The reactor is a purchase orderThe reactor is a purchase orderThe reactor is a purchase orderThe reactor is a purchase orderThe reactor is a purchase orderThe reactor is a purchase orderThe reactor is a purchase order

The nuclear pivot is real and the timeline is the entire story. Announced hyperscaler agreements now total somewhere between 10 and 13 GW across more than a dozen deals, split roughly evenly between power purchase agreements (PPAs — long-term contracts to buy output) and direct development partnerships. The marquee ones:

• Meta: up to 6.6 GW in the PJM market announced on 9 January, across Vistra, TerraPower and Oklo — including more than 2.1 GW from Vistra’s Beaver Valley, Perry and Davis-Besse plants.
• Amazon: an expanded offtake with Talen for close to 2 GW from Susquehanna in Pennsylvania, running through 2042.
• Microsoft: a $16 billion, 20-year agreement restarting Three Mile Island Unit 1, rebranded the Crane Clean Energy Center.
• Google: the first corporate agreement to develop a fleet of small modular reactors (SMRs) in the US — up to 500 MW across six or seven units with Kairos Power, first reactor targeted for 2030.

Notice the split, because it is the whole point. Most of that capacity is existing or mothballed plants being contracted, which is a financing event rather than a generation event: the electrons already existed and now have a named buyer at a guaranteed price. The genuinely new capacity — the SMRs — is not expected to shift the US data centre mix until somewhere between 2030 and 2035.

Why nuclear, specifically

The technical case is one number: US nuclear plants averaged a 92% capacity factor in 2024. Nothing else grid-scale is close. An AI training cluster is close to a constant load and a terrible match for intermittent supply without storage nobody has built yet. If you need firm, clean, gigawatt-scale power for twenty years, the list of options is genuinely short.

Which is why the bridge is gas

While the reactors are being licensed, the actual megawatts are coming from turbines. In January an alliance was announced to deploy 2 GW of dedicated power, primarily natural gas, for hyperscale AI infrastructure.

There is a sharp engineering footnote here that deserves more attention than it gets: gas turbines have been reported failing under AI workloads. A training run is not a smooth industrial load — it ramps hard, drops hard, and does it on the scheduler’s clock. Turbines optimised for steady baseload are being cycled in ways their duty models did not anticipate.

The honest read

The nuclear announcements are doing two jobs. One is procurement. The other is narrative cover for a decade of gas, because “we signed a 20-year clean power agreement” is a better headline than “we are burning methane until 2032”.

Both things are true. The agreements are real, the intent is probably sincere, and the carbon between now and the reactors coming online is also real and will be emitted regardless.

Watch what gets built, not what gets signed. The purchase order and the power plant are separated by a decade, and the load arrives next year.