The Grid Is the Bottleneck: Nuclear, SMRs, and the Power Race Behind AI

• Demand shock after two flat decades. US electricity demand barely grew from 2005–2020; utilities planned accordingly. Data centers — the IEA projects their global consumption roughly doubling toward ~945 TWh by 2030, with AI the driver — arrived as a step-change concentrated in a handful of metros (Northern Virginia, Dallas, Phoenix, Atlanta), each new campus demanding 100 MW to over 1 GW, the load of a mid-sized city, on a 2–3 year buildout clock.
• Supply can't clear on that clock. Interconnection queues run 5–8 years in constrained regions; new gas turbines are backordered into the next decade; transmission takes a decade to permit. The mismatch — 2-year demand meeting 8-year supply — is the entire crisis in one sentence.
• Prices did what prices do. Wholesale power near hyperscale clusters spiked up to 267%, and the politics arrived on schedule: ratepayer advocates documenting data-center cost-shifting onto households, states drafting special tariff classes, and "AI raised my power bill" becoming a 2026 election-season talking point. Social license, not engineering, may end up the binding constraint on siting.

Hence the hyperscaler pivot from buying power to procuring generation — bypassing the queue by contracting directly for dedicated, firm, carbon-free capacity. And for 24/7 firm carbon-free power at gigawatt scale, the menu has exactly one proven item: nuclear.

The nuclear-for-AI portfolio as of mid-2026 — 13 announced projects, 9.8+ GW committed — sorts into three strategies with very different risk profiles:

• Restarts (the near-term wins): Microsoft's landmark $16B, 20-year PPA to revive Three Mile Island Unit 1 (835 MW, targeted 2027) proved shuttered plants are the fastest gigawatt on the menu — existing licenses, grid connections, and workforces cut timelines to ~3–4 years. Palisades in Michigan led the way back; every other retired-but-intact US unit is now being courted.
• Existing-fleet PPAs and uprates: Meta's 20-year Clinton (Illinois) contract and its multi-GW portfolio across Constellation and Vistra capacity; Amazon's $20B+ campus adjacent to the Susquehanna plant. These monetize the operating fleet — the cheapest carbon-free electrons in America at $30–60/MWh LCOE — and effectively re-rate the entire US nuclear sector, whose equities have behaved like AI infrastructure plays since 2024.
• SMR bets (the 2030s pipeline): Google's 500 MW with Kairos Power, Amazon's $700M X-energy stake targeting up to 12 Xe-100 reactors, Meta's TerraPower and Oklo agreements. Small modular reactors promise factory fabrication, 3–5 year builds, and 50–300 MW increments — but none is commercially operating in the US yet, and first-of-a-kind LCOE estimates run $100–180/MWh versus $30–60 for existing nuclear and $20–50 for renewables. The hyperscalers are deliberately overpaying to buy down the learning curve — venture capital denominated in megawatts.

Strip a power deal to its skeleton and it's the same fixed-cost/contribution analysis as any capital project — which is exactly how the buyers model it:

Illustrative: 1 GW AI campus, 90% utilization (~7.9 TWh/yr)
Power cost at $50/MWh (fleet nuclear PPA):   ~$394M/yr
Power cost at $140/MWh (FOAK SMR):           ~$1.10B/yr
Δ = ~$710M/yr — the premium for new nuclear today

But: campus revenue at stake if power is late,
assuming $10–15B/yr from a 1 GW AI deployment
→ one year of delay costs 10–20x the annual power premium

That last line explains every seemingly irrational number on the deal sheet: for a hyperscaler, time-to-power dominates price-per-MWh. Paying a 2–3x premium for guaranteed 2027 electrons beats cheap 2033 electrons when the compute they enable earns billions annually and the AI land-grab is now. The corollaries:

• Restarts and existing-fleet PPAs win the 2020s (speed at fleet prices); SMRs are a 2030s option being purchased today (the deals are structured as learning-curve subsidies with offtake certainty).
• Gas remains the bridge nobody advertises — turbine backlogs notwithstanding, gas-plus-future-nuclear is the realistic portfolio behind many "clean AI" campuses, with the nuclear share arriving late-decade.
• Behind-the-meter structures (generation feeding the campus directly, bypassing the grid) are the endgame — they dodge interconnection queues and transmission charges entirely, and regulators are only beginning to decide what that means for everyone left sharing the grid's fixed costs.

• Ratepayers are the political fault line. When a data center contracts a plant's output, that supply exits the pool everyone else buys from; when campuses drive peak load, grid upgrade costs socialize across bills. The 267% wholesale spikes near clusters are becoming retail increases with a lag — expect special data-center tariff classes, exit fees, and "bring your own generation" mandates to spread through 2026–2027 state dockets.
• The nuclear supply chain re-rated: uranium (spot prices multiples of their 2020 lows), enrichment (a Russian-dependence problem being onshored with urgency), turbine and construction capacity, and the operating-fleet owners (Constellation, Vistra) whose plants became strategic assets. The trade spread well beyond the reactor builders.
• Energy became an AI competitive moat. Compute you can't power is inventory; the SoftBank-vs-CoreWeave-vs-hyperscaler race is increasingly narrated in megawatts secured rather than GPUs owned. Watch megawatts-under-contract the way you watched chip allocations in 2023 — it's the same scarcity, one layer down the stack.
• The efficiency counterweight is real but outpaced: per-token energy keeps falling (better chips, better models, SLM offload to devices), yet total demand rises faster — Jevons again, this time in joules. Efficiency delays the reckoning; it hasn't canceled it.

The through-line: the AI economy just discovered what heavy industry always knew — cheap, firm power is the ultimate upstream input, and securing it is a decade-scale capital project. The companies treating electrons with the same strategic seriousness as silicon are the ones whose 2030 looks like their pitch decks.