Nvidia and PreussenElektra Are Sizing Up Decommissioned German Nuclear Plants for AI Data Centers
A preliminary partnership announced October 21, 2026 would reuse shuttered nuclear sites' power and industrial infrastructure — skipping the hardest parts of building AI compute capacity from scratch.
The hardest problem in AI infrastructure right now isn't compute design — it's land, power, and the years it takes to get both permitted and connected. Nvidia and German energy company PreussenElektra think decommissioned nuclear plants might solve all three at once.
The two companies announced a preliminary partnership on October 21, 2026 to assess whether shuttered German nuclear sites could be repurposed to host high-performance AI data centers. No investment figure has been disclosed, and no construction approval or definitive development agreement has been finalized. What exists, for now, is a structured exploration — but the logic behind it is hard to argue with.
Why Dead Nuclear Plants Are Attractive Real Estate
Decommissioned nuclear facilities carry infrastructure that took decades and billions to build: heavy-duty grid interconnections, cooling systems, industrial-grade land already zoned for large-scale energy use, and physical structures built to handle enormous operational loads. For a conventional greenfield data center, assembling that combination — permits, grid access, water rights, structural foundation — can add years to a project timeline and substantial cost before a single server rack is installed.
Reusing existing power and industrial infrastructure, as the Nvidia-PreussenElektra plan intends, compresses both variables. The grid hookups are already there. The land is already cleared and compliant. The industrial backbone doesn't need to be engineered from nothing. For AI workloads that demand dense, continuous power delivery at scale, that starting position matters enormously.
Germany is a specific context worth naming. The country completed its nuclear phase-out in 2023, leaving a set of sites that are decommissioned but not yet fully dismantled — a narrow window where existing infrastructure retains usable value. PreussenElektra, as the operator of several of those plants, holds the site knowledge and energy expertise that Nvidia would need to evaluate feasibility. The partnership structure reflects that: this is an assessment exercise, not a build announcement.
What the Partnership Actually Is — and Isn't
The announcement is deliberately scoped. A preliminary partnership to assess repurposing is not a commitment to build. No capital has been committed publicly. No regulatory pathway has been cleared. German decommissioned nuclear sites carry complex legal and environmental obligations — the cleanup requirements alone shape what any future use could look like, and that due diligence is presumably part of what this assessment covers.
That restraint is appropriate. The gap between "this site has useful infrastructure" and "this site is permitted, safe, and shovel-ready for hyperscale AI compute" is significant. Grid capacity that served a nuclear plant's own generation needs may not map cleanly onto the import-side power demands of a data center. Cooling infrastructure designed for reactor thermal management is a different engineering problem than server cooling. These are solvable questions — but they are questions, not foregone conclusions.
What Nvidia brings to the table is the demand signal and the hardware roadmap. If the sites prove viable, Nvidia has every incentive to see large-scale GPU clusters deployed in Europe, where AI infrastructure buildout has lagged the United States and parts of Asia. PreussenElektra brings site access, operational history, and energy-sector relationships. The partnership is a logical pairing for an exploration that neither party could run as effectively alone.
The Bigger Infrastructure Bet
This move fits a pattern that is quietly reshaping where AI compute gets built. Data center developers and hyperscalers have been circling unconventional sites — retired fossil fuel plants, industrial brownfields, isolated grid nodes with stranded capacity — because the conventional site development pipeline cannot keep pace with AI demand. The appeal of nuclear sites specifically is that they tend to sit at the high end of the infrastructure quality spectrum: robust grid connections, serious physical plant, and locations that were deliberately chosen for power reliability.
The Nvidia-PreussenElektra announcement doesn't resolve the hard questions about whether any specific German site can be converted economically and at speed. But it signals that the search for AI compute capacity has moved well past the obvious locations. When one of the world's most valuable chip companies partners with a nuclear operator to study turning reactor sites into GPU clusters, the era of conventional data center siting is visibly ending.
The real shift isn't about nuclear nostalgia. It's about the industry finally reckoning with the fact that building AI infrastructure at the required pace demands a fundamentally different approach to land and power — one that starts with what already exists, not what can be permitted from scratch.