Google’s nuclear deal has two power numbers. Only one is new capacity

Google and Constellation’s 6 October agreement is an unusually useful case study in reading an AI energy announcement. It combines 890 megawatts of planned additional nuclear capacity with a separate 2,700 MW arrangement involving existing generation. Adding them produces a large procurement number. It does not produce the amount of new power being built.[1]

That distinction is not an argument against buying existing nuclear electricity. Keeping a plant financially viable and increasing its output can both matter. They answer different questions, however: what electricity is contracted, what supply is preserved and what physical capacity is added? A credible assessment keeps all three visible.

A nuclear uprate is an engineering project, not a new label

Google describes the work as upgrades at six operating nuclear plants. Using existing sites can make this a different proposition from constructing an entirely new facility. The relevant advantage is making more use of an established generating system, not obtaining unlimited additional output merely by signing a longer contract.[2]

The Nuclear Regulatory Commission distinguishes measurement-uncertainty recapture uprates, typically below 2%, stretch uprates generally up to 7%, and larger extended uprates that can require major equipment changes. Those categories concern plant-specific engineering and licensing; they are not a menu from which a buyer can select an arbitrary percentage increase.[3]

The NRC describes technical review of uprate applications before approval. That gives the commercial promise a second, independent chain of evidence: analyses, approvals and physical modifications. A purchase agreement can support investment, but it cannot substitute for the steps required to operate safely at the resulting output.[4]

For an AI company, this is a useful constraint rather than a minor footnote. A datacentre opening date and a generation-upgrade date are separate schedules. Where they do not align, the company must explain what supplies the intervening demand. A future clean-power asset does not travel backwards in time to serve an earlier load.

What 890 MW could produce—and what that calculation cannot tell us

Here is an illustrative calculation, not a forecast for these projects. Suppose the full 890 MW increment operates for a non-leap year at a 90% capacity factor. Multiplying 890 MW by 8,760 hours and 0.90 gives 7,016,760 MWh, or approximately 7.02 terawatt-hours. The assumed 90% is a scenario input, not a disclosed guarantee for the uprates.

Annual energy is a useful sizing quantity, but it does not establish supply in a particular hour. Two resources can deliver the same yearly total with very different availability during a stressed evening. Nor does electricity entering a regional network establish unconstrained delivery to a particular datacentre. Capacity, energy, location and timing all have to match the claim being made.

The arithmetic also shows why “enough to power a certain number of homes” is an incomplete comparison. Such a statement needs an assumed household consumption level and says little about the datacentre’s own hourly demand, backup arrangements or network costs. Reporting the physical quantities is clearer than asking a household analogy to carry the entire story.

Additionality needs a counterfactual, not just a contract

Our interpretation is that the new-capacity component has a more direct physical claim than the existing-supply component: an uprate can be checked against a previous plant output limit. Even then, assessing what the buyer caused requires asking what would have happened without that buyer. Was the upgrade already financed? Would it have happened later, at a smaller scale or not at all? The announcement alone does not settle that counterfactual.

The existing-fleet arrangement deserves a different test. If commercial certainty helps preserve otherwise threatened generation, the relevant comparison concerns retirement or reduced operation. That can be consequential, but it should be demonstrated with evidence about the asset rather than counted automatically as newly constructed capacity.

There is a distributional question too. A privately financed generation project may still interact with shared transmission, balancing and reliability costs. The appropriate question for neighbouring customers is not simply who signed the power contract. It is which costs are assigned to the new load, which benefits reach the wider system and which assumptions underpin those allocations.

The next announcement worth watching is a commissioning milestone

A useful public progress record would name each upgrade, show its licensed and operating baseline, identify the incremental output, and distinguish approval, equipment installation and commercial operation. Delays should remain visible. That would let readers observe the conversion of a financial commitment into physical supply rather than repeatedly counting the same planned megawatts.

For AI infrastructure, this is the more substantive story than a race for the biggest headline number. The agreement addresses a real lever: extracting additional output from an established fleet. Its significance will be earned as the new capacity becomes available, at the places and times where demand is growing—not when unlike quantities are added together.

Report an error or suggest a correction