Meta's 6.6-Gigawatt Nuclear Deal Won't Power a Single AI Server Until 2032

On January 9, 2026, Meta signed agreements with Vistra, TerraPower, and Oklo covering a combined 6.6 gigawatts of nuclear power, intended partly to support its Prometheus supercluster data center in New Albany, Ohio. It is one of the largest tech-nuclear deals ever announced. It is also, almost entirely, a bet on power that doesn't exist yet. No data center anywhere in the world is currently running on electricity from a small modular reactor (SMR) — the technology hyperscalers are betting billions on remains undeployed at commercial scale.
The breakdown of Meta's deal makes the timeline problem obvious. TerraPower's commitment covers up to eight Natrium reactors delivering 2.8 gigawatts of baseload power plus 1.2 gigawatts of integrated thermal storage — with the first unit not expected online until 2032. Oklo's share involves an advanced reactor campus in Pike County, Ohio, targeting up to 1.2 gigawatts by 2034. Only Vistra's 2.6-gigawatt contribution is real today, and it isn't new nuclear at all — it's existing Ohio and Pennsylvania plants, including 433 megawatts from capacity uprates on reactors already running.
The Pattern Repeats Across the Industry
Meta isn't an outlier. Amazon's partnership with X-energy, South Korea's KHNP, and Doosan aims to mobilize up to $50 billion in public and private investment to deploy X-energy's Xe-100 reactor design, each unit capable of producing 80 megawatts of electricity or 200 megawatts of high-temperature steam. Across the industry, SMR-for-AI announcements in 2026 share a structure: large headline gigawatt figures, multi-year construction and licensing timelines, and a first-unit delivery date somewhere between 2030 and 2035.
That structure isn't dishonest — it reflects how nuclear construction actually works. The U.S. Nuclear Regulatory Commission's licensing process for a new reactor design, even a standardized SMR, typically runs several years before ground is even broken, and construction of a first-of-its-kind unit adds several more. TerraPower's Natrium design uses a sodium-cooled fast reactor architecture that has never been deployed commercially at this scale in the U.S., meaning the 2032 date already assumes a relatively smooth regulatory and engineering path. Oklo's compact fast reactor design faces a similar first-deployment burden.
So What's Actually Powering AI Data Centers Right Now?
The honest answer is natural gas, grid power, and increasingly, on-site gas turbines purchased directly by hyperscalers to bridge the gap. Data center power demand from AI training and inference is growing faster than any new generation source — nuclear included — can plausibly be built. Nuclear's appeal isn't that it solves 2026's power problem; it's that it offers 24/7 carbon-free baseload at a scale solar-plus-storage struggles to match for industrial loads, and hyperscalers are willing to pre-pay for capacity nearly a decade out because the alternative is competing for scarce grid interconnection slots with everyone else building a data center in the same regions.
This is also why existing-fleet deals like Vistra's matter more in the near term than any SMR announcement. Capacity uprates on operating reactors, power purchase agreements with plants that would otherwise face retirement, and life-extension deals are the only nuclear-adjacent moves that affect electricity supply before 2030. The SMR contracts are options on the 2030s, priced and signed now because reactor manufacturing capacity, specialized components like TRISO fuel, and licensing review slots are themselves scarce resources that early movers are locking up.
What to Actually Watch
Three signals will tell you whether this cycle of announcements becomes real power rather than another round of press releases that quietly slip. First, watch NRC licensing milestones for TerraPower's Natrium and Oklo's Aurora designs specifically — a construction permit approval is a harder signal than a site selection announcement. Second, watch whether hyperscalers' near-term capacity gap gets filled by gas turbines bought in bulk (Meta, Microsoft, and others have already placed large turbine orders), because that tells you how much pressure is actually being absorbed by nuclear versus fossil generation in the interim. Third, watch cost overruns and schedule slips on these first-of-kind units — nuclear construction has a long history of both, and an SMR industry built on investor confidence rather than operating plants is unusually exposed to a single high-profile delay. The gigawatts are real commitments. The power, for now, is still mostly a forecast.