Deutsch
English
Español
Français
Nederlands
Português
In 2026, the primary bottleneck for AI infrastructure is no longer GPU supply or chip fabrication — it is electricity. Global data center power demand is set to surpass 1,000 TWh, driven by AI workloads, while grid interconnection delays, transformer shortages, and substation capacity gaps have pushed nearly half of planned U.S. AI data centers into delay or cancellation. In response, Microsoft, Google, Amazon, and Meta have committed over $100 billion collectively to direct power generation, anchored by 13 announced nuclear-powered data center projects totaling 9.8 GW, including the restart of Three Mile Island and first-of-kind small modular reactor (SMR) deployments. This strategic shift reshapes the energy transition, nuclear economics, and the geography of AI infrastructure globally.
The 1,000 TWh Threshold: Why Power Became the Binding Constraint
The International Energy Agency (IEA) projects that global data center electricity consumption will hit approximately 1,000 TWh in 2026 — equivalent to the entire annual electricity use of Japan or France. This marks a doubling from 460 TWh in 2024, driven overwhelmingly by the exponential growth of generative AI workloads. A single AI training rack can draw 80–140 kW, and a 10,000-GPU cluster requires power levels that most regional grids simply cannot deliver.
Goldman Sachs warns of a 49 GW U.S. generation shortfall by 2028 as data center demand surges 126 GW globally through that year — nearly matching Canada's entire installed capacity. The AI data center power crisis has become the defining infrastructure challenge of the decade. PJM Interconnection capacity prices have spiked nearly tenfold, with data centers driving a $9.33 billion increase in capacity payments across 13 states. Meanwhile, utilities in Northern Virginia, Ohio, and Texas have imposed interconnection moratoriums due to transformer shortages — lead times for high-voltage transformers have stretched from 24–30 months pre-2020 to 3–5 years in 2026.
Hyperscaler Nuclear Commitments: 13 Projects, 9.8 GW, $100B+
Facing these constraints, the four largest cloud providers — Microsoft, Amazon, Google, and Meta — have collectively committed over $100 billion to direct power generation, with nuclear energy at the center of their strategy. According to the Carnegie Endowment for International Peace, announced agreements could provide up to 13 GW of nuclear capacity, though the authors caution that tech companies have remained cautious due to cost, timing, and regulatory uncertainties.
Microsoft and Three Mile Island: A Historic Restart
Microsoft signed a landmark 20-year power purchase agreement with Constellation Energy in September 2024 to restart Three Mile Island Unit 1 — now rebranded as the Crane Clean Energy Center — in Pennsylvania. The plant, which shut down in 2019 for economic reasons, will provide 835 MW of carbon-free baseload power exclusively to Microsoft's AI data centers across Pennsylvania, Chicago, Virginia, and Ohio. The Trump administration approved a $1 billion federal loan in November 2025, accelerating the restart to 2027. Microsoft plans to spend approximately $1.6 billion on refurbishing the plant. This deal triggered a wave of similar agreements across the industry.
Amazon, Google, and Meta: SMRs and PPAs
Amazon Web Services expanded its power purchase agreement with Talen Energy to 1,920 MW and invested $700 million in X-energy to support up to 12 Xe-100 small modular reactors. Google committed 500 MW from Kairos Power's KP-FHR advanced reactors, while Meta signed agreements for up to 6.6 GW of nuclear capacity across multiple developers including TerraPower (Natrium sodium-cooled fast reactor), Oklo (Aurora microreactor), Vistra, and Constellation. In January 2026, Vistra announced agreements with Meta to support existing nuclear plants and develop new generation, signaling an industry-wide race to secure firm, 24/7 carbon-free power.
Small Modular Reactors: From Promise to First Permits
The Nuclear Regulatory Commission is expected to issue the first commercial SMR construction permits in 2026, marking a pivotal moment for the advanced nuclear industry. TerraPower's Natrium reactor received its NRC Construction Permit in March 2026 for the Kemmerer, Wyoming plant. Kairos Power became the first advanced reactor to receive an NRC construction permit in December 2023 and started nuclear construction in May 2025. NuScale Power remains the only SMR with full NRC design certification, having received Standard Design Approval for its 77 MWe US460 module in May 2025.
However, the small modular reactor deployment challenges remain significant. The HALEU (high-assay low-enriched uranium) fuel supply chain is the industry's biggest bottleneck, with current U.S. production at only ~900 kg per year, though a $2.7 billion+ DOE investment aims to scale capacity. First-of-a-kind costs remain 30–50% above nth-of-a-kind targets, a key risk for investors. China's Linglong One is on track to become the world's first commercial land-based SMR in the first half of 2026, while the BWRX-300 at Darlington, Canada marks the first SMR under construction in North America.
Grid Realities: Why Nuclear Is the Only Scalable Option
The fundamental challenge is that AI data centers require 24/7 carbon-free baseload power at a scale that renewables alone cannot reliably provide without massive overbuilding and storage. Solar and wind have capacity factors of 20–35%, while nuclear delivers >90%. Battery storage at grid scale remains prohibitively expensive for multi-day backup. Natural gas is filling the near-term gap, but this creates a contradiction for the carbon-neutral pledges made by all four hyperscalers.
Goldman Sachs identifies energy availability as the biggest infrastructure constraint for AI, displacing chip supply as the binding limit. Gartner predicts that by 2027, power shortages will restrict 40% of AI data centers. The energy transition and AI infrastructure are now inextricably linked, with nuclear power emerging as the only technology capable of meeting both the reliability and carbon-free requirements at the necessary scale.
Expert Perspectives
"The thesis is clear: AI clusters need 24/7 carbon-free baseload power at a scale only nuclear can provide. Existing restarts deliver fastest, while new SMRs offer longer-term scale," notes the SMR Intel tracker, which monitors all nuclear data center deals globally.
John Pendleton and Mackenzie Schuessler of the Carnegie Endowment write in their June 2026 analysis: "Hyperscalers prefer acting as energy offtakers over direct ownership, using two main pathways: power purchase agreements with utilities and direct partnerships with advanced reactor developers. The deals signal enthusiasm but it remains unclear whether big tech will make big bets on nuclear power."
Energy Secretary Chris Wright has stated explicitly that nuclear will be the largest single use of Loan Programs Office funds going forward, with the administration's 2026 budget request including $30 billion in new loan authority oriented toward nuclear projects.
FAQ
Why are hyperscalers turning to nuclear power for AI data centers?
AI data centers require massive, continuous, carbon-free electricity that renewables alone cannot reliably provide. Nuclear power offers 24/7 baseload generation with >90% capacity factor, making it the only scalable solution for the 1,000 TWh demand projected for 2026.
How much nuclear capacity have tech companies committed to?
Microsoft, Amazon, Google, and Meta have announced 13 nuclear-powered data center projects totaling 9.8 GW, with collective commitments exceeding $100 billion. Meta leads with up to 6.6 GW across multiple developers.
When will the Three Mile Island restart be operational?
Three Mile Island Unit 1, now called the Crane Clean Energy Center, is expected to resume operations in 2027 after a $1.6 billion refurbishment. Microsoft has a 20-year PPA for its entire 835 MW output.
What are small modular reactors and why are they important?
SMRs are nuclear reactors under 300 MWe designed for factory fabrication and modular deployment. They offer lower upfront costs, shorter construction timelines (3–5 years), and passive safety features. The NRC is expected to issue first commercial SMR construction permits in 2026.
Will power shortages really restrict 40% of AI data centers by 2027?
Gartner predicts that by 2027, power shortages will operationally constrain 40% of existing AI data centers due to grid interconnection delays, transformer shortages, and surging demand. This forecast underscores the urgency of the nuclear pivot.
Conclusion: The Geography of AI Infrastructure Is Being Rewritten
The nuclear pivot is not just an energy story — it is reshaping where AI infrastructure gets built. Traditional data center hubs like Northern Virginia face mounting grid pressure, while regions with existing nuclear plants (Pennsylvania, Illinois, South Carolina) or SMR development sites (Wyoming, Tennessee, Washington) are gaining attractiveness. The future of AI data center location will increasingly be determined by access to firm, carbon-free power rather than by fiber connectivity or tax incentives.
As the IEA, Gartner, and Goldman Sachs all warn, the window for action is narrow. With 1,000 TWh of demand already here and 40% of data centers facing power constraints by 2027, the hyperscaler nuclear pivot represents the most significant energy-infrastructure shift since the build-out of the interstate highway system. Whether the nuclear industry can deliver on its promises — and whether regulators and supply chains can keep pace — will determine the trajectory of AI itself.
Follow Discussion