The International Energy Agency’s April 2026 update put a number on the question that has been dominating energy strategy meetings since GPT-4 went into production: global data-center electricity consumption will roughly double from 485 TWh in 2025 to 950 TWh in 2030, taking data centers to about 3% of global electricity demand. The AI-focused subset of that load triples in the same window. Servers running inference workloads alone — not training, just inference — grow at roughly 30% per year and account for nearly half the net increase in global data-center consumption between 2024 and 2030.
In the United States, that translates to about 240 TWh of incremental annual demand by 2030, a 130% increase over 2024. China adds 175 TWh, up 170%. There is no scenario where the existing grid, plus new gas, plus the steady-state buildout of renewables, absorbs that load on the schedule the hyperscalers need it. The conclusion the hyperscaler CFOs reached in 2024 has only hardened in 2026: baseload nuclear is the only firm, low-carbon, twenty-four-hour electron that scales to a gigawatt site.
That is why, by May 2026, every major US hyperscaler has signed at least one nuclear power deal, and the pipeline of conditional offtake agreements between data-center operators and SMR projects has grown from 25 GW at the end of 2024 to 45 GW today. The question is no longer whether the deals get signed. The question is how much of that 45 GW actually shows up on the grid, and when.
The Signed Stack, Reconciled
Thirteen announced projects, four hyperscalers, roughly 9.8 GW of nameplate capacity that has progressed past the press-release stage. Sorted by first-electrons date:
Microsoft — Crane Clean Energy Center (former TMI Unit 1). 835 MW pressurized water reactor restart in Pennsylvania. Twenty-year PPA worth roughly $16 billion. Constellation has the facility at 65% staffing and has tested the main generator and turbines. Constellation secured a $1 billion DOE loan in late 2025 to fund the restart, and an early interconnection request is in front of PJM. First electrons targeted for 2027, a year ahead of the original 2028 schedule. This is the only deal in the stack that produces nuclear electrons for an AI data center before 2030.
Amazon — Susquehanna (Talen Energy). 1,920 MW PPA from the existing Susquehanna PWR pair. After FERC rejected the amended interconnection service agreement in November 2024 in a 2-1 vote, the companies restructured from behind-the-meter to a front-of-the-meter retail arrangement. Talen now serves AWS as a licensed Pennsylvania retail electricity provider, and the reconfiguration takes place during the Spring 2026 refueling outage. The 1.92 GW does not flow as private wires into the data center — it goes to the grid, and Amazon’s incremental consumption is netted against it.
Amazon — Cascade Advanced Energy Facility (X-energy + Energy Northwest). Four Xe-100 high-temperature gas-cooled reactors, 80 MWe each, totaling 320 MWe with an option to expand to 960 MWe. Part of a larger 5+ GW Xe-100 commitment Amazon made when it took an equity stake in X-energy. As of May 2026, X-energy has not yet filed for an NRC license on the Washington project. First electrons are not realistically expected before the mid-2030s.
Google — Kairos Power. 500 MW across multiple Kairos KP-FHR (fluoride-salt-cooled high-temperature) reactors. The Hermes demonstration unit in Oak Ridge is the bridge to the first commercial units. Online window: 2030 to 2035.
Meta — Vistra (existing fleet). 2.1+ GW under twenty-year PPAs covering the Beaver Valley plant in Pennsylvania plus Perry and Davis-Besse in Ohio. All operating PWRs. Power flow begins as soon as contracts execute, but no new nameplate is added — Meta is buying clean firm capacity from already-running units to feed its Prometheus AI supercluster in New Albany, Ohio.
Meta — TerraPower Natrium. Two 345 MWe sodium-cooled fast reactors with molten-salt thermal storage. First unit targets 2032 or later, dependent on the Kemmerer first-of-a-kind learning curve and NRC construction permit progression.
Meta — Oklo Aurora. 1.2 GW campus in Pike County, Ohio. Pre-construction and site characterization begin in 2026; first phase targeted as early as 2030.
The arithmetic is uncomfortable. Of the roughly 9.8 GW that has progressed past board approval, exactly 835 MW arrives in 2027. The next slug — Susquehanna’s 1.92 GW reconfiguration — is an accounting flow, not new capacity. Real new nameplate from new reactors does not start showing up until 2030 with Oklo, and the largest tranches (TerraPower, Kairos, Cascade) are 2032 to mid-2030s. The 45 GW conditional offtake number is real, but it is a 2035-and-beyond number, not a 2027 number.
Where the Schedule Actually Breaks
Three constraints determine which of these deals slip and which deliver. None of them are about reactor physics.
The NRC review queue. The Cascade project has not yet filed. The Dow-X-energy Long Mott station in Texas filed its construction permit application in 2025 and is the leading Xe-100 case study for what an NRC review actually looks like at the post-design-certification stage. Construction permits for Gen IV designs are taking 24 to 36 months even with the agency’s modernization efforts. Operating license reviews add another 18 to 24 months. The schedule math forces hyperscalers into either restart deals (Crane) or existing fleet PPAs (Susquehanna, Vistra) for any pre-2030 electron.
HALEU and TRISO fuel. Every advanced reactor in the stack — Xe-100, Aurora, Natrium, Kairos KP-FHR — runs on High-Assay Low-Enriched Uranium (HALEU, 5-20% U-235). US HALEU production capacity in 2026 is measured in single tonnes per year. The DOE HALEU Availability Program awards from 2024 began commercial scale-up, but a single Xe-100 four-pack requires roughly 1.5 tonnes of TRISO fuel particles for its first core load. Multiplying that across Cascade, Long Mott, the Meta Natrium pair, and the Kairos fleet produces a fuel requirement that current US capacity does not meet until late this decade. Russia is structurally out of the supply chain. Centrus’s Piketon cascade and the X-energy/TX-1 TRISO line in Oak Ridge are the two lines that determine whether the 2032 dates hold.
FERC and the behind-the-meter question. The November 2024 ISA rejection at Susquehanna established that the agency will not approve arrangements that look like one large customer absorbing capacity that the grid paid to build. The Talen-AWS restructure is the template — front-of-the-meter, licensed retail provider, no FERC approval required. Every co-located deal going forward has to thread that needle. Behind-the-meter SMR projects, where the reactor is sited next to or inside the data-center campus, may escape the issue, but only if they are entirely new generation and the data center is the only direct customer. The Oklo Pike County campus is designed exactly that way; the Cascade project less so.
What This Looks Like Inside the Plant
Every one of these new builds, whether a restart like Crane or a first-of-a-kind like Natrium, is a Process Safety Management site under 29 CFR 1910.119. The new entrants — Xe-100, Aurora, KP-FHR, Natrium — also live under 10 CFR 50 and 10 CFR 52 for the nuclear envelope, but the human-in-the-loop instrumentation, the alarm management, the management of change procedures, and the contractor safety programs are PSM-shaped problems with nuclear-grade documentation requirements layered on top.
The combination matters because the workforce that will operate, maintain, and inspect this fleet does not yet exist at the scale required. The Crane restart alone is hiring roughly 600 positions. Cascade, Long Mott, the Meta Natrium pair, and the Kairos fleet collectively need an order of magnitude more operators, mechanics, I&C technicians, and PSM/QA staff than the US nuclear workforce currently supplies. AI-augmented procedure execution, AI-assisted MOC review, and AI-driven incident-precursor detection are not optional augmentations to that buildout — they are the only way the operating envelope holds with a workforce of that age and tenure distribution.
That is the dual-use thesis behind the energy-plus-AI convergence the hyperscalers are now financing. The same compute that the data center sells to enterprise customers also runs the safety-instrumented systems, the predictive maintenance models, and the procedure-compliance verification that keeps the reactor next door inside its license. The IT load and the OT load become each other’s customers.
The Schedule Is the Story
The honest summary of the May 2026 nuclear-for-AI position: physics solved, regulatory pathway open, fuel constrained, workforce constrained, construction-schedule constrained. The 45 GW pipeline is real intent backed by real balance sheets. The 9.8 GW of nameplate that has cleared board review is the credible 2030-2035 number. The 835 MW arriving in 2027 from a restart is the only datapoint that proves any of this works on the hyperscalers’ actual timeline.
That asymmetry — twenty-six months from a single restart to the next credible electron — is the structural risk every hyperscaler CFO is now hedging with natural gas, geothermal, and grid-side renewables. It is also why the next twenty-four months of NRC docket motion, HALEU production milestones, and FERC interconnection rulings will set the slope of the curve for the rest of the decade. Watch the regulator, not the press release.
For operators, EPC firms, and AI infrastructure teams thinking about how to underwrite, design, build, or commission this fleet, the operating-safety side of the buildout is where the schedule risk and the workforce risk converge. Porritt Inc. builds AI-native tools for exactly that problem — PSM compliance, NORMEX standards interpretation, AI-augmented inspector workflows, and incident-precursor detection — across both the nuclear envelope and the conventional balance-of-plant. If you are scoping a co-located SMR project, a restart workforce ramp, or an AI-and-OT safety integration for a new hyperscaler site, we should talk. Reach out for a consultation.
Timothy Porritt is founder of Porritt Inc., building AI-powered tools for process safety, engineering compliance, and industrial operations across refining, nuclear, and AI infrastructure. Based in Salt Lake City, Utah.