AI and Nuclear — Microsoft Reopens Three Mile Island to Power Data Centers

Cos'è: An analysis of the link between explosive AI data center growth and the nuclear renaissance as an energy source: from big tech Power Purchase Agreements to Small Modular Reactors, through to the environmental paradox of an AI industry that calls itself green.

The Microsoft-Constellation Energy Agreement: Three Mile Island Comes Back to Life

In September 2024 Microsoft and Constellation Energy announced a twenty-year Power Purchase Agreement (PPA) providing for the reopening of Unit 1 at the Three Mile Island nuclear plant in Pennsylvania — the same plant that became a symbol of the most serious commercial nuclear accident in America in 1979, though Unit 1 was not involved in that incident and had operated safely until its economic closure in 2019. Unit 1, with a capacity of 835 megawatts, will supply power exclusively to Microsoft data centers through the PJM Interconnection grid. The commercial name of the project is 'Crane Clean Energy Center'. The cost of the agreement was not made public, but Constellation estimated approximately $1.6 billion in investments needed to bring the plant back to full operation by 2028. This is an unambiguous signal: the demand for 24/7 carbon-free energy from the AI industry is intense enough to make it economically rational to reopen nuclear plants closed for market reasons, not safety reasons.

Amazon and Google: A Parallel Race to SMRs and Existing Nuclear

Microsoft is not alone. Amazon Web Services announced two separate nuclear deals in 2024: an investment in X-energy, a startup specializing in fourth-generation Small Modular Reactors (SMR) with helium gas-cooled reactors, and an agreement with Talen Energy to purchase power from the Susquehanna nuclear plant, also in Pennsylvania, to power a 960 MW data center campus directly adjacent to the facility. This latter agreement was initially blocked by the Federal Energy Regulatory Commission (FERC) over concerns about the impact on the regional electricity grid, a signal of the regulatory problems accompanying this transition. Google instead chose the SMR route by signing an agreement with Kairos Power, a California startup developing molten salt-cooled modular reactors (fluoride salt-cooled), with the goal of having the first reactors operational by 2030 and aggregate capacity of 500 MW by 2035. The choice of different technological architectures by the three hyperscalers reflects uncertainty about which SMR supply chain will first reach commercial maturity.

Why Renewables Are Not Enough: The 24/7 Power Problem

The logic behind the return to nuclear is technical before ideological. Large-scale AI data centers require continuous, stable, high-density power. A 500-1000 MW campus hosting clusters of H100 or B100 GPUs cannot tolerate interruptions of a few seconds without losing training jobs that take weeks. Solar and wind are intermittent sources by definition: they produce when there is sun or wind, not when there is demand. Integration with battery storage systems (BESS) resolves hourly variability but not seasonality or prolonged periods of low production. Nuclear, conversely, offers a capacity factor above 90% — it works practically all the time, regardless of weather conditions — with CO₂ equivalent emissions per kWh produced comparable to offshore wind. For big tech companies with carbon neutrality commitments (Microsoft has committed to being carbon negative by 2030), purchasing nuclear energy through PPAs is the only way to sustain data center growth without exploding the carbon footprint. Renewables remain essential in the energy mix, but cannot be the sole source for continuous critical loads.

IEA and Doubling Consumption: The Environmental Paradox of AI

The International Energy Agency (IEA) published a report in January 2024 estimating the potential doubling of global data center electricity consumption by 2026, rising from approximately 200 TWh annually to 400-800 TWh. For context: the entire Italian electricity grid consumes approximately 300 TWh per year. This estimate includes the contribution of generative AI, which has significantly higher energy consumption per query than a traditional web search: according to some estimates, a ChatGPT query consumes approximately 10 times the energy of a standard Google search. The paradox is explicit: the AI industry positions itself as a tool for solving climate problems — power grid optimization, accelerating battery materials research, reducing industrial waste — while its own operation requires unprecedented energy infrastructure at scale. Microsoft, Google and Amazon all have carbon neutrality commitments that become progressively harder to maintain as AI workloads grow faster than renewable or nuclear energy procurement.

SMR as Future Solution vs Existing Nuclear as Immediate Solution

Small Modular Reactors represent the long-term technological promise: reactors with power below 300 MW (versus 1000-1600 MW for conventional reactors), factory-buildable with reduced timelines, placeable at smaller sites and potentially less expensive through industrial economies of scale. However no commercial SMR is yet operational in the West in 2024: NuScale, the most advanced project in the USA, cancelled its first commercial plant (the Carbon Free Power Project in Idaho) in November 2023 due to out-of-control costs — $89 per MWh projected, more than double the initial estimates. Rolls-Royce in the UK, X-energy and Kairos in the USA are all in development phases with a 2030-2035 horizon. Existing nuclear, conversely, is the immediate solution: already built, licensed and amortized plants that can be reactivated or operated under PPAs at relatively contained marginal costs. The time window in which AI demand will be met primarily by existing nuclear (2024-2032) and that in which next-generation nuclear will come into play (2032+) define two very different energy scenarios for the AI industry.

---