Where do the gigawatts come from?
The grid is the default and slow. Behind-the-meter is fast and often dirty. Nuclear is patient and clean. Hyperscalers are placing all three bets at once.
The fastest path to 1 GW of new on-site power in 2026 is a natural-gas turbine farm. The cleanest path with comparable lead time is a Bloom-style fuel cell array. Everything else — utility interconnection, SMRs, geothermal — sits on a 5–15 year timeline.
The menu, sorted by lead time
Every gigawatt of new AI power capacity in 2026–2030 is some weighted blend of the options below. Hyperscalers are not picking one and committing — they are signing parallel contracts, so that whichever path clears regulatory and supply-chain hurdles first is the path that gets used.
Grid: the default, and the slow path
If you ask a utility for 1 GW of new load, the answer is some combination of "yes, eventually" and "you'll pay for the upgrades." The interconnection queue — the line of projects waiting to physically connect — averages 5+ years in most US markets, longer in PJM and CAISO. The wait is not the utility being lazy; it's transformer manufacturing lead times, transmission upgrades, and federal/state permitting that nothing in the system can compress.
For hyperscalers, the grid is what you build the site around but rarely what you depend on for the first 1–3 years of operation. The interconnection eventually arrives; the cluster needs to be live before then.
Natural gas: fastest, dirtiest, most permitted
If you want 500 MW of generation on-site within 18 months, the answer is a row of natural-gas turbines. The technology is mature. The supply chain is mature. The permitting path is well-understood. Combined-cycle gas is ~55% efficient and emits roughly half what coal does per kWh — clean, in the relative sense, but not in the absolute sense.
Hyperscalers with public 2030 net-zero commitments have a problem here: every gigawatt of behind-the-meter gas they sign for makes the 2030 target harder. The current pattern is to sign gas as bridge generation (3–7 year contracts) with the explicit assumption that nuclear or fuel cells will replace it on the back end.
Fuel cells: on-site, modular, expensive
Solid-oxide fuel cells (Bloom Energy, FuelCell Energy) and proton-exchange-membrane fuel cells (Plug Power) convert natural gas or hydrogen directly to electricity at ~50–60% efficiency, with no combustion. They install in modular containers, scale from MW to hundreds of MW, and skip the grid interconnection entirely.
The downside is cost-per-kWh — fuel cells run roughly 2–3× the marginal cost of grid electricity, and the capex is heavy. But for a hyperscaler whose binding constraint is time-to-power, not cost-per-kWh, fuel cells are a meaningful fraction of the answer. Bloom alone announced more than 1 GW of new data-center deployments through 2026.
Solar plus battery: cheapest energy, hardest to schedule
Utility-scale solar produces electricity at ~$25–40/MWh, the cheapest source on the grid before storage. The problem for AI loads is that AI clusters do not idle when the sun goes down — training runs are 24/7 by design. Solar must be paired with multi-hour battery storage, gas firming, or grid backstop.
The math gets more attractive each year as battery prices fall — utility battery storage is at ~$200/kWh for 4-hour systems in 2026, half what it was in 2022. But solar+battery alone cannot yet firm a 1 GW continuous load economically. It is a meaningful slice of the mix, not the whole answer.
Nuclear: small modular, patient money
The 2024 Amazon-Talen Cumulus deal — directly attaching a 960 MW data center to an existing nuclear plant — was the opening move. Since then: Microsoft's Three Mile Island restart, Google's Kairos Power SMR offtake, Amazon's X-Energy investment. Hyperscalers want nuclear because it is dispatchable, carbon-free, and matches the 24/7 profile of training compute precisely.
The catch is timeline. Existing reactor restarts can deliver in 24–36 months. Greenfield SMRs (Oklo, NuScale, X-Energy, Kairos) are at 5–10 years from contract to first power, and the supply chain for the first commercial fleet is constrained. Nuclear is real, but it is the 2030+ answer, not the 2026 answer.
Geothermal: the dark horse
Enhanced geothermal — drilling deeper than conventional, using fracking-style techniques — is the one source that could deliver dispatchable, carbon-free, gigawatt-scale power on a 3–7 year timeline. Fervo Energy is the named leader, with operating projects in Nevada and Utah and announced deals with Google. The total addressable resource in the US is comfortably above 100 GW.
Geothermal is real but underbuilt. The bottleneck is drilling capacity and project finance, not physics. If the next two years deliver execution on the announced pipeline, geothermal becomes the cleanest 2028–2032 buildout option.
Why this is the best moment in history to invest in energy
For decades, building new low-carbon generation required subsidies. The economics of nuclear, solar, fuel cells, and geothermal did not work against grid electricity that was already underwriting its own capex. Governments had to pay developers to build, or pay buyers to pay developers to build. Investment-tax credits, production-tax credits, feed-in tariffs, renewable-portfolio standards — the entire policy toolbox existed because the underlying market would not price low-carbon power high enough to fund the buildout.
That premise is now inverted. Jensen Huang put the new frame plainly at Stanford CS153: this is the best chance in the history of humanity to invest in sustainable energy, because market forces are doing the work that subsidies used to do. AI loads will pay $80–$150 per MWh for firm, reliable power on a 5–10 year contract. Almost any clean source — nuclear, geothermal, fuel cells, solar+battery with firming — clears those numbers at scale. The subsidy stack still helps, but the marginal investment decision no longer depends on it.
The downstream implication is broader than which sources get built. The full energy stack — transmission, transformers, switchgear, storage, hydrogen, grid software — has the same demand pull. The constraint is now industrial capacity and supply chains, not policy. The buildout that follows is the deepest grid upgrade since the postwar electrification cycle.
Source: Jensen Huang, Stanford CS153 Frontier Systems lecture, May 13, 2026.
The math hyperscalers are running
In 2026, the typical hyperscaler power stack for a new 1 GW campus looks like:
- 40–60% behind-the-meter mix of gas turbines and fuel cells, online in 12–24 months.
- 20–30% solar+battery for daytime offset and ESG accounting.
- 20–40% grid power on a 3–5 year arrival, ramping as the cluster grows.
- Optional nuclear/geothermal 2028–2032 wedge that replaces the gas portion as it comes online.
The strategic read: the constraint is no longer “does the chip exist.” It is “can you site, finance, and energise 5 GW of compute on the timeline your roadmap demands.” Whoever wins the 2027–2030 capacity race will look back and identify the power-stack decision — three years earlier — as the move that made it possible.