- Google’s €13 billion investment in Finland includes a 22-year power purchase agreement with Fortum covering up to 50% of the Loviisa nuclear plant’s output from 2030 to 2049, making it Google’s first nuclear PPA in Europe.
- The deal gives Fortum the revenue certainty to proceed with a €1 billion life-extension program for Loviisa, roughly €755 million of which previously lacked a final investment decision.
- Microsoft, Amazon and Meta have signed comparable long-term nuclear agreements, with Amazon paying approximately $650 million for a campus connected directly to the Susquehanna plant, a sign that grid bypass, not just clean power, is now a primary procurement objective.
A 22-year power purchase agreement between Google and Finnish energy company Fortum, disclosed this week as part of a €13 billion Finland investment, will fund the life extension of the Loviisa nuclear plant through 2050. It is Google’s first nuclear PPA in Europe and its first deal anywhere structured around keeping an existing plant running rather than building a new one. The underlying logic is straightforward: AI data centres need power that never stops, and most grids cannot supply it fast enough.
Why the Grid Cannot Keep Up
The International Energy Agency estimates global data centre electricity consumption could reach 945 TWh by 2030, roughly double current levels and close to 3% of total global electricity use. AI-optimised server racks are a large part of the reason: they can draw 30 kW to over 100 kW per rack, compared with 5 to 15 kW for conventional server racks. The gap between what AI infrastructure demands and what local grids can deliver is already visible in wholesale electricity prices, which have risen sharply near major U.S. data centre clusters.
Grid interconnection queues make the problem worse. Large projects in the U.S. routinely wait more than three years for connection approval, long enough to delay entire data centre buildouts. For hyperscalers running AI training clusters at close to full utilisation around the clock, that kind of uncertainty is operationally unacceptable.
Nuclear’s Specific Appeal
Wind and solar are cheap, but they are intermittent. A training cluster running a frontier model cannot pause because the wind dropped. Nuclear plants deliver firm, carbon-free baseload power regardless of weather, which is why AI operators are willing to pay a premium for it. That premium, locked in over decades through a PPA, is also what makes older or marginal nuclear assets economically viable again, the plants get a long-term revenue stream; the data centre operator gets guaranteed power. Both sides are buying certainty.
Small modular reactors add another dimension. Their compact footprint allows placement closer to data centre campuses, cutting transmission losses and reducing dependence on overloaded public infrastructure. OpenAI CEO Sam Altman has publicly backed nuclear investment as a necessary part of scaling AI, according to public statements, and Amazon has committed to developing 12 SMRs with X-energy and Energy Northwest targeting a combined output close to 1 GW.
How the Deals Are Structured
Google’s Fortum agreement is the clearest example of how these arrangements work in practice. The 22-year PPA covers up to 50% of Loviisa’s output from 2030 to 2049. That contracted revenue allows Fortum to proceed with a €1 billion life-extension investment programme, roughly €755 million of which had no final investment decision before this deal. The agreement also supports an additional 10 MW capacity increase at Loviisa, beyond a 38 MW increase already planned for 2028.
Microsoft took a similar route in 2024, signing a 20-year PPA with Constellation Energy to restart Unit 1 of the Three Mile Island plant in Pennsylvania. The 835 MW unit is scheduled to resume generation in 2027. Amazon went further: it acquired the Cumulus data centre campus adjacent to Talen Energy’s Susquehanna nuclear plant for approximately $650 million in March 2024, giving AWS a direct physical connection to the plant and bypassing the public grid entirely for a significant share of its load. Oklo has also secured a 1.2 GW agreement with Meta for advanced reactor capacity, according to reports. The U.S. Department of Energy has added public financing to the mix, committing $1.9 billion to NextEra Energy for restarting Iowa’s Duane Arnold plant, with Google as the anchor customer through a 25-year PPA.
Power as the New Constraint
Chip supply and model capability get most of the attention in AI infrastructure coverage, but the binding constraint for the next phase of buildout is power. The U.S. data centre sector currently consumes around 180 TWh annually; projections suggest that could reach 400 to 600 TWh by 2030, with high-end scenarios going further still. Co-locating directly beside power plants, as Amazon has done at Susquehanna, is one response. Long-term PPAs that fund life extensions, as Google has done at Loviisa, are another. Both strategies amount to the same thing: hyperscalers moving upstream into energy infrastructure because the alternative, waiting for public grid capacity to catch up, is not compatible with their build schedules.
The pattern across Google, Microsoft, Amazon and Meta suggests this is no longer a niche procurement strategy. Whether it extends to new-build reactors at scale, rather than life extensions and restarts, depends on how quickly SMR developers can move from commitments to operational plants, a timeline that remains measured in years, not quarters. For more on how AI infrastructure demands are reshaping silicon and power strategies, visit our AI Hardware section.
