → dedicated feed to AI facilities (Russia hydro-for-aluminum precedent)
  • Fastest deployable: natural-gas turbines + grid-scale batteries

Timeline & Paradigm Shift

  • Acute crunch window: 2026–2028 (demand outruns maturing infrastructure orders)
  • Post-2028 → acceleration resumes as backlogs clear
  • Shift from "scale is all you need" → efficiency, distillation, making-do-with-what-exists

Bottom line (Shapiro): "You cannot purchase your way past the laws of thermodynamics, the lead times on specialized equipment, or the fundamental constraints of electrical grid infrastructure built over decades for a very different world."


Cross-check sources: S&P Global/451 Research (Oct–Dec 2025), BloombergNEF (late 2025), Wood Mackenzie (Q2–Q4 2025 surveys). Numbers hold under scrutiny; no major revisions noted as of Jan 29 2026.