Analysis
Google signed a 20-year power purchase agreement with Constellation Energy for up to 3,590 megawatts of electricity, including roughly 890 megawatts of new nuclear capacity, The Register reported Wednesday. Constellation, which the outlet calls the largest purveyor of atomic energy in the US, will invest $4.3 billion to uprate at least 11 existing reactors across Illinois, Pennsylvania and New Jersey — modernizing turbines, steam generators and digital control systems rather than building new plants. The first uprated reactor is expected online in 2028.
Alphabet's Google, founded in 1998 by Larry Page and Sergey Brin, has spent the past three years converting its cloud and search-advertising profits into one of the largest AI infrastructure buildouts anywhere, with 2026 datacenter and compute capex alone set to hit up to $205 billion. That scale of spending runs into a hard physical limit: the US grid cannot add gigawatts of new generation as fast as hyperscalers can pour concrete for data centers.
Competing for the same reactors
Google is not alone in this race. Microsoft and Meta have each pursued their own nuclear power-purchase agreements over the past two years to secure baseload electricity for AI compute, and Amazon has made similar moves through its own utility partnerships. The Constellation deal puts Google in direct competition with its own hyperscaler peers for a finite set of US reactors capable of being uprated — there are only so many plants in Illinois, Pennsylvania and New Jersey that qualify, and Microsoft and Meta have already been shopping the same inventory.
The deal size is not out of step with the market: the San Francisco Fed's own president has warned that AI-driven electricity demand could extend the current energy-price shock, a dynamic Axios reported this week is starting to show up in regional inflation data. Power markets, not GPU allocations, are now the scarce resource in the AI buildout — a theme that also shows up in Pulse's coverage of Type One Energy's $200 million fusion raise this week, aimed at the same underlying shortage a decade out.
What the deal doesn't solve is the timeline mismatch. Google is spending at a 2026 run rate, but the reactors it just locked up won't deliver incremental power until 2028 at the earliest — and only after Constellation completes a multi-year uprating process across a dozen sites. If AI compute demand plateaus before then, Google will have committed to two decades of above-market power purchases for capacity it no longer needs at the scale it planned; if demand keeps accelerating, the 2028 timeline means data centers sit power-constrained for at least another two years regardless of how much capital Google has available.
For hyperscaler-adjacent investors, the read is that power procurement is now a competitive moat alongside chip allocation — and a slower-moving one. A startup or REIT that can originate uprate-able nuclear capacity, or firm long-duration storage, has a negotiating position against Google, Microsoft and Meta simultaneously that didn't really exist two years ago.