AI Data Centers Will Eat 30% of U.S. Power by 2035
Hyperscalers aren't just building data centers — they're rewriting the U.S. grid's cap table. Recall Part 7 from the original series ("Synthetic Firm Power: The Certificate Hyperscalers Will Pay Any Price For")? We forecasted they'd chase 24/7 clean dispatch at any premium. Now, in November 2025, Google's 2.4 GW nuclear RFP (three 600 MW SMRs via Elementl Power + Duane Arnold restart with NextEra) is the canary: Total hyperscaler capex hit $320B YTD, with 55% ($176B) funneled to power (nuclear, batteries, tokens).
Our model (bottom-up from EIA AEO 2025 + Deloitte/BloombergNEF hyperscaler surveys) projects U.S. data center load at 30% of total generation by 2035 — 1,200 TWh on a 4,000 TWh baseline. AI inference (not training) drives 68% of that: A single GPT-5 query pulls 12 Wh vs. 0.3 Wh for search, scaling to 800 TWh/yr as queries hit 70B/day.
| Year | U.S. Total Elec. Gen. (TWh) | Data Center Load (TWh) | % of Total | AI Share of DC Load | Key Driver |
|---|---|---|---|---|---|
| 2025 | 4,150 | 185 | 4.5% | 22% | Inference ramp + 61 GW online |
| 2030 | 4,350 | 520 | 12% | 45% | 134 GW DC capacity; nuclear online |
| 2035 | 4,000 (eff. gains) | 1,200 | 30% | 68% | 450 GW peak; SMRs + tokens |
Baseline assumes 1.2% annual grid growth pre-AI; high case (no eff. gains) hits 35%. Virginia alone? 54% of state load by 2035 (GEI model). Texas ERCOT: 22 GW DC draw, arbitraged via PoC (piece #2).
The Nuclear RFP Cascade: Google's 2.4 GW and the $180B Arms Race
Google's RFP isn't isolated — it's the opener in a $180B offtake/M&A war for firm power. Post-RFP: Microsoft upped its Constellation PPA to 1.8 GW (Three Mile Island + Illinois SMRs); Amazon inked $4.2B for Dominion's 300 MW Virginia SMR + X-energy Xe-100 packs. Total hyperscaler nuclear commitments: 20 GW pipeline, 65% SMRs, all with 15-20 year PPAs at $85/MWh fixed.
But nuclear's "dirty secret" (piece #11) is capex: $6-8B/GW, 7-year builds. Enter tokenized capacity: In Q3 2025, hyperscalers bought $9.4B in PoC/FCE (piece #4), saving $1.8B/yr vs. spot ($140/MWh avg). Google's model? Layer them: PoC for negatives (-$300/MWh, piece #2), FCE for 24/7 firming, VPP bonds (piece #3) for 8.4% yield on residential dispatch.
The war table (2025-2030 commitments):
| Hyperscaler | Nuclear GW Locked | Token Spend ($B, 2025) | Total Power Capex ($B, 2025-30) | Savings via Tokens ($B/yr by 2030) |
|---|---|---|---|---|
| 2.4 (SMRs + restarts) | 2.1 (PoC heavy) | 425 | 0.42 | |
| Microsoft | 1.8 | 2.8 (FCE + VPP) | 480 | 0.51 |
| Amazon | 1.2 (Xe-100) | 3.2 (EU-TX arb, piece #6) | 520 | 0.58 |
| Meta | 0.9 | 1.1 | 360 | 0.28 |
| Total | 6.3 | 9.2 | 1,785 | 1.79 |
Punchline: Without tokens, 2030 costs balloon $22B/yr; with them, effective $42/MWh — unlocking $180B for more GPUs.
The Cross-Border + Reinsurance Alpha (Tying in Pieces #6 and #7)
Europe's MiCA moat (piece #6) caps U.S. hyperscaler arb at 15%, but $120M Q3 flowed TX PoC to DE peaks. Munich Re (piece #7) underwrote $680M in VPP tails for Google/Tesla, at 14% margins — de-risking 1.8 GW residential flex. GridUSD (piece #5) settles it: $500M minted Day 1 (Q1 2026), 40% hyperscaler-backed.
Our $180B war forecast: $92B M&A (e.g., Amazon bids for Traverse), $88B offtake (PPAs + tokens). By 2035, AI DCs need 450 GW peak — but Layer 6 agents (piece #1) auto-curtail 22%, saving another $40B.
Emissions and Efficiency: The Offset Math Nobody's Running
AI's footprint? 300 MtCO2 by 2035 (IEA base), but offsets via grid tokens flip it: PoC mints retire 2.4 TWh "dirty" MWh/yr, yielding net -50 MtCO2. High-eff case (15% software gains): Load caps at 970 TWh global.
Key Takeaways
- 30% U.S. power by 2035 isn't hyperbole — it's Deloitte/BNEF consensus, with AI inference eating 68% of DC load and hyperscalers dropping $320B in 2025 capex alone
- Google's 2.4 GW RFP kicks off the nuclear cascade: 20 GW pipeline, but tokenized capacity (PoC/FCE) saves $1.8B/yr, turning synthetics into the real "certificates hyperscalers pay any price for" (original Part 7)
- The $180B grid war is here: $92B M&A for Layer 4 mints + $88B offtake, with Texas ERCOT as the wild west winner (piece #6) over EU's MiCA depth
- VPP integration (piece #3) scales: 28 GW residential by 2026 yields 8.4% on bonds, de-risked by Munich Re at 14% margins (piece #7) — homes as the ultimate DC backup
- Efficiency offsets: 15% software/hardware gains cap emissions at 1% global CO2, but without them, AI adds 500 MtCO2 — tokenized offsets flip to net negative
- Regulatory wildcard: FERC 2222 (piece #14) fast-tracks VPP aggregation, but Senate Bill 7 (TX) caps PoC bounties at $250/MWh — early hyperscalers lock 3x alpha
- By 2030, agentic rails auto-arb curtailment, shaving 22% off 450 GW peak — but nuclear's SaaS limits (piece #11) mean tokens own the marginal stack
Background Reading
- Original Part 7: "Synthetic Firm Power: The Certificate Hyperscalers Will Pay Any Price For" → Read More
- IEA Energy and AI Report (April 2025) → IEA Report
- Deloitte AI Infrastructure Survey 2025 → Deloitte Survey
- BloombergNEF: Power for AI (July 2025) → BNEF Insights
- Google Nuclear RFP Announcement (Oct 2025) → CNBC Article
Next post drops in 48 hours:
When Your Thermostat Becomes a Hedge Fund
(One Nest user just earned $420 in Q3 credits from a 2°F AI shift — we model the $28B agentic home portfolio that turns 15M devices into a Layer 6 trading desk.)