Every Hyperscaler Has a Nuclear Deal—Here Is What Each Actually Bought

Nearly every large cloud and AI company now has a nuclear announcement. The logos make the market look settled: Meta has 6.6 gigawatts, Google has 1.8 GW with Elementl, Amazon has up to 5 GW with X-energy, Microsoft is restarting a reactor, and Nvidia is helping nuclear developers use AI.

The problem is that these numbers describe different things. A 20-year power purchase agreement from an existing plant is bankable demand. An option to buy power from a potential reactor is useful, but the reactor may not have a license, final site, financing, or construction contract. A software partnership has no purchased megawatts at all.

This reference table translates every headline into the commercial instrument that was actually announced.

The nuclear deal table

Capacities are not additive without adjustment. Some are existing plants, some include uprates, and some are maximum development rights that may overlap later project announcements. The table is a pipeline, not a generation meter.

Why “9.8 GW committed” needs a footnote

An industry total near 9.8 GW is directionally useful: technology companies are underwriting nuclear demand at a scale utilities can no longer ignore. It becomes misleading when presented as power that has been bought, permitted, or scheduled to turn on.

Use four buckets:

1. Operating power under contract. A buyer signs for electricity from an existing plant.
2. Restart or uprate. A contract helps add capacity to, extend, or reopen an established facility.
3. Development and offtake option. A buyer funds milestones and may purchase power if the project reaches them.
4. Technology partnership. Companies provide software, compute, design, or regulatory support without buying electricity.

Only the first category is close to ordinary procurement. The second still carries regulatory and engineering risk. The third carries full first-of-a-kind reactor risk. The fourth may accelerate the industry but should count as zero megawatts purchased.

Meta: the broadest portfolio, not one 6.6 GW purchase

Meta's June announcement combined three strategies.

Vistra offers the near-term foundation. Meta contracted around existing generation from the Perry and Davis-Besse plants in Ohio and Beaver Valley in Pennsylvania, while supporting license extensions and uprates expected to add 433 MW. Existing reactors reduce technology risk, although the uprates and extensions still need approvals and work.

Oklo adds up to 1.2 GW of advanced reactor development in Ohio. TerraPower adds two initial Natrium units and rights to support more units, with the broader portfolio reaching up to 2.1 GW. These are consequential demand signals. They are not equivalent to plugging a data center into 3.3 GW of new reactors today.

The 6.6 GW headline is therefore best read as Meta's maximum nuclear supply and development portfolio through 2035.

Google: a fleet agreement plus three site options

Google's Kairos agreement covers up to 500 MW from multiple small reactors. The first unit targets 2030, with additional deployments through 2035. It is important because a fleet order can spread first-of-a-kind learning across repeated builds.

The Elementl announcement is earlier-stage. Google agreed to fund development work at three potential sites, each with at least 600 MW of capacity, and obtained an option for power. That justifies the 1.8 GW pipeline label. It does not justify “Google bought 1.8 GW that will definitely be online.” Sites, reactor technology, permits, financing, and final offtake remain to be resolved.

Microsoft: a real PPA and a separate software alliance

Microsoft's agreement with Constellation is the easiest to classify: a 20-year power purchase agreement supports restarting the 835 MW reactor now called the Crane Clean Energy Center, formerly Three Mile Island Unit 1. The target is 2028, subject to Nuclear Regulatory Commission approval and the physical work required for restart.

Microsoft's work with Nvidia belongs in another column. Their “AI for Nuclear Energy” effort applies AI, simulation, digital twins, and cloud tools to design, operations, and licensing. The U.S. Department of Energy is also exploring AI to shorten reactor-review timelines. These efforts may unlock capacity, but they do not disclose an Nvidia nuclear PPA.

That distinction is precisely why a reference table is needed: “Nvidia has a nuclear deal” can be true as partnership language and false as purchased-power language.

Amazon: the largest maximum, the longest runway

Amazon's X-energy arrangement targets more than 5 GW of small modular reactor projects by 2039. It includes an investment in X-energy and project-development pathways with utilities and developers. An initial Energy Northwest configuration in Washington has been discussed at roughly 320 MW with expansion potential.

The maximum is large enough to shape manufacturing capacity. It is also far enough out that regulatory design approval, high-assay fuel availability, site permits, supply chains, project finance, and construction execution remain major variables.

Treat 5 GW as a strategic portfolio target, not a delivery forecast with the certainty of an operating gas plant.

What the July federal push changes

The Trump administration's reported July 21 $200 million push involving Oklo, X-energy, and supporting technology is another acceleration signal. Public support can fund demonstrations, supply chains, or licensing work. It cannot erase the sequence required to operate a reactor:

design → site → environmental review → license → financing → procurement
→ construction → fuel → commissioning → grid service

AI-assisted permitting may compress document review. It cannot make incomplete safety evidence complete. The right success metric is review time at equal or better safety quality, not simply fewer days.

Why hyperscalers want nuclear

Data centers need round-the-clock power. Wind, solar, storage, gas, transmission, and demand response all play roles, but hyperscalers value nuclear's high capacity factor and low operational carbon emissions. Long contracts also make the cost more predictable.

The demand is not purely environmental. Grid queues are long, gas turbines are scarce, and communities increasingly resist campuses that appear to compete with households for electricity. A dedicated clean-power story can improve siting prospects, though it does not remove local questions about water, transmission, waste, land, or subsidies. See our data center backlash map for the consent test.

How to read the next nuclear announcement

Ask seven questions before repeating the gigawatt number:

1. Is the capacity existing, an uprate, a restart, or new construction?
2. Is there a signed PPA, an option, an investment, or only a memorandum?
3. Who owns construction and cost-overrun risk?
4. Has the site been named and controlled?
5. What NRC design and site approvals remain?
6. Is the date for first unit, full fleet, or merely a development milestone?
7. Does “up to” depend on expansion rights that may never be exercised?

Then label the number in plain language: operating, contracted, under development, optioned, or partnership-only.

The practical verdict

The nuclear turn is real. Big technology companies are providing the long-duration demand, early capital, and credibility that reactor developers have lacked. Microsoft has a defined restart PPA. Meta combines existing purchases with advanced-reactor development. Google is supporting repeat deployment and site formation. Amazon is aiming at manufacturing-scale volume.

But the current fleet of headlines is several project stages pretending to be one unit of measurement. Gigawatts do not become interchangeable until plants are licensed, financed, built, fueled, connected, and delivering.

For AI customers, the near-term price still reflects today's grid. New reactors may improve capacity and carbon intensity in the 2030s; until then, efficiency remains the fastest capacity source. That is why choosing the right-sized model matters now.

How this table should be updated

Each project needs two dates: the announcement date and the latest verified milestone. Update capacity only when a company, utility, regulator, or project owner clarifies whether the number represents existing output, incremental uprate, initial units, or an expansion maximum. Preserve the earlier number in a note rather than silently rewriting history.

Move a project between stages only with evidence. A named site is not a permit; a submitted license is not approval; an approval is not a final investment decision; construction is not commercial operation. Record delays in the same table so the target year does not appear certain simply because no new press release exists.

Finally, keep partnership-only announcements at zero purchased megawatts until an offtake or ownership commitment is disclosed. That convention may make the industry total less exciting, but it makes the table citable. Readers can then sum operating, contracted, optioned, and aspirational capacity according to the question they actually have.

Related on explainx.ai

• The AI data center backlash, mapped
• Virginia's data center electricity tax
• Data centers, water, and active lawsuits
• AI energy and sustainability guide
• AI token pricing, explained
• Nvidia–OpenAI $250B backstop for an Ohio 10GW campus

Capacities and target dates are based on company announcements and public reporting available July 26, 2026. “Up to” values are maximum development ambitions, not guaranteed output. Nuclear schedules remain subject to licensing, financing, fuel, construction, interconnection, and commercial milestones.