AI Infrastructure Now Arrives In Gigawatts, But Announced, Permitted, Contracted And Delivered Capacity Are Four Different Things.

Washington recently unveiled two separate gigawatt-scale artificial intelligence (AI) power projects within a span of nine days, yet neither announcement carries a firm obligation for anyone to actually construct the proposed facilities. This disparity highlights a critical distinction in the rapidly evolving landscape of AI infrastructure development: the vast difference between an announced capacity and one that is truly delivered.

On July 20, the National Nuclear Security Administration (NNSA) made headlines by announcing it had selected Amentum to enter negotiations for a phased lease at its Savannah River Site in South Carolina. This ambitious project envisions a 1 gigawatt (GW) AI data center, supported by approximately 2 GW of dedicated on-site generation. Just nine days later, the Department of Energy (DOE) followed suit, naming Brookfield and NextEra Energy as its partners for the Paducah site in Kentucky. This partnership was touted with an eye-watering figure of over $100 billion in announced private investment, underpinning a 1.8 GW campus.

While both press releases prominently feature impressive gigawatt figures, the substance behind these numbers varies significantly. The term "gigawatt" itself remains constant, but the commitment to its delivery changes dramatically depending on its stage of development. Capacity can be merely announced, which effectively commits no party to any specific action. It can be permitted, meaning regulatory bodies have granted approval for construction, but there’s no inherent obligation to build. More concretely, it can be contracted, indicating that a party has agreed to financial terms for its development. Finally, and most tangibly, it can be delivered, signifying that the infrastructure is fully built, operational, and actively supplying power. These four stages represent distinct levels of obligation, even though the numerical value—a gigawatt—remains identical in all cases. This crucial distinction now forms the core narrative of the AI power story.

The Savannah River project, for instance, is currently in the "announced" stage, with NNSA’s statement clearly indicating a "selection to enter negotiations," not a final lease award. Amentum, the selected partner, also noted that the project would be "initially powered by natural gas bridging to advanced nuclear energy," with DC BLOX confirmed as the digital infrastructure partner. However, details such as investment figures, rental terms, specific phasing schedules, or concrete construction dates remain undisclosed by any of the involved parties. The September 2025 solicitation for the Savannah River site had set a demanding standard, requiring the chosen partner to "build, operate and decommission the project at its own cost" and secure its own utility interconnection agreements—a significant obligation, but still a step removed from a binding commercial settlement.

The Paducah project, while appearing more structured on paper, faces similar commercial uncertainties. The arrangement designates Brookfield to develop and operate the campus, while NextEra Energy is responsible for building and owning the generation assets. Three Kentucky utilities were also mentioned in the release. The proposed energy package includes roughly 2 GW of new grid-connected natural gas generation, supplemented by up to 2.6 GW of battery storage, with an estimated completion date of 2031. Crucially, the Department of Energy specified that the power service agreement "remains subject to approval by the Kentucky Public Service Commission," and the entire transaction still requires definitive documentation. Adding to the complexity, the DOE describes a 1.8 GW campus, while the coalition’s own announcement cites a compute load above 1.2 GW, supported by 1.8 GW of utility capacity, and up to 4.6 GW of dedicated generation by its full build-out in 2032. These discrepancies underscore the fluid nature of announced figures. Furthermore, the absence of a named anchor tenant for a campus carrying a $100 billion headline signals another layer of commercial incompleteness.

These two projects are part of a broader four-site federal program announced by the Department of Energy in July 2025, which also includes the Idaho National Laboratory and the Oak Ridge Reservation. To date, only Paducah and Savannah River have been matched with partners. Idaho’s first applications were due in November 2025, with rolling admissions thereafter, while Oak Ridge’s proposals were due on December 1. Neither of these sites has seen a publicly announced selection since. The federal program isn’t necessarily failing; rather, the sites were never intended as a uniform package. Idaho and Oak Ridge left capacity parameters to applicants, with Oak Ridge even offering two separate parcels. Interestingly, Paducah’s solicitation initially prioritized small modular reactors (SMRs), microreactors, geothermal, and underground thermal storage, yet the winning bid arrived with a more conventional solution of gas and batteries.

Beyond these four sites, the DOE has demonstrated faster progress elsewhere. In March, it announced a substantial 10 GW campus at the Portsmouth site in Ohio, partnering with SoftBank’s SB Energy and AEP Ohio. This project is backed by 9.2 GW of new gas generation, with construction on an initial 800-megawatt phase anticipated to begin this year. The DOE itself points to the Portsmouth model as a precedent for the Kentucky deal, suggesting it represents a more advanced stage of commitment. Thus, the federal program currently features one site (Portsmouth) with impending construction, two sites (Savannah River and Paducah) with named partners but no signed leases, and two sites (Idaho and Oak Ridge) with no public announcements at all. The federal government contributes land, a structured solicitation process, and political priority, but developers are still responsible for engineering, interconnection, capital, power contracts, and timelines. A site announcement merely starts the clock; it doesn’t close the gap.

The private sector exhibits a similar dynamic, often at an even larger scale, with the terminology evolving as projects progress. OpenAI’s "Stargate" initiative serves as a prime example. In July 2025, OpenAI and Oracle announced an "agreement to develop 4.5 gigawatts" of additional Stargate capacity. By September, this figure shifted to "up to 4.5 gigawatts" within a broader announcement of "nearly 7 gigawatts of planned capacity and over $400 billion in investment." A month later, Michigan’s involvement pushed the running total to "over 8 gigawatts and more than $450 billion." By April 2026, OpenAI claimed to have "already passed its 10 gigawatt commitment," having added over 3 GW in just 90 days, with no public updates since. These are corporate statements about future plans and agreements, not independently verified utility capacities.

The Abilene, Texas, Stargate site illustrates the journey between these "columns." Phase one has been built and is operational; an independent assessment by Epoch AI in April estimated roughly four of eight buildings were live, accounting for about 0.3 GW, with the full 1.2 GW projected by late this year. However, a 600-megawatt expansion, initially flagged in the September 2025 announcement, was reportedly dropped in March 2026. While Oracle publicly disputed this report, the core of the dispute concerned the expansion tranche rather than the overarching 4.5 GW agreement. Interestingly, Crusoe subsequently announced a 900-megawatt campus "next door" for Microsoft, indicating that the capacity didn’t vanish but merely changed tenants, highlighting the fluid nature of demand and supply in this sector.

Memphis, Tennessee, offers a vivid illustration of all four columns simultaneously. SpaceX, which acquired xAI in February 2026, has secured approval for approximately 300 megawatts of grid supply from the Tennessee Valley Authority (TVA) across two board decisions. The Shelby County air permit for the initial "Colossus" site covers 15 turbines and 247 megawatts, expiring in January 2027. However, SpaceX admitted on July 30 that it is "currently operating 69 gas turbines" across its Colossus sites, and critically, it will not remove the unpermitted ones until July 2027. Further complicating the picture, Mississippi permit filings describe a permanent 1.2 GW plant comprising 41 turbines. These Memphis numbers demonstrate a gigawatt of capacity materializing through a combination of self-build efforts and direct permit battles, rather than through a traditional utility contract. It’s important to note that generation capacity is distinct from data center load and should not be directly added or compared. While none of this renders the gigawatts "fake," it underscores that the unit has become a language of planning before it translates into a record of actual delivery.

The political and economic implications of AI power are increasingly centered on "ratepayer exposure." The construction, interconnection, transmission reinforcement, and backup power all come with costs. If these costs are disproportionately borne by the general customer base, AI demand could rapidly transform into a household electricity problem, swiftly turning public and political sentiment against such developments. Both federal site announcements specifically invoke the White House’s ratepayer protection pledge. Savannah River includes a pre-contractual "own-cost obligation" in its solicitation, while Paducah’s reliance on Kentucky Public Service Commission (PSC) approval ensures a public record and a regulatory decision on cost allocation.

Louisiana stands as a prime example of this contested terrain. Entergy Louisiana’s March 2026 agreement with Meta is structured to ensure Meta pays its full cost of service, with Entergy projecting approximately $2 billion in customer savings over 20 years, in addition to $650 million previously announced. However, this figure remains an Entergy projection, not a definitive regulatory finding. The generation capacity backing this deal is progressing through multiple stages concurrently. Three power plants were approved in August 2025, with two of them breaking ground in Richland Parish in December 2025 and slated for online operation in late 2028. An additional seven plants, totaling over 5,200 megawatts, were subsequently agreed upon with Meta and filed for approval. While the Louisiana PSC voted 4-1 in April to fast-track the schedule, this was not an approval of the plants themselves, with a final decision targeted for December. The disclosure fight is now paramount: an administrative law judge referred a subpoena dispute to the full commission on July 31, with intervenors arguing they cannot accurately assess ratepayer exposure due to extensive confidential filings. Entergy maintains this characterization is false.

Congress is also addressing this issue from a legislative angle. The Ratepayer Protection Act cleared the House Energy and Commerce Committee with a unanimous 52-0 vote on July 21. If enacted, this legislation would mandate state utility commissions to consider a standard under which data center customers consuming 100 megawatts or more would bear the full incremental cost of any grid upgrades they necessitate, along with financial assurances against stranded costs.

Despite the skepticism surrounding announced capacities, a strong counter-argument lies in the fact that the "ladder converts." Meta’s project in Richland Parish, Louisiana, serves as a compelling example. What began as a plan for roughly 2,600 megawatts evolved into a declared 5 GW target, culminating in a utility request for more than 5,200 megawatts, all underpinned by a signed full-cost-of-service agreement and with two plants already under construction since December 2025. This demonstrates a real and rapid conversion from announced capacity to delivered infrastructure, occurring within approximately eighteen months. The Abilene case further complicates the cautionary narrative; while an expansion planned by Oracle/OpenAI did not proceed, the capacity was swiftly picked up by a rival hyperscaler, Crusoe, for Microsoft, indicating that announced capacity can shift sideways rather than simply evaporating.

If the myriad "planned gigawatts" of projects like Stargate convert at even a fraction of the rate seen in Louisiana, the current skepticism will quickly seem outdated. Utilities, turbine manufacturers, gas suppliers, nuclear developers, and state regulators would then be facing an unprecedented construction boom rather than merely a cycle of press releases. However, a more nuanced reading of the situation still holds: conversion is demonstrably real at the regulated end of the spectrum, where the scrutiny of a commission, a defined tariff, and explicit cost allocation force the issue into the open. It remains largely unproven at the pure "announcement" end, where no such binding mechanisms exist.

The coming weeks promise to move more capacity between these critical columns than any recent announcement. Amentum is scheduled to report its third-quarter results on August 11, which will be the first opportunity for stakeholders to inquire about the financial specifics and true costs associated with the Savannah River project. The Louisiana commission is set to address the contentious Meta disclosure dispute on August 12, a decision that will shed light on the transparency and accountability of large-scale energy deals. Finally, the Tennessee Valley Authority board will convene on August 20, with a proposal expected that could strip data centers of their preferential industrial manufacturing rate, potentially altering the economic landscape for future developments in the region. None of these upcoming events will announce new gigawatts, but each will play a pivotal role in determining who is truly obliged to build them.