Westinghouse and Google Cloud Put a 3D Digital Twin at the Center of Nuclear Construction

The number that frames the whole project: 60%

The announcement names a specific target. Construction has historically accounted for 60% of a reactor's cost, and until recently that construction was managed with spreadsheets and paper documents. Delays in one task cascaded across thousands of interdependent ones.

That is the problem the Google Cloud partnership is pointed at. The joint system combines both companies' models and prediction tools with WNEXUS, Westinghouse's 3D digital twin of its reactors. Fed current and historical data, it is meant to predict bottlenecks, resequence construction tasks, adjust staffing, and account for supply chain constraints.

This is a concrete claim about where the money and the delay actually sit — not a general statement that AI will help. The stated goal it serves is having 10 AP1000 reactors under construction by 2030, enough combined power for 7.5 million households.

Westinghouse brought its own foundation to the table

The detail that stands out is that Westinghouse did not arrive empty-handed. Before the partnership, it had built Hive, proprietary AI infrastructure designed for nuclear's regulatory and export-control requirements, and Bertha, a generative assistant with access to 75 years of nuclear documentation.

The source states Google engineers were surprised that a 140-year-old company had already assembled the exact foundation needed to deploy AI securely in a heavily regulated environment. That framing matters: the constraint in nuclear is not model quality, it is whether a system can operate inside export controls and regulatory review at all.

In practice, Westinghouse supplied the compliance-ready substrate and the domain data; Google Cloud supplied models and prediction tooling. The digital twin is the connective layer that turns decades of documents into something a model can reason over.

The 'technology brick' claim of reuse

CTO Lou Martinez Sancho describes the construction optimization as a 'technology brick approach,' meaning the same tools extend past the first project. The article says they are already being applied to shorten licensing processes and to optimize operations.

energy for AI and AI for energyMontana Labs

On the operations side, the specific use is finding the fastest path through maintenance and refueling tasks to minimize reactor downtime. That is the same scheduling-and-sequencing problem as construction, applied to a different phase of the reactor lifecycle — which is what makes the reuse claim plausible rather than aspirational.

What the digital-twin dependency means for this build

The implication specific to this announcement is that the value here is bounded by the quality of Westinghouse's existing records and its WNEXUS twin, not by the Google models layered on top. The system predicts bottlenecks and resequences work only as well as the historical construction data it ingests.

The results so far are described only as 'significant time and cost savings' from 'early pilots' — no figures are given. The verifiable milestone remains external: whether AI-assisted scheduling actually moves those 10 reactors into construction against timelines that have slowed for decades. The AP1000 units at Vogtle were completed in 2023 and 2024, so there is a concrete construction baseline to measure the next builds against.