Nuclear Deals and Gigawatt Clusters: The 2026 AI Infrastructure Power Map

7 min read

The crisis framing is over. Three nuclear plants. Four hyperscalers. A combined 10 gigawatts of power contracted in 18 months. Twelve months ago, analysts were warning that the electrical grid could not keep pace with AI’s appetite for AI power infrastructure. The question in May 2026 is not whether power will flow — it is who contracted the right electrons at the right price, at what scale, and through which technology stack. The answer is a detailed map: nuclear restarts in Pennsylvania, stranded gas in West Texas, fuel cells in Oracle server farms, and small modular reactors still a construction permit away. As we covered in April, the grid stress precipitated by AI data center load growth set the conditions for this procurement surge. This is the 2026 AI power infrastructure deal map.

The Nuclear Anchor: Three Mega-Deals Reshaping the Grid

The most consequential AI power infrastructure contracts of the past 18 months are nuclear, and they cluster in Pennsylvania’s PJM footprint — the grid region that also hosts the highest concentration of hyperscaler data centers on the East Coast.

Constellation / Three Mile Island (Microsoft): The pivot point for the entire sector arrived when Microsoft signed a 20-year, 835-megawatt power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1 — the plant that shut in 2019 because it could not compete on price. Rebranded the Crane Clean Energy Center, it is now roughly 80 percent staffed with more than 500 workers on-site, and restart has been accelerated from 2028 to a probable 2027. The federal government backstopped the $1.6 billion restart with a $1 billion DOE loan, with the first advance expected in Q1 2026. The reported contract price — approximately $110 to $115 per megawatt-hour — is well above wholesale market rates, but Microsoft pays it because nuclear offers something no wind or solar PPA can: guaranteed around-the-clock, carbon-free megawatts matched continuously to data center load. That premium is now the market reference price for 24/7 carbon-free energy in PJM.

Talen / Susquehanna (Amazon): AWS went further. After acquiring Talen Energy’s co-located campus adjacent to the Susquehanna nuclear plant in Pennsylvania — a deal that drew Federal Energy Regulatory Commission scrutiny over grid reliability — the two companies restructured and expanded in June 2025. The new agreement covers 1,920 megawatts, runs through 2042, and is projected to deliver roughly $18 billion in total contract revenue for Talen, up to $1.4 billion per year at full volume. As of spring 2026, Susquehanna completed its refueling outage and transmission reconfiguration, transitioning from a behind-the-meter arrangement to a front-of-the-meter PPA — a design that FERC views more favorably for system reliability. It is the single largest nuclear-to-data-center supply agreement ever executed by volume.

Vistra / Comanche Peak (AWS and Meta): Texas-based Vistra Energy landed on both sides of the hyperscaler ledger. It signed a 20-year PPA with an investment-grade counterparty — widely reported as AWS — for 1,200 megawatts from Comanche Peak Nuclear Power Plant in Hood County, Texas, with delivery ramping from late 2027 to full capacity by 2032. In January 2026, Meta announced separate agreements with Vistra for up to 6.6 gigawatts from three nuclear plants in Ohio and Pennsylvania within PJM by 2035, plus an option on a 300-megawatt SMR. These deals position Vistra as the primary nuclear supplier to two of the four largest AI spenders simultaneously.

The Gas Turbine Race: Speed Over Efficiency

Nuclear deals take years to negotiate and years more to deliver. Data center buildout schedules are measured in months. That gap is filled by natural gas.

GE Vernova’s HA-class gas turbines — capable of more than 64 percent combined-cycle efficiency — have become the workhorse of AI power infrastructure. In Q1 2026 alone, GE Vernova signed 21 gigawatts of new gas turbine agreements, pushing its total contracted backlog from 83 to 100 gigawatts in a single quarter. Data center customers contributed $2.4 billion in electrification segment orders in Q1, exceeding GE Vernova’s full-year 2025 data center order total. Delivery slots are booked through 2030.

For deployments that cannot wait even for HA-class turbines, GE Vernova’s aeroderivative LM2500XPRESS offers a modular alternative. Crusoe Energy took delivery of 29 of these packages to power its AI data centers — behind-the-meter, without a grid interconnection queue, in months rather than years.

The Permian Basin is proving to be a specific arbitrage opportunity within this gas story. The basin produces so much associated natural gas that pipeline egress is regularly overwhelmed, and spot prices frequently trade negative. Crusoe, partnering with Engine No. 1 and Chevron, secured 4.5 gigawatts of natural gas power for AI data centers, using this stranded gas supply. Its Abilene, Texas campus is expanding to 1.2 gigawatts across eight buildings — completion targeted for mid-2026. Chevron separately announced a data center power campus in the Permian. The GW Ranch Project in Pecos County aims for 7.65 gigawatts. The Permian is becoming a gigawatt cluster in its own right.

Gas turbine capital costs, however, are rising steeply. The cost to build a combined-cycle gas turbine plant surged from under $1,500 per kilowatt in 2023 to $2,157 in 2025 — a 66 percent increase — driven by the same data center procurement that makes turbines so attractive. At the high end of the capital cost range, LCOE for gas reaches $49 per megawatt-hour, narrowing the gap with nuclear PPAs and making the premium for 24/7 carbon-free power look less extreme in comparison.

Fuel Cells: The Off-Grid Option

Between the multi-year lead times of nuclear and the carbon exposure of gas, Bloom Energy’s solid oxide fuel cell technology has found a market niche: fast-deploy, on-site power for data centers that cannot wait for either.

Bloom’s SOFC systems run on natural gas (or hydrogen blends), achieve approximately 60 percent electrical efficiency, require no water for cooling, and can be deployed in modular increments. In January 2026, Bloom and Oracle expanded a strategic partnership to 2.8 gigawatts — 1.2 gigawatts already under contract and in deployment for Oracle’s AI infrastructure build. A $5 billion partnership with Brookfield is rolling out Bloom technology across AI data centers globally. Between October 2025 and January 2026, fuel cell companies closed $7.65 billion in binding agreements to power AI facilities. Bloom’s 2026 Power Report found that one-third of data centers expect to be fully off-grid by 2030 — a structural bet on on-site generation rather than utility power that makes fuel cells’ grid-bypass speed advantage permanent rather than transitional.

The strategic logic mirrors the gas arbitrage: data centers facing years-long grid interconnection queues can bypass the queue entirely with on-site fuel cells, at a cost premium compared to wholesale power but a time savings that may be worth hundreds of millions in accelerated revenue for a hyperscaler racing to deploy AI capacity.

The Grid Stress Test: PJM and ERCOT Under Pressure

The deals being struck in boardrooms are registering in real time on capacity markets.

In PJM’s December 2025 capacity auction — setting prices for the 2027/28 delivery year — clearing prices hit the $333.44/MW-day cap for the third consecutive auction, a new record. Data centers accounted for $6.5 billion, or 40 percent, of the $16.4 billion total capacity cost burden according to PJM’s independent market monitor. The 2027/28 load forecast is approximately 5,250 megawatts higher than the prior year’s forecast, with 5,100 megawatts of that increase traceable directly to data center additions. The auction procured 6,623 megawatts less than PJM’s 20-percent reserve margin reliability target — a shortfall the market monitor called unprecedented in its uncertainty profile. MISO capacity auctions are showing similar tightening, with large-load interconnection requests from hyperscalers driving queue backlogs across the Midwest grid.

In ERCOT, summer 2026 peak demand is expected to reach 90.5 to 98 gigawatts, approaching but not yet exceeding the all-time record. As of February 2026, Texas had 158 proposed projects to build or expand gas power plants — roughly 40 gigawatts targeted directly at data center power supply. ERCOT itself has cautioned that aggressive longer-range data center demand projections may not materialize at the projected pace, introducing significant uncertainty into both the grid planning and capital deployment calculus.

Small Modular Reactors: The Next Map Layer

The deals signed in 2026 are mostly for power from existing nuclear plants or gas infrastructure deployable in the near term. The SMR layer is the longer-dated bet — and it is being placed now.

Amazon has committed to deploying 5 gigawatts of X-energy’s Xe-100 pebble-bed reactors by 2039, an anchor order that enabled X-energy’s April 2026 IPO at $16 to $19 per share — the largest commercial deployment commitment of any next-generation reactor design. Google signed the first US corporate SMR fleet deal with Kairos Power — 500 megawatts of fluoride salt-cooled high-temperature reactors, with the first site targeted by 2030. TerraPower’s Natrium reactor has begun non-nuclear construction at a retiring coal plant in Kemmerer, Wyoming; an NRC construction permit decision is expected in the first half of 2026. Meta is funding two Natrium reactors for 690 megawatts of delivery by 2032.

The NRC is expected to issue its first two commercial SMR construction permit decisions during 2026 — a regulatory milestone that will either validate or delay the entire sector timeline. SMRs are not a 2026 power source. They are a 2030 to 2035 power source being contracted and funded now, while existing nuclear plants and gas turbines hold the line.

The Cost Map

The economics of AI power infrastructure in 2026 reveal a convergence: the gap between the cheapest (gas) and the most reliable (nuclear) is narrowing as turbine procurement costs inflate.

Nuclear PPA prices for existing plant restarts are running $100 to $115 per megawatt-hour — a scarcity premium over the $40 to $60 wholesale market range, but converging with new-build gas under current turbine cost inflation. Goldman Sachs modeling puts new-build nuclear at approximately $77 per megawatt-hour with a $100-per-ton carbon price — cheaper than a near-100-percent-renewable solution ($87/MWh) or gas combined cycle ($91/MWh) under the same carbon scenario. Bloom SOFC fuel cells land at roughly $80 to $110 megawatt-hour equivalent depending on gas prices and installation cost.

The practical conclusion: lock in long-term nuclear PPAs now at $100 to $115 per megawatt-hour, deploy gas turbines and fuel cells to fill the gap while nuclear projects come online, and hold SMR optionality for post-2030 base load. The companies that secured power contracts in 2023 and 2024 at terms unavailable today have a structural advantage in AI infrastructure buildout for the next decade.

What This Map Means

The 2026 AI power infrastructure map is not a crisis document — it is a procurement ledger. Three structural shifts are now visible. First, power contracts are moving from sub-500-megawatt renewable deals to gigawatt-scale commitments tied to specific physical plants — nuclear plants, gas campuses, fuel cell deployments. Second, site selection has inverted: power availability now precedes every other criterion, including fiber connectivity, land cost, and proximity to customers. Third, the grid is being partially bypassed — not abandoned, but supplemented — by on-site generation at a scale that would have seemed implausible before AI data centers crossed the gigawatt threshold.

The next layer of deals — SMR contracts and Permian Basin mega-campuses — will determine who controls AI power infrastructure in the 2030s. That map is being drawn now.

---

Arjun Mehta covers AI infrastructure and semiconductor hardware for NextWave Insight.