Washington announced two gigawatt-scale AI power projects in nine days. Neither one obliges anybody to build anything.

On July 20 the National Nuclear Security Administration said it had selected Amentum to enter negotiations over a phased lease at the Savannah River Site in South Carolina, covering a 1 gigawatt AI data center and about 2 gigawatts of dedicated on-site generation. Nine days later the Department of Energy named Brookfield and NextEra Energy as its partners at the Paducah site in Kentucky, behind more than $100 billion of announced private investment and a 1.8 gigawatt campus.

Both releases lead on a gigawatt figure. Very little stands behind either number.

A gigawatt is a gigawatt. What changes from one announcement to the next is who is obliged to deliver it. Capacity can be announced, which commits nobody to anything. It can be permitted, which means a regulator has allowed it to be built but nobody has to build it. It can be contracted, which means someone has agreed to pay for it. Or it can be delivered, which means it exists and is running. The figure looks identical in all four cases. The obligation behind it does not. Four columns, one number.

That distinction is now doing most of the work in the AI power story. Savannah River is announced. Paducah is announced, with a Kentucky commission proceeding still ahead of it. Stargate's headline totals are announced. What has actually been built and energized, in Texas, in Louisiana and in Memphis, is a small fraction of what has been said out loud, and it arrived by routes that had nothing in common.

For utilities, regulators and investors, the obligation is now the whole question. The measurement fight over how much energy AI actually uses and the grid constraint AI compute keeps running into both assume the power gets built. The column a gigawatt sits in is what tells you whether it will.

Two Deals, Nine Days, Two Structures

The catch at Savannah River is the verb. Selected to enter negotiations comes before a lease, before permits, before safety and security approvals, before financial terms and before any timetable. NNSA said as much itself: selection "does not constitute a final lease award."

Amentum, announcing the same day, described the project as initially powered by natural gas bridging to advanced nuclear energy. DC BLOX confirmed itself as the consortium's digital infrastructure partner. No investment figure, no rent, no phasing schedule and no construction date has been published by any of them.

The September 2025 solicitation set a harder test than a press release. The chosen partner has to build, operate and decommission the project at its own cost and secure its own utility interconnection agreements. That is a real obligation, and still not a commercial settlement.

Paducah is further along on paper and no further along commercially. Brookfield develops and operates the campus, NextEra builds and owns the generation, and three Kentucky utilities are named in the release. The energy package is roughly 2 gigawatts of new grid-connected gas plus up to 2.6 gigawatts of battery storage, with completion expected in 2031. The Department of Energy says the power service agreement "remains subject to approval by the Kentucky Public Service Commission," and the transaction still requires definitive documentation.

The columns are already blurred in the paperwork. DOE describes a 1.8 gigawatt campus. The coalition's own announcement that day puts compute load above 1.2 gigawatts, with 1.8 gigawatts the utility capacity standing behind it, up to 4.6 gigawatts of dedicated generation, and full build-out in 2032. Those are three different quantities in one project.

So Paducah discloses more structure than Savannah River. It also has no named anchor tenant. A campus on that scale carrying a $100 billion headline with nobody publicly committed to filling it is a different kind of incomplete.

Half The Federal Program Has No Partner

Both sites sit inside the four-site program the Department of Energy announced in July 2025: Idaho National Laboratory, Oak Ridge Reservation, Paducah and Savannah River.

Two of the four are now matched to partners. Idaho's first applications were due in November 2025, with later ones accepted on a rolling basis, and Oak Ridge's proposals were due on December 1. Neither has produced a publicly announced selection since.

The program isn't failing. The four sites were simply never one package. Idaho and Oak Ridge left capacity to the applicants, and Oak Ridge offered two separate parcels. Paducah's solicitation prioritized small modular reactors, microreactors, geothermal and underground thermal storage. The winning bid arrived with gas and batteries.

The Department has also moved faster outside these four sites. In March it announced a 10 gigawatt campus at the Portsmouth site in Ohio with SoftBank's SB Energy and AEP Ohio, backed by 9.2 gigawatts of new gas, and construction on an initial 800 megawatt phase is expected this year. DOE's own Paducah release says the Kentucky deal builds on the Portsmouth model. So the federal program has one site with steel in its future, two sites with named partners and no signed leases, and two sites with nothing announced at all.

The federal government supplies land, solicitation structure and political priority. Developers still supply engineering, interconnection, capital, power contracts and time. A site announcement starts the clock. It does not close the gap.

Announced, Permitted, Contracted, Delivered

Private campuses show the same ladder at larger scale, and the wording drifts as they climb it.

In July 2025 OpenAI said it and Oracle had "entered an agreement to develop 4.5 gigawatts" of additional Stargate capacity. By September the same company was describing the same deal as "up to 4.5 gigawatts" inside an announcement that put Stargate at "nearly 7 gigawatts of planned capacity and over $400 billion in investment". A month later Michigan took the running total to over 8 gigawatts and more than $450 billion. In April 2026 OpenAI said it had already passed its 10 gigawatt commitment, with more than 3 gigawatts added in 90 days, and has published no running total since.

Those are company statements about plans and agreements. None of them is utility-verified capacity.

Abilene shows what happens between columns. Phase one was built and is running: an independent tally by Epoch AI in April put roughly four of eight buildings live, about 0.3 gigawatts, with the full 1.2 gigawatts projected for late this year. The 600 megawatt expansion flagged in the September 2025 announcement was reported dropped in March 2026. Oracle publicly rejected that reporting, though the dispute concerned the expansion tranche rather than the 4.5 gigawatt agreement. Crusoe then announced a 900 megawatt campus next door for Microsoft. The capacity did not disappear. It changed tenant.

Memphis shows all four columns at once. SpaceX, which acquired xAI in February 2026, has roughly 300 megawatts of grid supply approved by the Tennessee Valley Authority across two board decisions. The Shelby County air permit for the first Colossus site covers 15 turbines and 247 megawatts and expires in January 2027. SpaceX said on July 30 that it is currently operating 69 gas turbines across the Colossus sites and will not remove the unpermitted ones until July 2027. Mississippi permit filings describe a permanent 1.2 gigawatt plant of 41 turbines.

Generation is not data center load, and the two should never be added together or compared like for like. What the Memphis numbers show is a gigawatt arriving through self-build and permit fights rather than through a utility contract.

None of this makes the gigawatts fake. The unit simply became a planning language before it became a delivery record.

Who Carries The Cost, And Who Gets To Know

Ratepayer exposure is now the political center of AI power. Somebody pays for generation, interconnection, transmission reinforcement and backup. If that cost leaks into the general customer base, AI demand becomes a household electricity problem, and the politics turn fast.

Both federal sites invoke the White House ratepayer protection pledge. Savannah River adds the solicitation's own-cost obligation, which is pre-contract evidence. Paducah adds something harder to walk away from: a Kentucky Public Service Commission proceeding, which produces a record and a decision.

Louisiana is furthest along and most contested. Entergy Louisiana's March 2026 agreement with Meta is structured so that Meta pays its full cost of service, with about $2 billion in projected customer savings over 20 years on top of $650 million previously announced. That figure is an Entergy projection, not a regulatory finding.

The generation behind it sits in three stages at the same time. Three plants were approved in August 2025, of which two broke ground in Richland Parish in December 2025 and are due online in late 2028. Seven more, totaling more than 5,200 megawatts, were agreed with Meta and filed. The commission voted 4-1 in April to fast-track the schedule, not to approve the plants, and a decision is targeted for December.

The disclosure fight is now the live one. An administrative law judge referred a subpoena dispute to the full commission, reported on July 31, and the commissioners can take it up on August 12. Intervenors say they cannot put a number on ratepayer exposure because so much of the filing is designated confidential. Entergy says that characterization is false.

Congress is moving on the same question from the other end. The Ratepayer Protection Act cleared House Energy and Commerce 52-0 on July 21. It would require state utility commissions to consider a standard under which data center customers of 100 megawatts or more carry the full incremental cost of the upgrades they trigger, with financial assurances against stranded costs.

The Countercase: The Ladder Converts

The strongest argument against all of this is Meta.

Richland Parish went from a plan of roughly 2,600 megawatts to a declared 5 gigawatt target to a utility request of more than 5,200 megawatts, with a signed full-cost-of-service agreement around it and two plants in construction since December 2025. That is what conversion looks like, and it happened in about eighteen months.

Abilene complicates the cautionary case differently. Phase one is built and occupied. The expansion that fell through was picked up next door by a rival hyperscaler within weeks. Announced capacity moved sideways rather than evaporating.

If Stargate's planned gigawatts convert at even part of Louisiana's rate, today's skepticism will look static. Utilities, turbine makers, gas suppliers, nuclear developers and state regulators would be facing a construction wave rather than a press cycle.

The narrower reading survives that. Conversion is real at the regulated end, where a commission, a tariff and a cost allocation force the question into the open. It stays unproven at the announcement end, where nothing forces it at all.

Three dates in the next two weeks will move capacity between columns, which is more than any of these announcements did. Amentum reports third-quarter results on August 11, the first scheduled moment anyone can ask what Savannah River costs. The Louisiana commission can take up the Meta disclosure fight on August 12. The Tennessee Valley Authority board meets on August 20, with a proposal expected to strip data centers of the industrial manufacturing rate.

None of the three will announce a gigawatt. Each of them decides who is obliged to build one.