The AI infrastructure race has reached its next bottleneck: electricity. Reuters reported on April 10 that major tech companies including Microsoft, Google, Amazon, and Meta are putting substantial financial weight behind nuclear energy projects, signing long-term power purchase agreements that provide nuclear startups with capital and commercial credibility. For years, nuclear projects struggled to find buyers willing to commit to 20-year contracts. AI data centres, with their insatiable and predictable demand for power, have changed this calculation entirely. Energy procurement is becoming a strategic competitive advantage — and Big Tech is moving to lock it in.
Direct Answer: Why are Microsoft, Google, and Amazon investing in nuclear energy? AI data centres require enormous, continuous electricity — and conventional grid power is insufficient, unreliable, or too expensive at the scale required. Microsoft, Google, Amazon, and Meta are signing long-term power purchase agreements with nuclear energy companies to secure dedicated, carbon-free power for their AI data centre buildouts. Nuclear offers 24/7 baseload power (unlike solar and wind which are intermittent), very high energy density (small land footprint relative to output), and — critically — long-term price certainty via 20-30 year contracts. The Iran war has raised energy market uncertainty further. AI’s electricity demand is projected to grow 160% by 2030 (Goldman Sachs). Companies that lock in dedicated power supplies now gain a structural infrastructure advantage that cannot be easily replicated.
The Scale of the Problem
To understand why Big Tech is turning to nuclear, start with the numbers.
AI electricity consumption:
- US AI data centres consumed over 10% of total US electricity in Q1 2026 — up from approximately 4% in 2023
- Goldman Sachs projects AI data centre electricity demand will grow 160% by 2030
- A single large-scale AI data centre running 50,000+ Nvidia H100 GPUs consumes approximately 150–200 megawatts continuously — equivalent to powering 150,000 average US homes
- TSMC’s planned Arizona fabrication plants (manufacturing AI chips) will each require 1+ gigawatts of dedicated power
The grid capacity problem: US electricity grid infrastructure was not designed for concentrated, dense, always-on loads of this scale. Data centres in northern Virginia (the world’s largest data centre hub) have already triggered grid capacity warnings. Utilities in data-centre-heavy regions are reporting that new connection requests are backed up 5–10 years. The existing grid cannot absorb the AI infrastructure buildout on the timelines AI companies require.
The conventional power problem: Solar and wind are intermittent — AI data centres cannot run on power that is only available when the sun shines or the wind blows without massive battery storage (which introduces its own costs and supply chain constraints). Natural gas is reliable but carbon-intensive and subject to price volatility — exactly the volatility the Iran war has amplified.
Nuclear addresses all of these problems simultaneously: 24/7 baseload, very low carbon, predictable long-term pricing via contract, and high energy density.
The Deals Being Signed
The template was established in 2024 with Microsoft’s landmark agreement to restart Three Mile Island’s Unit 1 reactor (decommissioned since 2019) specifically to power Microsoft data centres in Pennsylvania. That deal — announced September 2023, operational 2024 — provided the proof of concept that Big Tech could structure and execute nuclear power agreements.
The 2026 wave:
Google — Kairos Power: Multi-reactor contract with Kairos Power for molten salt fluoride small modular reactors (SMRs). Kairos’s KP-FHR design operates at atmospheric pressure (safer than traditional pressurised water reactors) and can be factory-manufactured in smaller modules. Google’s contract provides Kairos with the offtake agreement needed to secure project financing for its first commercial reactors.
Amazon — X-energy: Investment in X-energy’s Xe-100 high-temperature gas-cooled reactor. X-energy is building a commercial plant in Washington state. Amazon’s participation provides both capital and a committed power buyer — the two things nuclear projects need that they have historically struggled to find simultaneously.
Microsoft — Helion Energy and others: Beyond Three Mile Island, Microsoft has invested in Helion Energy (fusion) and is in discussions with multiple SMR developers for additional contracted power. Microsoft’s energy team is now one of the most active nuclear power procurement operations outside of utilities.
Meta — Undisclosed partners: Reuters identified Meta as among the companies actively pursuing nuclear deals, though specific partnerships have not been publicly named as of April 12.
Amazon — Nuclear-powered data centre announcement: AWS announced in early 2026 that it is building a data centre campus with dedicated nuclear power in the mid-Atlantic United States, the first AI data centre designed from the ground up around nuclear power supply.
Why Nuclear Over Other Alternatives
The energy alternatives to nuclear each have specific limitations that make them less suitable for AI data centre baseload:
Solar: Intermittent. Requires battery storage for continuous operation. Land-intensive. Costs have fallen dramatically but storage costs offset the advantage for 24/7 industrial loads. Good supplement; not suitable as primary power for data centres without enormous storage investment.
Wind: Same intermittency problem as solar. Offshore wind is more consistent but expensive to build and maintain. Transmission from wind-rich regions (Midwest, offshore) to data centre clusters (northern Virginia, the Bay Area) adds cost and grid complexity.
Natural gas: Reliable and dispatchable, but carbon-intensive (creating ESG reporting problems for companies with net-zero commitments) and subject to price volatility. The Iran war’s impact on energy prices has made gas contracts riskier in 2026.
Geothermal: Extremely limited by geography. Iceland, Iceland-adjacent regions, and specific US locations (California, Nevada, parts of the Pacific Northwest) have viable geothermal. Cannot be deployed near most existing data centre clusters.
Nuclear (existing plants): Reliable, low-carbon, predictable long-term cost via contract. Limited supply of existing plants — many are nearing end of life, and new plant approvals have been limited in the US since Three Mile Island (the 1979 accident) and Chernobyl. The restart of Three Mile Island was a political and regulatory achievement, not just a commercial one.
Nuclear (SMRs): Small modular reactors represent the most interesting opportunity — factory-manufactured, smaller footprint, designed to be deployed near load centres rather than in remote locations. Still pre-commercial (no SMR has been built at commercial scale in the US as of April 2026), but the tech companies’ long-term contracts are the financing unlock that SMR developers need to build first-of-kind plants.
The Strategic Moat: Energy as Infrastructure
The deeper significance of Big Tech’s nuclear deals goes beyond sustainability. Energy access is becoming a strategic competitive moat alongside chip access.
The structural advantage of long-term contracts: Nuclear power purchase agreements typically run 20–30 years. A company that signs a 25-year contract for 500 megawatts of nuclear power in 2026 has secured a fundamental AI infrastructure input — electricity — through 2051. New entrants to the AI infrastructure market in 2030 or 2035 will need to compete for the same power, which will be scarcer and more expensive if grid capacity has not expanded proportionally.
The compounding timeline advantage: Nuclear plants take 5–15 years to build (depending on technology and regulatory environment). Companies that signed contracts in 2024–2026 are initiating construction now. Plants will begin operating in 2030–2035. The AI infrastructure advantage from this energy supply will be most acute precisely when frontier AI compute demands are highest.
The parallel with chip supply: The Broadcom-Google-Anthropic chip deal and the Samsung/TSMC capacity commitments represent the same strategic logic applied to silicon. Lock in the critical inputs — chips, energy — before competitors can, at terms that reflect today’s pricing rather than future scarcity pricing.
For startups and smaller AI companies: The energy deals being signed now by Microsoft, Google, Amazon, and Meta create a structural barrier that goes beyond capital. You cannot replicate a 25-year nuclear power contract quickly. The AI infrastructure hierarchy — frontier models, custom chips, dedicated energy — is being consolidated by the same small group of hyperscalers.
The Nuclear Safety and Waste Context
Nuclear energy’s benefits come with well-documented trade-offs that any honest assessment must include.
Safety: Modern reactor designs (SMRs in particular) are significantly safer than the plants involved in Three Mile Island (1979) and Chernobyl (1986). Kairos’s molten salt design operates at atmospheric pressure — the reactor cannot overpressurise, eliminating the explosion risk that damaged Chernobyl. Xe-100 uses TRISO fuel particles that are physically incapable of melting down. Modern SMRs are designed for passive safety — cooling without external power or operator intervention.
Waste: Nuclear waste remains the unresolved challenge. High-level radioactive waste (spent fuel) must be stored for thousands of years. The US has no operational permanent storage facility (Yucca Mountain was politically blocked). Spent fuel currently sits in dry cask storage at reactor sites — a temporary solution that has become permanent by default. None of the current AI-nuclear deals resolve this; they add to the waste that future generations must manage.
Cost: Nuclear remains the most expensive form of new electricity generation by upfront capital cost. The economics only work with long-term contracts (which provide financing certainty), government support (US loan guarantees, production tax credits), and at scale. First-of-kind SMRs will be expensive; nth-of-kind SMRs may be cost-competitive with gas peakers. The first movers are paying to establish the supply chain.
The Sovereignty Dimension
For readers who think about data sovereignty and infrastructure independence, Big Tech’s nuclear energy strategy raises a specific concern: the energy layer of AI infrastructure is being locked up by the same companies that control the model and chip layers.
A country or organisation that wants to build sovereign AI infrastructure — running its own models on its own hardware — faces a stacking set of dependencies:
- Chips: Manufactured by TSMC (90% of advanced nodes), sourced via Nvidia, AMD, or Google/Amazon custom silicon
- Energy: Increasingly contracted long-term by Microsoft, Google, Amazon, Meta
- Models: Developed by OpenAI, Anthropic, Google, Meta
- Infrastructure: Hosted on AWS, Azure, Google Cloud
Nuclear energy deals make the second layer harder for independent players to access. For European governments trying to build sovereign AI infrastructure, for India’s AI mission, for any organisation that does not want full dependence on US hyperscalers — energy supply is now part of the problem.
The practical response: small modular reactors, if they achieve commercial deployment and reasonable cost by the mid-2030s, may democratise nuclear power access. Until then, sovereign AI infrastructure at scale requires contracting energy independently — which means countries and large organisations need to run their own energy procurement, not rely on hyperscaler-negotiated contracts.
FAQ
Why can’t AI data centres just use solar and wind? Solar and wind are intermittent — power output varies with weather. AI data centres run 24/7 at consistent load, requiring baseload power. The cost of battery storage sufficient to smooth solar/wind intermittency for a 200-megawatt data centre is currently prohibitive. Nuclear provides 24/7 power without storage.
What is a small modular reactor (SMR)? An SMR is a nuclear reactor smaller than traditional large-scale plants (typically under 300 MW vs 1,000+ MW for conventional plants), designed to be factory-manufactured and assembled on-site. Smaller size means lower upfront capital cost, faster construction, and more deployment flexibility. No commercial SMR has been built at scale in the US yet — the tech companies’ deals are funding the first commercial plants.
Is nuclear energy clean? Nuclear is low-carbon during operation — emissions comparable to wind and solar on a lifecycle basis. The waste issue (radioactive spent fuel requiring thousands of years of storage) is unresolved and represents a genuine long-term environmental liability. Tech companies count nuclear as clean energy for ESG reporting purposes; environmentalists debate this.
How much electricity does an AI data centre use? A large-scale AI data centre running 50,000 Nvidia H100 GPUs consumes approximately 150–200 megawatts continuously — enough to power 150,000 average US homes. TSMC’s planned Arizona chip fabrication plants each require 1+ gigawatts. The total AI data centre electricity demand in the US exceeds 10% of national consumption as of Q1 2026.
Does this affect electricity prices for consumers? Potentially yes, in data centre-heavy regions. Data centres competing with residential and commercial users for grid capacity can contribute to electricity price increases and grid stress. Northern Virginia, where much of the world’s data centre capacity is concentrated, has already seen grid capacity warnings. Nuclear deals that add dedicated power supply — not drawing from the shared grid — are designed to mitigate this.
Related Articles
- TSMC Q1 2026: $35.7B Record Revenue — The AI Chip Chokepoint That Controls Everything
- OpenAI Pauses Stargate UK: Energy Costs and Regulation Kill £Billion AI Data Centre
- Amazon’s Chip Business Is Worth $50 Billion — Jassy Threatens Nvidia
Sources & Further Reading
- MIT Technology Review — AI Section — In-depth coverage of AI research and industry trends
- arXiv AI Papers — Pre-print research papers on AI and machine learning
- EFF on AI — Civil liberties perspective on AI policy
