The Fastest New Nuclear Power Isn’t a New Reactor

In the first week of October, two separate announcements quietly redrew the fastest path to new nuclear power in the United States — and neither one involved building a new reactor. Google agreed to pay Constellation Energy more than $4.3 billion to “uprate” eleven existing reactors across six plants for an extra 890 megawatts, while the Department of Energy offered Vistra Corp. a conditional $4.2 billion loan to squeeze 433 MW more out of three aging units in Ohio and Pennsylvania. Combined, that’s roughly $8.5 billion committed in a single week to a technique the nuclear industry has used since 1977 but that AI-era power buyers are only now treating as a frontline procurement strategy. For anyone planning data center capacity against a grid that can’t keep up with demand, uprates deserve a harder look than they’ve gotten.

The deals: Google’s $4.3 billion bet and Washington’s $4.2 billion loan

Google’s agreement with Constellation is actually two deals stacked together. A 15-year contract covers 2,700 MW of power the company already generates in the PJM Interconnection, and a 20-year power purchase agreement layers on top of it for 890 MW of brand-new capacity created by uprating 11 reactors at six plants in Illinois, Pennsylvania, and New Jersey. Constellation expects to spend more than $4.3 billion on the physical upgrades, and the first new megawatts aren’t expected until 2028. Google says the deal brings the new U.S. nuclear capacity it has helped enable — counting uprates and reactor restarts together — past 1.5 gigawatts.

The DOE-Vistra deal, announced the same week through the department’s Office of Energy Dominance Financing, is a conditional loan commitment of up to $4.2 billion tied to upgrades at the Beaver Valley plant in Pennsylvania and the Davis-Besse and Perry plants in Ohio. Those projects would add 433 MW of new capacity while preserving almost 4 GW of existing baseload generation and extending plant operating licenses by 20 years. Much of that power is already spoken for: Vistra signed 20-year PPAs in January covering more than 2,600 MW from those same three plants to serve Meta’s PJM load, with purchases starting in late 2026 and additional capacity phasing in through 2034. Amazon has a separate 20-year PPA for up to 1,200 MW from Vistra’s Comanche Peak plant in Texas, which the new DOE financing could also help upgrade.

What an uprate actually buys you — and what it doesn’t

A nuclear uprate is a regulatory and engineering process, not a new power plant. The Nuclear Regulatory Commission has approved roughly 144 uprates since the first requests in 1977, adding more than 6,500 megawatts total — the rough equivalent of six new reactors, spread across five decades. They come in three sizes: measurement uncertainty recapture uprates add less than 2% of capacity by recalibrating how precisely a plant measures its own output, stretch uprates add 3% to 7% through moderate equipment changes, and extended power uprates can add up to roughly 20% but require replacing major components like turbines, generators, and feedwater systems. Most of a reactor’s uprate potential tops out around 20% of its original licensed capacity — this is optimization at the margins of plants already built, not a new source of baseload.

That matters for how buyers should read the “fastest path” framing now attached to this strategy. Uprates genuinely skip the decade-plus timeline of licensing and constructing a new reactor, and they avoid the unresolved supply chain and cost overruns that have dogged small modular reactor projects. But “faster than new-build” is a low bar — Google’s own capacity doesn’t arrive until 2028, two years from now, and Vistra’s CEO has said the real uprate push at its fleet is likely to land after 2028 as well. If your capacity planning horizon is 2026 or 2027, uprates don’t solve your problem; they solve 2029’s problem, assuming nothing slips.

The catch: taxpayer-backed loans, restart troubles, and an unclosed demand gap

Google has emphasized that its deal with Constellation puts no cost burden on other ratepayers — a claim worth scrutinizing given how much of this broader nuclear buildout runs through federal financing. The DOE’s $4.2 billion commitment to Vistra is a loan, and loans backed by the Office of Energy Dominance Financing are ultimately government-guaranteed exposure, meaning taxpayers carry risk if a project underperforms. Execution risk isn’t hypothetical, either: Holtec’s effort to restart the Palisades plant in Michigan — a separate but related piece of the “bring old nuclear back” trend — has been delayed by an NRC investigation into a tipped fuel assembly during refueling, and safety advocates continue to raise concerns about rehabilitating Three Mile Island Unit 1. Neither Google’s nor Vistra’s announcements included a side-by-side comparison of the new capacity against actual regional demand growth, and that omission is the real tell: 890 MW and 433 MW are meaningful numbers, but they are a rounding error against the 32 GW power shortfall Morgan Stanley has projected for U.S. data centers through 2028. Uprates are a genuine tool, not a fix for the structural gap between what the grid can deliver and what AI buildout wants to consume.

There’s also a vendor dynamic worth flagging for anyone negotiating similar deals: Constellation agreed to use Google’s own agentic AI tools for planning and managing the uprate construction work, tying a utility’s execution timeline to the same company that’s buying the power. That’s not necessarily bad, but it’s a dependency buyers should ask about when a hyperscaler counterparty is also supplying the project-management software for its own power source. This is the same pattern of entangled, circular arrangements that’s shown up elsewhere in AI infrastructure financing, and it deserves the same scrutiny before anyone treats a headline MW figure as a done deal.

What this means if you’re planning capacity now

Three practical takeaways. First, uprates are a real, additive source of firm, carbon-free power, and they’re worth asking about if you’re negotiating directly with a utility or a nuclear-adjacent IPP in a region like PJM, ERCOT, or the Midwest — but treat 2028-2029 as the realistic earliest delivery window, not next year. Second, if a counterparty’s power story leans on uprate or restart megawatts, ask for the plant-by-plant breakdown and the NRC filing status, not just the aggregate number; the gap between “capacity we’ve enabled” (Google’s 1.5 GW framing, which includes restarts) and “capacity physically online” can be large and easy to gloss over in a press release. Third, don’t let a flurry of nuclear deal headlines change your near-term risk posture — this news cycle is a signal that utilities and hyperscalers are getting more serious about firm power, which is good for the market over a five-year horizon, but it does nothing to close the gap buyers are underwriting against for 2026 and 2027 capacity. Our earlier look at Big Tech’s broader nuclear bet still holds: the question isn’t whether nuclear power is coming back for data centers, it’s whether it arrives on your timeline or two cycles after it.

Uprates are the cheapest, fastest incremental megawatts the U.S. nuclear fleet can offer right now, and the October deals prove hyperscalers are willing to pay a premium to lock them in early. But “fastest” in nuclear still means years, not quarters, and the deals announced this month cover a small fraction of the power gap buyers are actually facing. Model your power contracts, interconnection timelines, and site-selection decisions against the grid you’ll actually have in 2027 and 2028 — not the one a press release describes. Start with our Data Center Buyer’s Toolkit to pressure-test your own power assumptions before you sign.

Written by

Raajeev Ratra

Data Center Infrastructure Expert | 15+ Years in DC Design, Operations & Project Management

Raajeev is a seasoned data center professional with hands-on experience in hyperscale facilities, colocation design, power & cooling infrastructure, and global DC operations. He shares practical insights to help engineers and IT leaders build better infrastructure.

Connect on LinkedIn →