Seven years after emerging from bankruptcy, Westinghouse Electric has become the anchor of America's nuclear revival, backed by an $80 billion government contract and a new AI platform built with Google Cloud to make reactor construction faster and cheaper than it has ever been.
The turnaround is striking enough to deserve a moment. In 2017, Westinghouse filed for bankruptcy under the weight of catastrophic cost overruns building four AP1000 reactors in Georgia and South Carolina. The South Carolina project was abandoned outright. The two Vogtle reactors in Georgia eventually came online in 2023 and 2024, but they arrived seven years late and $18 billion over budget. That was supposed to be the lesson that killed the large-reactor business in America. It wasn't.
This month, the US Department of Energy conditionally committed $17.5 billion in loan facilities to finance the long-lead equipment for 10 new Westinghouse AP1000 reactors, part of an $80 billion agreement signed with Westinghouse, Brookfield Asset Management, and Canadian uranium producer Cameco. The goal, as reported by ANS Nuclear Newswire, is to have all 10 reactors under construction by 2030. In exchange for subsidizing the financing, the federal government will claim 20 percent of cash distributions or profits above the $17.5 billion threshold once profitability conditions are met. According to Brookfield, the DOE loan package could pull construction and commercial operation timelines forward by as much as three years.
The direct cause of all this is AI. Data center electricity consumption is on track to hit 1,000 TWh in 2026, roughly equivalent to Japan's annual usage, and it's growing at 15 to 20 percent annually. The North American Electric Reliability Corporation has already warned of elevated risk of summer electricity shortfalls across all three US grid regions. Wind and solar can't solve this. They're intermittent. Data centers run around the clock and need firm, dispatchable power that doesn't disappear when the sun sets or the wind drops. Nuclear is the only carbon-light baseload source that can scale to meet that kind of demand. The $80 billion contract is, in practical terms, the grid's answer to what GPUs need.
The AI platform built to fix what sank Vogtle #
Separate from the government deal but directly connected to it, Google Cloud announced a partnership with Westinghouse to build a custom AI platform for reactor construction. It combines Westinghouse's WNEXUS digital twin of the AP1000 with Google Cloud's Vertex AI, Gemini, and BigQuery. The training base: 75 years of Westinghouse nuclear engineering data. Early proof-of-concept testing showed the system can automatically interpret design models, generate thousands of construction work packages, and simulate real-world disruptions. Then re-optimize sequencing in minutes. That last part matters: one of the core reasons Vogtle ran so far over budget was schedule chaos, with delays in one section cascading into delays everywhere else. An AI system that can re-sequence thousands of interdependent construction tasks on the fly is a direct attack on the root cause of the last failure.
It's worth being clear-eyed about what's promised versus proven. Early pilot testing shows cost and time savings, but no AP1000 has ever been built on schedule in the United States. The Vogtle experience is the only domestic data point, and it's a brutal one. The optimism around the new programme rests heavily on the argument that the first build is always the hardest, that modular construction techniques have matured, and that the AI tooling will do what it couldn't do before. Those are plausible arguments. They are not guarantees.
What the big-tech pivot to nuclear actually means #
Still, the structural logic is hard to argue with. The IEA projects that data centers, AI workloads, and crypto mining together could double their share of global electricity consumption from roughly 2 percent in 2022 to 4 percent by 2026. Nuclear is projected to supply at least 5 GW of dedicated data center capacity by 2030, but that figure could look small quickly if AI compute demand keeps compounding the way it has. Microsoft has already locked in a 20-year power purchase agreement with Constellation Energy for the restarted Three Mile Island Unit 1, covering 835 MW of output. The big-tech pivot to nuclear is not theoretical at this point. It's contracted.
Westinghouse's position in all of this is genuinely unusual. It's not a newcomer with a promising SMR design and a funding deck. It's a company with a commercially proven reactor, a 75-year engineering base, and now an $80 billion government backstop. The AP1000 is the design that's already running at Vogtle. The question isn't whether the technology works. The question is whether the construction execution has actually improved - or whether the industry is about to repeat the same mistakes with more money behind them. If the Google Cloud AI platform does what Westinghouse says it can do, the answer could be different this time. The 2030 construction start date will be the first real test.
Also read: Congress moves to make AI model distillation a sanctionable offense as Chinese labs face theft accusations • AMD bets $5 billion on Anthropic and gets tens of billions in chip orders back • IBM's CEO says the mainframe isn't dying but the numbers are doing him no favors