Microsoft and Amazon are turning to nuclear power to fuel their massive AI operations, but not by building new reactors from scratch. Instead, they're restarting decades-old plants and investing in smaller, factory-built designs that can be deployed faster than traditional nuclear facilities or the fusion systems still years away from commercial viability.

Why Are AI Data Centers Suddenly Draining the Power Grid?

The explosion in artificial intelligence workloads has fundamentally changed how data centers consume electricity. Unlike traditional cloud computing, which allows operators to dial power consumption up and down based on demand, training large language models (LLMs) requires continuous, uninterrupted power. A single AI training cluster can draw between 20 and 100 megawatts around the clock for weeks or months. This constant demand has made power the primary constraint on data center expansion, surpassing even land availability and fiber connectivity.

The pressure is most acute in Northern Virginia, home to the world's largest data center market. The regional grid operator, PJM Interconnection, has received interconnection requests exceeding 4 gigawatts from new facilities alone and warned that reserve margins could fall below reliability thresholds between 2026 and 2028 unless new firm capacity enters service. In Texas, the grid operator ERCOT projects data centers could add 37 gigawatts of load by 2030 under high-growth scenarios. These forecasts have pushed utilities to reconsider nuclear power, since renewable energy paired with batteries cannot yet provide the 24/7 carbon-free electricity that major tech companies need to meet their sustainability commitments.

How Are Microsoft and Amazon Securing Nuclear Power?

Microsoft signed a landmark 20-year agreement to restart Three Mile Island Unit 1, the Pennsylvania reactor that has sat idle since 2019. The plant is expected to supply 835 megawatts starting in 2028, marking the first restart of a U.S. reactor closed for economic reasons. The contract includes performance guarantees tied directly to Azure availability metrics and escalation clauses that allow Microsoft to scale up from the initial 835 megawatts without renegotiating core terms. Constellation Energy, which operates the plant, has already begun recruiting operations staff and ordering long-lead components like steam generators, demonstrating how hyperscaler capital can compress traditional nuclear project timelines.

Amazon took a different approach, committing $500 million to a small modular reactor (SMR) project developed by X-energy, with first units targeted for the early 2030s near existing Amazon Web Services facilities in Washington state. The partnership also includes joint development of a dedicated fuel fabrication line capable of producing high-assay low-enriched uranium pellets at commercial volumes, reducing reliance on foreign supply chains. Amazon is also pursuing power purchase agreements with existing nuclear plants in Ohio and Pennsylvania.

These announcements set a timeline that current fusion companies cannot match. No fusion device has yet reached sustained net electricity production at utility scale, leaving next-generation fission reactors as the nearer-term solution while fusion teams continue their technical work.

Why Small Modular Reactors Over Fusion Energy?

Fusion energy has long been pitched as the ultimate clean power solution, and venture capital has flowed into fusion startups at record levels, reaching $6.2 billion globally in 2023. However, the timeline for commercial fusion remains distant. Most public roadmaps show fusion reaching utility-scale deployment after 2035, while small modular reactors can be deployed in the 2030 to 2035 window.

Small modular reactors use established fission fuel and can be factory-built in modules. Several designs have reached the licensing review stage with the Nuclear Regulatory Commission. The NuScale VOYGR plant, for instance, received design certification in 2023. In contrast, fusion approaches, including tokamaks and pulsed systems, still require further gains in plasma confinement and materials durability. Commonwealth Fusion Systems aims to demonstrate net energy gain with its SPARC device, while Helion Energy has signed an agreement with Microsoft targeting electricity delivery by 2028, though that timeline remains uncertain.

Small modular reactors benefit from decades of operating experience with pressurized-water reactor technology, allowing regulators to review designs with reference plant data. Fusion systems must satisfy novel safety questions around tritium inventory, magnet quench behavior, and neutron-activated structural materials. The NuScale design uses passive cooling systems validated over 50 years of commercial operation, whereas fusion projects must demonstrate that superconducting magnets can survive repeated thermal and radiation cycling without degradation.

Steps to Understanding the Nuclear-AI Power Timeline

• Immediate Action (2026-2028): Microsoft's Three Mile Island restart and Amazon's existing nuclear power purchase agreements provide near-term capacity, with Three Mile Island expected to deliver 835 megawatts by 2028.
• Near-Term Deployment (2030-2035): Small modular reactors from vendors like NuScale and GE Hitachi will begin commercial deployment, with factory fabrication rates expected to reach one module every six months once supply chains mature.
• Long-Term Vision (2040+): Fusion energy companies continue development work, with most roadmaps showing commercial deployment after 2035, though some projects like Helion have made earlier commitments to Microsoft.

Cost trajectories differ significantly between the two approaches. Small modular reactor vendors now emphasize overnight capital costs below $4,000 per kilowatt by the fifth-of-a-kind unit, leveraging factory fabrication learning curves. Fusion teams project similar cost trajectories only after multiple demonstration plants have operated, with some roadmaps showing $3,000 per kilowatt achievable by 2050 under aggressive deployment assumptions.

The funding landscape reflects these different risk profiles. Small modular reactor companies like NuScale and GE Hitachi have secured government cost-sharing agreements under the U.S. Advanced Reactor Demonstration Program, relying on public funds to de-risk first deployments. Fusion, by contrast, attracts private venture capital betting on scientific breakthroughs that could unlock larger returns. Private equity funds focused on energy transition are now structuring hybrid vehicles that hold both SMR development rights and early-stage fusion minority stakes, allowing investors to hedge across both timelines.

U.S. data center electricity consumption is projected to rise from 4 percent of total grid load in 2023 to roughly 8 percent by 2030, driven primarily by large language model training runs. This surge has forced utilities and grid operators to model two parallel procurement tracks: one anchored by small modular reactors for the 2030 to 2035 window and a secondary fusion track for 2040 and beyond. Regional transmission organizations have begun publishing sensitivity cases that assume 10 to 15 gigawatts of new nuclear capacity co-located with hyperscale data center campuses.

The nuclear-AI partnership represents a fundamental shift in how the tech industry approaches infrastructure. Rather than waiting for the perfect long-term solution, hyperscalers are securing immediate capacity through reactor restarts while positioning themselves for next-generation technologies. This pragmatic approach acknowledges that artificial intelligence's power demands cannot wait for fusion breakthroughs, even if those breakthroughs could eventually provide cleaner, more abundant energy.