9.8 GW of nuclear capacity. That is how much power the four largest technology companies have committed to buying for AI data centers in the last 18 months. Enough to power 7 million homes. Enough to restart a shuttered reactor at Three Mile Island. Enough to make Meta the single largest corporate buyer of nuclear energy in American history.
AI infrastructure has stopped being a compute problem. It has become an energy problem.
AI data centers are projected to consume over 1,000 terawatt-hours of electricity annually by the end of 2026 — more than Japan's entire power grid. A single ChatGPT query draws roughly 2.9 watt-hours against 0.3 for a standard search. At scale, that tenfold gap breaks the economics of conventional grid supply.
PJM, America's largest grid operator, warned it will have barely enough capacity starting summer 2026. Capacity prices have spiked tenfold. Utilities requested $31 billion in rate hikes in 2025 alone. The grid was not designed for 100+ kilowatt server racks running 24/7.
Four Hyperscalers, 13 Deals, 9.8 GW
Every major hyperscaler has now signed at least one nuclear power purchase agreement. The deals break down by strategy.
Microsoft fired the starting gun. Its 20-year, $16 billion PPA with Constellation Energy to restart Three Mile Island Unit 1 — now renamed the Crane Clean Energy Center — created the template. At 835 MW and targeting 2027, a year ahead of the original schedule, it proved that a tech company with enough balance sheet can resuscitate a dormant nuclear asset. The DOE backstopped the $1.6 billion renovation with a $1 billion loan.
Amazon went deeper into next-generation technology. Its $700 million investment in X-energy through the Climate Pledge Fund backs up to 12 Xe-100 small modular reactors. The company also expanded its co-located campus adjacent to the Susquehanna nuclear plant to 1,920 MW, a deal that will deliver an estimated $18 billion in total contract revenue to Talen Energy through 2042.
Google signed the first corporate small modular reactor agreement with Kairos Power, committing to 500 MW of capacity from fluoride salt-cooled reactors. The deal uses an order-book model where Google commits to purchasing power from multiple units as they come online, providing the revenue certainty Kairos needs for financing.
Meta built the largest portfolio: up to 6.6 GW across four partners. TerraPower will supply Natrium sodium-cooled reactors. Oklo will build Aurora micro-reactor campuses. Vistra and Constellation will provide power from existing nuclear fleets. The goal: power Meta's Prometheus AI supercluster in Ohio entirely off the grid.
The SMR Inflection Point
Small modular reactors are the bridge between today's natural gas backup and a fully decarbonized AI infrastructure. But the gap between promise and delivery remains wide.
In March 2026, the Nuclear Regulatory Commission issued the first construction permit for a commercial non-light-water reactor in more than 40 years — TerraPower's 345 MW Natrium design, backed by Bill Gates and partially owned by Meta. It was a landmark. It was also a permit for a plant that will not produce power before 2030.
SMR economics are unproven at scale. First-of-a-kind levelized cost of energy estimates range from $100 to $180 per MWh, compared to $30 to $60 for existing nuclear and $20 to $50 for renewables. NuScale's Carbon Free Power Project, the only SMR design to achieve NRC certification, was terminated in 2023 after the target offtake price rose from $58 to $89 per MWh.
The industry raised $1.3 billion in equity in 2025 — real money, but a fraction of what a single hyperscaler spends in a quarter. HALEU fuel supply is constrained. Transformer lead times stretch 2 to 4 years. Less than 10% of the needed nuclear capacity will be available by 2030.
That gap will be filled by natural gas. The climate math is uncomfortable but the physical math is inescapable.
Who Pays
The cost of AI's power demand is not borne equally. Data center electricity costs have risen 42% since 2019. In Virginia, where data centers consume roughly 40% of electricity, families could face an additional $70 per month on power bills by 2028. Communities in Ohio, Georgia, and Missouri are pushing back against new data center construction.
As we wrote in June, the energy-AI convergence is reshaping infrastructure economics. But the nuclear dimension adds a layer the previous analysis did not fully capture: the hyperscalers are not just buying power — they are building a parallel energy system. Behind-the-meter generation, direct PPAs with reactor operators, and co-located SMR campuses all point toward a future where the largest AI clusters are electrically sovereign.
The Timeline Reality
The first nuclear electrons for an AI data center arrive in 2027 from the Crane Clean Energy Center — Three Mile Island, rebranded and rebooted. Everything after that depends on regulatory speed, supply chains, and the willingness of ratepayers to subsidize infrastructure they did not ask for.
What is not in doubt: the era of assuming the grid will absorb AI's growth is over. The hyperscalers have voted with their balance sheets. Nuclear is no longer optional for AI infrastructure. The question is whether it arrives fast enough.