Thesis Labv0.2.0
Home › Knowledge › Themes › AI power bottleneck

themes

Theme: The AI build-out is power-constrained, not chip-constrained

This is the underlying narrative behind thesis T001. Captured here as a longer reference document.

The core observation

Hyperscalers (Microsoft, Google, Meta, Amazon) are spending capex faster than electricity infrastructure can be deployed to serve their AI data-center load. The chip side of the equation moves on a chip-design cadence (~18-24 months). The electricity side moves on a regulatory + physical infrastructure cadence (5-15 years for new generation, 5-10 years for transmission). The gap is structural and won't close on a business-cycle timeline.

The data points (as of May 2026)

  • US data centers draw ~41 GW — a 150% increase in five years. (Belfer Center)
  • Goldman estimates AI infrastructure capex at $500-700B in 2026 alone. (Goldman Sachs)
  • Utilities plan $1.4T capex through 2030 for AI-driven load growth — a 27% surge over prior plans. (Tech Insider)
  • Interconnection queues exceed 5 years in PJM, MISO, ERCOT. Hyperscalers are building dedicated generation alongside DCs.
  • Hyperscaler nuclear PPA activity is unprecedented:
  • Microsoft-CEG (Crane/TMI, 20-year)
  • Meta-CEG (Clinton)
  • Amazon-TLN (Susquehanna, 1.92 GW through 2042)
  • Meta-Vistra/TerraPower/Oklo/Constellation (up to 6.6 GW)
  • Google-NextEra (Duane Arnold restart)
  • Microsoft-Three Mile Island (20-year)

What it means

The cleanest market expression is the operators of existing dispatchable, low-carbon, in-load-pocket generation. Their assets are unreplicable in any reasonable timeframe and their negotiating leverage with hyperscalers is growing, not eroding.

Tiers of expression:

  1. Pure-play nuclear utilities — CEG (largest fleet), TLN (Amazon template), VST (post-Calpine equivalent). Most direct.
  2. Diversified IPPs with significant nuclear — CEG post-Calpine fits here too. Adds ERCOT gas optionality.
  3. SMR developers — OKLO, SMR. Speculative; 5-10 years from material commercial output.
  4. Picks-and-shovels — uranium miners (CCJ), enrichment (LEU), HALEU producers. Indirect but cleaner exposure to "more nuclear period."

The bear case (held as thesis T007)

The strongest bear case is AI capex ROI mismatch: if hyperscaler AI revenue doesn't inflect fast enough to justify the capex, the build-out decelerates and the demand-side support for AI-power evaporates. Goldman has framed this as a $450B consensus revenue vs $1T required.

Key markers to watch for the bear: - Any hyperscaler explicitly cuts AI capex guidance. - Hyperscaler aggregate capex guidance cut >15%. - Training compute demand plateaus (a major lab reports diminishing returns on scaling).

These are the falsifiers on T001 and T002, and they're the support evidence for T007.

Specific risks to monitor

  • PJM capacity auction trajectory. Sustained pricing supports the thesis; a meaningful drop falsifies it. See pjm capacity market.
  • NRC license renewals. 80-year license extensions are the long-tail terminal-value driver. Slow NRC progress = lower NPV.
  • Geopolitical event affecting fuel supply. HALEU and enriched uranium come heavily from Russia/Kazakhstan; supply shock is a tail risk.
  • Hyperscaler vertical integration. If Microsoft / Meta / Amazon directly acquire generation assets (Talen is rumored), the equity-market exposure compresses.

Educational layer for the user

The user is learning markets through this work. The teaching arc for this theme:

  1. Capacity factor — why 1 GW of nuclear ≠ 1 GW of solar from the buyer's perspective (capacity factor)
  2. PPAs — how hyperscalers contractually capture power (power purchase agreement)
  3. Capacity markets — why grid operators pay for availability, separate from energy (pjm capacity market)
  4. Nuclear PTC — the federal floor under existing reactor cash flows (nuclear production tax credit)
  5. Interconnection queues — why building new power takes 5+ years (to write)

Once these five are internalized, the user can read any utility 10-K and reason about it independently — which is the actual goal.

Historical analog

The closest historical analog might be the 1990s telecom backbone build-out. Massive capex into fiber and switching infrastructure ran ahead of demand for several years (the bandwidth glut), then the demand caught up (post-2000) and the surviving infrastructure owners (Level 3, Williams, eventually fiber acquired into Verizon/AT&T) earned excess returns for decades. The risk in the analog is the intermediate bust — the 2001-2003 telco collapse — even though the long-run thesis was right.

The lesson: even if the multi-decade thesis is correct, the path may include a violent intermediate drawdown if capex overshoots demand. The cash reserve in the portfolio (~33%) is partly designed for that contingency.