Chips and capital are no longer the binding constraint on the AI buildout. Firm power is. Here is how developers are closing the gap between now and 2032, and what each option costs in carbon.
Last week showed where the AI buildout is actually constrained. In Utah, Fervo Energy's Cape Station delivered first power from enhanced geothermal. In Washington, the Department of Energy awarded $1.9 billion to squeeze more capacity out of existing transmission lines. In North Carolina, regulators rejected a $584 million gas turbine because they did not trust the data center forecasts behind it. And in New Mexico, Oracle reportedly sent a force majeure notice on a campus whose gas supply plan fell through.
Chips and capital matter, but the scarce input is firm electricity: power that is available every hour, delivered to a specific site, on a timeline that matches construction. This piece maps how developers are getting it, how long each path takes, and what each path costs in emissions.
The International Energy Agency estimates data centers used 485 TWh of electricity worldwide in 2025, up 17% in a single year, and projects roughly 950 TWh by 2030. AI-focused facilities are on track to triple over the same period.
In the US, Lawrence Berkeley National Laboratory put data center use at 176 TWh in 2023, or 4.4% of national electricity, and projected 325 to 580 TWh by 2028. EPRI's February 2026 analysis puts data centers at 9% to 17% of US electricity by 2030. BloombergNEF projects US data center demand reaching 106 GW by 2035, up from about 25 GW in 2024.
The grid is not keeping pace:
The fastest path to firm power today is natural gas, often built on site or dedicated to a single campus.
Gas is fast relative to the grid, but it has costs beyond carbon. Oracle's Project Jupiter shows the fragility: New Mexico rejected the gas pipeline extension, and the campus is now seeking 2 GW of renewables to replace it. Every onsite turbine also adds Scope 1 emissions that no renewable certificate can offset.
Nuclear is the clean firm option with the most capital behind it, but the timelines split sharply between existing plants and new ones.
Restarts and existing plants (2027 to 2032).
Advanced reactors (2030 onward).
The restarts are the rare case of new clean firm capacity arriving this decade. Everything else lands between 2030 and 2035, after much of the current wave of campuses is energized.
A third category is moving from pilots to contracts.
Geothermal. Fervo's first 33 MW block at Cape Station is online, with the rest of the 100 MW Phase I expected by January 2027 and a 400 MW Phase II targeted for 2028. Google signed a 396 MW PPA with Fervo on September 1, with an option for about 600 MW more.
Multi-day storage. In February, Google, Xcel Energy, and Form Energy announced 300 MW and 30 GWh of 100-hour iron-air batteries in Minnesota, paired with 1.9 GW of wind and solar. Crusoe followed with a 12 GWh agreement with Form Energy in March.
These resources address the hardest part of the problem: clean power at night and through multi-day weather lulls, which is exactly where annual renewable matching falls short.
The cheapest megawatt may be one the data center agrees not to use at peak times. A February 2025 study from Duke University's Nicholas Institute found that existing US grids could absorb 76 GW of new load if that load curtailed 0.25% of the time (about 85 hours a year), and 126 GW at 0.5% curtailment.
Operators are starting to commit to this. In March, Google said it had written 1 GW of data center demand response into utility contracts with Indiana Michigan Power, TVA, Entergy Arkansas, Minnesota Power, and DTE. Texas SB 6 requires new loads of 75 MW or more to install remote shutoff equipment and allows curtailment during grid emergencies. In June, FERC ordered all six grid operators to show cause on their large-load rules, including flexible-load service.
Training workloads can pause, checkpoint, and resume. Inference is less flexible, but it can shift between regions. Flexibility turns a load that the grid has to plan around into a resource that helps balance it.
| Path | Example | Scale | Earliest at scale | Carbon profile |
|---|---|---|---|---|
| Onsite or dedicated gas | Meta Hyperion, Stargate Abilene | 0.3 to 7.5 GW | 2026 to 2028 | High, Scope 1 |
| Nuclear restart | Crane, Duane Arnold | 0.6 to 0.8 GW | 2027 to 2029 | Near zero |
| Existing nuclear PPA | Talen and Amazon | Up to 1.9 GW | Ramps to 2032 | Near zero, but not new supply |
| Advanced nuclear | Kairos, X-energy, Oklo | 0.05 to 1.2 GW | 2030 to 2035 | Near zero |
| Enhanced geothermal | Fervo Cape Station | 0.1 to 0.5 GW | 2026 to 2028 | Near zero |
| Renewables plus multi-day storage | Xcel and Form Energy | 0.3 GW / 30 GWh | 2027 onward | Low |
| Demand flexibility | Google demand response | 1 GW contracted | Now | Avoids new peakers |
The policy trend is consistent: large loads will pay for what they use and prove they are real.
Firm clean power is arriving, but most of it arrives after 2030. The decisions made about the bridge period will lock in emissions for a decade. We will track every major power agreement in our buildout tracker as it happens.
Every hyperscaler matches 100% of its electricity with renewables, and every one of them reported higher emissions this year. The rules behind that paradox are being rewritten. Here is the standard we think the industry should hold itself to.
The energy per AI query has fallen more than 30-fold in a year, and total AI emissions are still climbing. For CIOs and chief AI officers, the levers that matter most are model choice, workload design, and procurement.