The Pulse
As of May 2026, every major AI hyperscaler has signed at least one nuclear power deal. Thirteen announced projects now commit more than 9.8 gigawatts of nuclear capacity to AI infrastructure.
Microsoft, Amazon, Google, and Meta are not just buying electricity anymore. They are buying power plants, restarting reactors, and in some cases building gas turbines on site because the grid cannot connect them fast enough.
The nuclear AI power plants story in 2026 is really an energy story wearing a technology costume. AI demand did not create a new energy crisis. It exposed one that utilities have been slow-walking for two decades.
Core significance
Why it matters:
- Thirteen deals, 9.8 gigawatts, and counting: Meta leads with up to 6.6 GW across TerraPower, Oklo, Vistra, and Constellation.[SMRIntel nuclear data center deal tracker]
Microsoft secured 835 MW from the Three Mile Island restart. Google committed to 500 MW from Kairos Power SMRs plus 600 MW from the Duane Arnold restart in Iowa.
- Microsoft revived America’s most infamous nuclear site: In September 2024, Microsoft signed a 20-year, $16 billion power purchase agreement with Constellation Energy to restart Three Mile Island Unit 1, renamed the Crane Clean Energy Center.[Introl nuclear power AI data centers report]
The 837 MW Pennsylvania facility is expected back online in 2028. Microsoft will take 100 percent of its output for AI data centers.
- Energy, not chips, is now the binding constraint: Goldman Sachs has identified energy availability as the single biggest infrastructure constraint for AI, displacing chip supply.[Spaziocrypto nuclear power AI data centers analysis]
Nvidia itself has slowed cluster expansion in some locations, not because of GPU shortages, but because of power shortages. Sundar Pichai admitted to Bloomberg that AI’s growth rate was not anticipated.
Deep context: From grid customer to power producer
For seventy years, technology companies were grid customers. They built data centers near cheap power and reliable infrastructure, then signed standard commercial contracts with local utilities.
That model broke around 2023. US electricity demand had been flat for two decades. AI data centers arrived needing hundreds of megawatts per site, sometimes gigawatts, on timelines utilities could not match.
Interconnection queues in data-center-heavy regions like Northern Virginia now run five to ten years for new large connections.[Vucense big tech nuclear AI data centers 2026]
Seven companies, Amazon, Google, Meta, Microsoft, xAI, Oracle, and OpenAI, responded the same way: stop waiting for the grid and build dedicated generation instead.[Abhishek Gautam big tech own power plants 2026]
The Three Mile Island symbolism
Three Mile Island Unit 2 partially melted down in 1979, becoming American nuclear power’s most recognisable disaster. Unit 1, a separate reactor on the same site, kept running safely until 2019.
It shut down not for safety reasons but because cheap natural gas and renewables made it commercially unviable.
OpenAI’s Sam Altman has called for an energy breakthrough and personally invested in nuclear startup Oklo, reflecting how central nuclear has become to frontier AI strategy.[TechCrunch Microsoft Three Mile Island report]
Microsoft’s 2024 deal gave Constellation the revenue certainty to justify the restart. The plant Microsoft is reviving was never involved in the accident. Five years from shutdown to AI-powered restart is a remarkably fast reversal for an asset everyone assumed was permanently retired.
Data insights
By the numbers:
All figures from named primary sources and industry trackers cited inline.
- 12+: Small natural gas turbines OpenAI, Oracle, and SoftBank strung together to power the Abilene, Texas Stargate site while waiting for grid connections.[Insurance Journal AI gas turbines report]
xAI and Alphastruxure are doing the same at other sites. Single-cycle gas turbines were largely retired technology before AI revived demand for fast-deploy generation.
- 700 to 900 MW: Capacity of Oracle’s two natural gas microgrids proposed for Project Jupiter in New Mexico, part of the $500 billion Stargate initiative.[MEXC Oracle Project Jupiter permits]
The 1,400-acre Santa Teresa campus will eventually house up to 800,000 GPUs across four buildings, primarily serving OpenAI’s next-generation training needs.
- 10%+: Share of total US electricity that AI data centers consumed in Q1 2026, up from roughly 4% in 2023.
A single large AI data center running 50,000-plus Nvidia H100 GPUs draws 150 to 200 MW continuously, equivalent to powering 150,000 US homes.[Epoch AI Stargate site by site capacity tracker]
Table 1: Hyperscaler nuclear deals by company 2026
| Company | Total commitment | Lead deal | Online date | Structure |
| Meta | Up to 6.6 GW | TerraPower, Oklo, Vistra, Constellation portfolio | Late 2020s to 2030s | Multiple PPAs plus development agreements |
| Microsoft | 835 MW confirmed | Crane Clean Energy Center, Three Mile Island Unit 1 | 2027 to 2028 | $16B 20 year PPA, 100% output |
| 1.1 GW combined | Kairos Power SMRs plus Duane Arnold restart | 2029 to early 2030s | PPA plus SMR development deal | |
| Amazon | Multiple GW | Talen Energy Susquehanna plus X energy SMRs | Phased, underway | 17 year PPA plus $700M SMR investment |
Table 2: Nuclear versus gas versus grid for AI power
| Dimension | Nuclear restart or SMR | On site gas turbines | Standard grid connection |
| Time to power | 3 to 5 years for restarts; SMRs early 2030s | Months to 1 to 2 years | 5 to 10 years in constrained regions |
| Carbon profile | Near zero operational emissions | High emissions, local air quality concerns | Depends on regional grid mix |
| Capital structure | Long term PPA, 17 to 20 years | Capex for turbines plus fuel supply | Standard utility tariff |
| Companies using it | Microsoft, Google, Amazon, Meta | OpenAI, Oracle, xAI, SoftBank | All companies as baseline |
| IMAGE PROMPT BOX Prompt: A modern data center campus at dusk in the Texas desert with rows of natural gas turbine generators visible alongside server warehouse buildings. Industrial photography style with dramatic sky, no text overlays. Alt text: Stargate Abilene Texas AI data center natural gas turbines bridging power gap 2026 Placement: After Table 2, before The business case section |
The business case: What this means for enterprise AI buyers
Energy procurement is now a frontier AI capability, not a back office function. Companies that locked in nuclear PPAs in 2024 are securing 2027 and 2028 capacity that latecomers cannot buy at any price.
For enterprises buying AI compute from these providers, the practical effect is capacity reliability. A cloud region backed by a dedicated 800 MW nuclear PPA is less likely to face the throttling or price spikes that grid-constrained regions will see by 2027.
As covered in our AI data center power consumption report, the gap between announced gigawatts and operational gigawatts remains the single most important number to track through 2027.
Expert nuance: The carbon contradiction nobody is resolving
The same seven companies, Amazon, Google, Meta, Microsoft, xAI, Oracle, and OpenAI, that have made public carbon neutrality commitments are simultaneously the largest builders of new gas generation in the United States. [Fortune big tech climate goals data centers fossil fuels]
Nuclear deals get the headlines because they fit the sustainability narrative. Gas turbines get built because they are the only technology that can be deployed in months rather than years.
The honest picture is a barbell: near zero carbon nuclear power purchase agreements for 2028 onward, bridged by emissions heavy gas turbines running today. [MLQ AI gas powered data centers emissions analysis]
A Wired analysis of permits found 11 gas fired data center projects connected to OpenAI, Meta, Microsoft, and xAI could release more than 129 million tons of greenhouse gases yearly at full capacity, more than Morocco’s entire 2024 national emissions.
Strategic outlook
- Watch the SMR delivery timelines closely: Every SMR deal, Kairos, X energy, TerraPower, Oklo, targets early 2030s delivery.
None has an operating commercial reactor yet. The first one to actually deliver power, rather than sign another PPA, resets the entire sector’s credibility.
- Grid Strategies projects 15%+ demand growth over five years: This follows two decades of flat US electricity demand.[IEEE Spectrum big tech nuclear power AI]
That reversal is structural, not cyclical. Utilities planning on old demand curves will be the next bottleneck after interconnection queues.
- Natural gas is the swing fuel through 2028: Every nuclear restart has a multi-year gap that gas fills.
Enterprises evaluating AI providers on sustainability credentials should ask specifically what powers today’s clusters, not just what powers 2029’s.
Key question answered
Why are Microsoft, Amazon, Google, and Meta buying nuclear power plants for AI?
Because AI data centers now require hundreds of megawatts to multiple gigawatts per site, and US grid interconnection queues in major data center regions run five to ten years.
Nuclear power purchase agreements, including Microsoft’s Three Mile Island restart and Meta’s 6.6 GW portfolio, lock in carbon free baseload power for the late 2020s. In the near term, the same companies are building natural gas turbines on site because nuclear restarts and SMRs cannot deliver power before 2027 at the earliest.
The takeaway
The nuclear AI race is really a story about US energy infrastructure failing to keep pace with a demand curve nobody modelled five years ago.
Thirteen nuclear deals and 9.8 gigawatts sound like a climate win, and by 2029 they may be. But the gas turbines running at Abilene, Memphis, and soon New Mexico are the honest 2026 picture.
For enterprises, investors, and policymakers, the number worth tracking is not gigawatts announced. It is gigawatts actually delivering electrons, and through 2027 that number will be overwhelmingly natural gas wearing a nuclear company’s press release.
