The electricity exists.
Getting to it is the problem.

The Gap, Stated Plainly

A data center needs electricity. Somewhere there is a generator that makes electricity. Those two facts are the easy part, and almost every article on this subject treats them as the whole story. They are not. The hard part is the connection between them, and the connection is where the years go, where the money goes, and where nearly all of the confusion lives.

Here is the thing worth holding on to before anything else. Electricity is not a commodity you can put on a truck. It has to travel down a wire that someone built, under a contract someone signed, with permission from a regulator who had to approve both. A megawatt that exists in Pennsylvania and a server hall in Virginia that wants it are not connected by wanting.

Where the Load Actually Is

Data center demand is not a national average sprinkled evenly across the country. It arrives as a small number of very large campuses in a small number of places, and that concentration is the reason a national statistic about spare generating capacity tells you almost nothing about whether any particular campus can be served.

Note the right-hand column, because it answers the question this page is about more directly than anything else on it. Of the eight largest operating campuses in the country, three generate power on their own site, and in every one of those cases the fuel is natural gas. The rest take grid power. None of them is served by a small modular reactor, because there is not one to be served by.

Meta's Hyperion in Richland Parish, Louisiana, is missing from that table on purpose. It is the largest campus in the country by announced capacity and it measures nothing at all, because it is still under construction. It belongs in the next section rather than this one, and keeping the two apart is the whole discipline of this page.

What Is Planned, and How Firm It Is

Announced campuses are much larger than operating ones, and they are announced in total facility power at full build-out, which is typically several years and several construction phases away. Both things can be true at once: the pipeline is genuinely large, and the headline numbers are the softest figures in the subject.

The site capacities in that table come from Epoch AI's April 2026 review of where the Stargate sites stand, and they sum to about 9.6 GW. Treat the total with the same suspicion as any other announced figure: OpenAI's own release describes the program as nearly 7 gigawatts of planned capacity and describes two of the sites jointly rather than separately. When the developer and the analyst disagree about the total, both numbers are announcements. For scale: Hyperion at its announced 5 GW would be comparable to the entire Plant Vogtle site in Georgia, which at close to 5 GW is the largest nuclear plant in the United States and took from 2009 to 2024 and more than $30 billion to expand.

The most useful public illustration of how soft a pipeline figure can be comes from a utility rather than a developer. Dominion Energy Virginia disclosed 53.8 GW of data center capacity under contract in July 2026, up from 16.5 GW three years earlier. It also broke that figure into stages, which almost nobody else does.

The speed of revision is its own warning. Georgia Power's projected incremental load over roughly seven years went from 400 MW in its 2022 filing to 6,600 MW in 2023 and 8,500 MW in 2024. That is a twenty-one-fold revision at one utility inside three years, and it is the cleanest single illustration of how uncertain these forecasts are — in both directions.

The procurement followed. On 19 December 2025 the Georgia Public Service Commission certified resources totalling 9,885 MW with commercial operation or delivery dates between 2027 and 2030, a figure Georgia Power has said is expected to serve data centers to the extent of roughly 80 percent. Note that certifying resources is not the same as building generation — it covers purchases and contracts too — and note the scale: Georgia Power's entire existing statewide capacity is around 22,000 MW.

The Four Routes

There are four ways to join a data center to a source of electricity, and every arrangement in this subject is one of them or a combination of them. They are not variations on a theme. They differ in who waits, who pays, who approves it, and — the difference that matters most and gets mentioned least — whether any new electricity exists at the end.

That is not a criticism of the contracts. They are real, they are large, and they keep existing plants running that might otherwise have closed — which has genuine value. It is a criticism of how they are described, because "powering AI with nuclear" is used for both kinds without distinction, and only one of them changes what the country can generate.

Route Two: The Generator Queue

If you want a new power plant to serve the grid, it joins an interconnection queue. The grid operator studies what your plant does to the system, works out what has to be reinforced, assigns you the cost, and eventually offers an agreement. This process is federally regulated and reasonably well documented, which is why it is the part of this story people can actually cite numbers for.

The last three rows are the ones that change behavior. A developer who plans around a specific queued plant is planning around something with roughly a one in eight chance of existing — and signing the interconnection agreement does not settle it, because two fifths of the capacity that got that far still withdrew. That is the counterparty risk that makes building your own generation look reasonable rather than eccentric.

Route One: The Queue Nobody Had a Process For

The ordinary route is the one nobody writes about: the data center asks the utility for service, like any other customer, and takes power from the grid. For a load the size of a small city, that turns out to be a much less settled process than connecting a power plant.

• Generator interconnection is federal. FERC has jurisdiction, there is a standard process, and the data is published annually.
• Large load interconnection was not. It has historically been handled state by state, with no standard national process and no equivalent dataset.
• That gap has now been closed, very recently. The Department of Energy directed FERC in October 2025 to open a rulemaking to accelerate large load interconnection, and the docket drew over 3,500 pages of comments. On 18 June 2026 FERC issued show-cause orders to all six organized markets — PJM, MISO, SPP, CAISO, ISO New England and NYISO — requiring each to justify or reform its tariff on application and study processes, cost transparency and cost shifting, co-location and behind-the-meter generation, transmission service for flexible large loads, and study processes for generators serving nearby large loads. Responses were due within 60 days.

Meanwhile the request volumes are extraordinary, and they are the single weakest number in this entire subject. ERCOT reported roughly 474 GW of large-load interconnection requests in August 2026, about 90 percent of them data centers, against a record actual peak of about 85.5 GW set in August 2023. The Governor of Texas ordered an audit of those projects on 3 August 2026, covering whether they bring their own generation, where their water comes from, and what public assistance they have received.

There is a real signal underneath the noise, and it points the other way from the headline. PJM cut its summer 2028 peak forecast by 4,414 megawatts, about 2.6 percent, citing stricter vetting of planned data centers and large loads. When an operator starts checking, the number goes down.

Route Three: Sitting Next to the Plant

If the wire is the problem, remove the wire. Put the data center on the same site as the generator, run a short private connection, and never touch the public grid: no queue, no transmission charges, no waiting. This is called behind-the-meter co-location, and for about two years it was widely assumed to be the answer.

1. 1
   November 2024 — FERC says no

   Amazon had bought a 300 MW data center campus physically co-located at Talen's Susquehanna nuclear plant in Pennsylvania, and PJM filed an amended interconnection agreement to expand it. FERC rejected the filing two to one, finding the non-standard terms had not been justified. Rehearing was denied in April 2025.
2. 2
   June 2025 — the market answers before the regulator finishes

   Rather than keep litigating, Amazon and Talen restructured the whole arrangement to front-of-the-meter: Susquehanna sells into PJM, Talen acts as Amazon's retail supplier, and the local utility handles delivery. The largest co-located nuclear data center in America de-co-located itself and accepted grid charges.
3. 3
   December 2025 — FERC writes a rulebook

   FERC found PJM's tariff unjust and unreasonable for lack of clarity, and ordered it to create four defined transmission service options for co-located load, along with reforms to behind-the-meter netting rules.
4. 4
   April 2026 — partially settled

   FERC accepted some of PJM's compliance filings and rejected others, notably PJM's attempt to narrow the definition of co-located load. Questions about pairing with storage, and about how the costs are allocated, remain open.

The reason regulators care is not obstruction. A plant that served every customer on the system and now serves one customer has left the common pool, while in practice still relying on the grid to back it up when the unit trips or refuels. Somebody pays for that backup and for the transmission that provides it. Deciding who is a real question, not a delaying tactic.

Route Four: Buying What Already Exists

The fourth route skips construction entirely. Find a reactor that is already running, or one that has been shut down and could be restarted, and sign a long-term contract for its output. This is the route almost every famous deal in this subject has actually taken.

8.3 GW across six contracts, with named plants and named sellers. This is a real and substantial market. Every megawatt of it comes from a reactor that already exists or one being brought back from retirement, and none of it is a small modular reactor.

Note also what the restart deals reveal about the timescale. Two of the six contracts are for plants that have not restarted yet. Three Mile Island Unit 1 is targeted for 2027; Duane Arnold for the first quarter of 2029. Restarting a reactor that already exists, on a licensed site, with the transmission already built, is the fastest way to add firm capacity anyone has found — and it still takes years.

Who Pays for the Wire

Connecting a very large load usually means building something: a substation, a transmission upgrade, sometimes a line. That cost is real and somebody is assigned it. Every co-location dispute, every large-load rulemaking and every state proceeding on this subject is, underneath, an argument about who.

• The developer's position. We pay our interconnection costs like any customer, we bring load that spreads fixed costs across more sales, and existing customers benefit.
• The consumer advocate's position. If the upgrade is built and the campus never arrives, or arrives at a fraction of its stated size, ratepayers hold an asset they did not need and did not ask for.
• The reason the second position has force. Wood Mackenzie expects operators to commit to only about 28 percent of the capacity being requested, which leaves roughly 72 percent that it does not expect to be committed to. Note the precise claim: that is one consultancy's estimate, and being committed to is not the same event as being built. The risk of building for a load that never shows up is not hypothetical even so.

States have started writing the answer into the tariff rather than arguing it case by case. Georgia is the most developed example: the Public Service Commission ordered a base rate freeze through 2028 in July 2025, imposed minimum billing and longer contract terms on large loads, required new data center contracts to be filed thirty days before execution, and required Georgia Power to backstop the costs financially through 2031 if the contracts do not materialize.

The Answer Most Developers Actually Chose

Faced with four routes that are all slow or constrained, a great many developers took the most aggressive version of route 3: not co-locating beside somebody else's plant, but building their own generation on site and waiting for nobody. That generation is overwhelmingly natural gas, not nuclear, and it is happening now rather than in the 2030s. It is also the version of route 3 that genuinely adds supply, because the plant is new.

• The International Energy Agency projects 15 to 27 GW of on-site natural gas serving data centers by 2030, predominantly in the United States, explicitly because developers are navigating slow grid connections.
• Global Energy Monitor counted 252 GW of US gas capacity in development in January 2026, with more than a third — about 84 GW — slated to power data centers directly on site.
• For comparison, total small modular reactor capacity under construction in the United States for any customer at all is about 0.4 GW, and none of it is contracted to a data center.

Even this has a queue of its own, just not a regulatory one. Gas turbines are sold out. As of manufacturer earnings reported in August 2026, GE Vernova held a 116 GW backlog with delivery slots into 2031; Siemens Energy quoted about three years and held 69 GW; Mitsubishi held 35 GW. Order intake is running three to seven times shipment rate, which is what a sold-out factory looks like.

Where Small Modular Reactors Fit

The case for a small modular reactor next to a data center is a connection argument, not a generation argument. It is not that nuclear electricity is cheaper — it is not. It is that a reactor on your own site is route 3 in its supply-adding form, and it sidesteps both queues at once: no waiting for the grid to reach you, and no waiting for somebody else's plant to be built. It is the same argument that put gas turbines on eight campuses, with a different fuel.

That is a coherent argument. The difficulty is entirely in the present tense. As of August 2026, no small modular reactor is operating anywhere in the United States, and none is under construction with a data center as its contracted customer. The two most advanced construction permit applications in the country are TVA's Clinch River, for the Tennessee Valley grid, and X-energy's Long Mott, for a Dow chemical works at Seadrift, Texas. Neither is an AI project.

This is why the sector's own scoreboard is worth watching more closely than its announcements. The thing that would change this picture is not another framework agreement. It is a construction permit application that names a data center as the customer, or an operating licence application from one of the companies already building. Neither had happened by August 2026.

How to Read a Claim About This

Almost every misleading claim in this subject comes from describing one route in language borrowed from another. The decoder is short. When you read that a company is powering AI with some source of electricity, work out which of the four routes is being described, and then ask the question in the right-hand column.

Where This Comes From

• Queue duration and completion rates. Lawrence Berkeley National Laboratory, Queued Up, 2026 edition. Read its scoping note: it covers generators, not large loads.
• Co-location. FERC's November 2024 order rejecting the Susquehanna interconnection agreement amendment, the April 2025 rehearing denial, the December 2025 order directing PJM to establish service options, and the April 2026 compliance order.
• Large-load requests. ERCOT filings and Texas state reporting, August 2026; Wood Mackenzie's 1,066 GW estimate, August 2026; US generating capacity from EIA.
• Contracted capacity. Summed from the six announced technology company agreements, each with a named plant and a stated megawatt figure.
• On-site gas. International Energy Agency projections to 2030 and Global Energy Monitor's January 2026 count of US gas in development.