By ByteBot
AI’s real bottleneck isn’t GPUs—it’s gigawatts. On January 9, Meta announced nuclear energy deals with three companies totaling 6.6 GW of capacity to power Prometheus, its AI supercluster in Ohio. This is the largest corporate nuclear purchase in US history, and it signals a shift the industry hasn’t fully processed: the AI race is now an energy race.
The Energy Ceiling
Meta needs 6.6 GW for a single AI datacenter—more power than the entire state of New Hampshire consumes. That’s not a compute problem. That’s a fundamental constraint on how fast AI can scale.
Training GPT-4 consumed 50 gigawatt-hours, enough to power San Francisco for three days. AI datacenter demand is projected to surge 160% by 2030, from 10 GW to 68 GW. And Meta isn’t alone: Google, Microsoft, and Amazon are all pursuing nuclear deals. The pattern is clear—energy infrastructure, which takes 5-10 years to build, is now the long pole. Not chips.
For developers, this means cloud pricing will increasingly reflect energy costs, not just compute costs. Regional AI availability depends on local power infrastructure. Energy efficiency isn’t an optimization afterthought anymore; it’s a first-order architectural concern.
Three Bets, Three Timelines
Meta’s strategy hedges across technologies and delivery timelines with three distinct partnerships.
Vistra provides 2.6 GW from existing nuclear plants in Ohio and Pennsylvania, plus 433 MW of uprates—the largest corporate-backed nuclear uprates ever. Purchases begin late 2026, with full capacity by 2034. This is the safest bet: proven technology, operating plants, near-term delivery.
Oklo offers up to 1.2 GW from new Aurora small modular reactors at a Pike County, Ohio nuclear campus. First reactors could come online as early as 2030, with full capacity by 2034. The catch? Oklo has no operating reactors yet. Hacker News skeptics call it “total vaporware.” This is the risky, mid-term bet.
TerraPower, Bill Gates’ nuclear venture, will supply up to 2.8 GW from eight Natrium reactors. Each unit provides 345 MW of baseload power with built-in energy storage that can ramp to 500 MW. Initial units target 2032. TerraPower’s first commercial Natrium reactor is already under construction in Wyoming, expected in 2030. This is the long-term innovation bet.
The logic: spread bets across immediate availability, near-term deployment, and future innovation. If one partner delays, the others cover.
Why Not Solar or Wind?
Datacenters need 24/7 consistent power. Solar and wind can’t deliver that without massive battery storage.
Nuclear plants run at 90%+ capacity. Solar averages 23%. Wind averages 34%. The US has 37.4 GW of battery storage today, growing to 187 GW by 2030. But even with storage, renewables can only serve about 80% of datacenter demand. Baseload generation—nuclear or gas—is still required for the other 20%.
Goldman Sachs projects natural gas will power 60% of incremental US datacenter demand through 2030. Nuclear is preferred for its carbon-free baseload, but the difficulty of building new plants means gas and renewables are more realistic short-term solutions. Meta’s choosing nuclear anyway: Vistra provides immediate carbon-free baseload, while Oklo and TerraPower aim for long-term capacity without emissions.
The Ohio Corridor
Prometheus, Meta’s supercluster in New Albany, Ohio, will be the world’s first gigawatt-capable datacenter. Meta’s been in New Albany since 2017 on a 740-acre campus. The location isn’t random.
Ohio has existing nuclear infrastructure—Vistra’s Perry and Davis-Besse plants are nearby. It has access to the PJM grid, regulatory support (Ohio approved a 200 MW natural gas plant for the campus), and Oklo’s new nuclear campus is also planned for Ohio.
AI development is clustering near energy sources. That creates regional disparities in compute access, talent migration to energy hubs, and cloud region availability tied to energy geography. If you need low-latency access to cutting-edge AI infrastructure, colocation in energy-rich regions like Ohio may become necessary.
The Hyperscaler Energy Arms Race
Meta’s 6.6 GW is the largest deal, but it’s not the only one. Microsoft signed a 20-year agreement to restart Three Mile Island by 2028. Google has a 500 MW SMR fleet deal with Kairos Power, targeting 2030-2035. Amazon is pursuing 5+ GW from X-energy SMRs by 2039.
Only hyperscalers can afford gigawatt-scale nuclear deals. Startups are reliant on cloud pricing, with no direct access to energy markets. If energy costs spike, small players face a competitive disadvantage. This reinforces the big tech moat.
The counter-argument: hyperscaler investments add total capacity to the grid. Meta’s 6.6 GW benefits all users, not just Meta. Cloud economies of scale still favor startups over self-hosting.
There’s a near-term bottleneck, though. Vistra provides immediate relief starting late 2026, but Oklo and TerraPower are delayed until 2030-2034. That’s a four- to eight-year gap where AI scaling may hit energy constraints before new capacity comes online.
What This Means
Meta’s nuclear deal confirms what the industry is starting to realize: AI progress isn’t limited by compute anymore. It’s limited by the grid. Cloud pricing, regional availability, and even architectural decisions will increasingly reflect energy constraints.
The AI race is now an energy race. And energy takes time.
