As data-center developers add on-site generation to bridge utility delays, the grid queue is being joined by queues for turbines, air permits, gas infrastructure, storage, and EPC capacity.
For two years, the AI infrastructure contest centered on scheduled, contracted, deliverable grid power. A power purchase agreement was never the same thing as an energized campus, but it became the shorthand: whoever could reserve the largest blocks for the longest periods held the advantage. As SCN argued, that power moat was largely already held by hyperscalers with the balance sheets and project pipelines to secure it.
That contest has not ended. It has acquired an upstream queue.
JLL projects nearly 100 GW of new global data-center capacity between 2026 and 2030, effectively doubling the installed base. Enverus puts grid-interconnection timelines at five to six years in key markets and says projects with 2025 first-power dates spent an average of more than 2,100 days in the queue. Developers are therefore no longer relying exclusively on utility schedules. They are funding dedicated generation, storage, and electrical infrastructure to bring compute online before permanent grid service arrives.
Behind-the-meter generation can bypass the grid for first power. It does not necessarily bypass it forever. Many campuses still anticipate a later utility connection, and self-supply replaces one critical path with several others: an air permit, a gas connection, an equipment reservation, storage and controls, EPC capacity, financing, and eventually a grid position.
That is the real change in 2026. The scarce asset is no longer only a contracted megawatt. It is control of the entire power-delivery stack.
One of the clearest expressions of that strategy sits on more than 8,000 acres in Pecos County, Texas, approximately 17 miles north of Fort Stockton. Amazon has confirmed acquiring the GW Ranch data-center campus, while Pacifico GW LLC—not Amazon—is the named applicant in the state air-permit record.
The project’s capacity requires careful qualification. The Texas Commission on Environmental Quality determination describes a natural-gas-fired plant with 35 simple-cycle turbines and a nominal output of 5,000 MW. Pacifico markets the permitted configuration as having 7.65 GW of gross gas-generation capacity and says the wider campus is planned to include 1.8 GW of battery storage and up to 750 MWac of solar. The public documents do not explain the difference between the 5-GW nominal figure and the 7.65-GW gross figure, so they should not be treated as interchangeable.
The initial configuration is a private power system. TCEQ’s determination says the plant will serve an on-site AI data center and will not be capable of buying from or selling to the local utility distribution system. Amazon, however, says the campus is designed to transition to grid-connected service and potentially add supply as interconnection timelines allow. The bypass is therefore a first-power strategy, not necessarily the permanent architecture.
Two qualifiers remain decisive. GW Ranch is permitted, not built. Pacifico says it has secured turbines and expects first power from an initial 1-GW phase in the first half of 2027, but that is a developer forecast rather than an operating milestone. The distance between an air permit, a reserved machine, and a running plant is exactly where the new power thesis will be tested.
A project targeting first power in 2027 cannot solve its problem with a large-frame turbine delivered in 2031. Near-term bridge power and permanent multi-gigawatt generation are related markets, but they are not the same market.
For projects seeking service within the next 18 to 36 months, Enverus points to fuel cells, fleets of reciprocating engines, and aeroderivative turbines. These modular technologies can be installed incrementally, carry lower single-unit outage risk, and may be available faster than large combined-cycle systems. Their trade-offs include fuel cost, maintenance, efficiency, emissions, and the risk that temporary equipment remains in service much longer than planned.
The distinction is visible in an August 7 agreement announced by Energy Vault. The company will supply battery energy storage, grid-forming power-conversion systems, and control software for an initial 1.25 GW of integrated power infrastructure backed by an unnamed hyperscaler customer in Texas. A national EPC partner will provide the generation layer, including Caterpillar gensets. Energy Vault says the first deployments are expected within four to 12 months. The customer and EPC contractor remain undisclosed.
The 1.25-GW figure describes the integrated deployment, not necessarily the battery system’s power rating, and no battery duration in megawatt-hours has been disclosed. Gas engines are already dispatchable; the batteries are not what makes them firm. Their role is to manage fast workload transients, voltage and frequency stability, ride-through, generator cycling, fuel efficiency, and the transition to an eventual utility connection.
Bridge generation buys speed. Large-frame turbine reservations buy scale.
Confusing the two produces the central contradiction in many AI power announcements: a company promises compute in 2027 while citing equipment supply that does not arrive until the following decade.
For permanent gas generation at multi-gigawatt scale, the equipment queue is real.
GE Vernova’s second-quarter 2026 results showed 116 GW of gas-turbine capacity under contract, consisting of 53 GW in backlog and 63 GW covered by slot-reservation agreements. That is not a 116-GW backlog, and it is not all attributable to AI: GE said approximately 20% of the contracted capacity was for data centers, with the balance spread across utilities and other customers in 26 countries.
The delivery signal is still stark. GE has agreements signed into 2031, describes itself as mostly sold out through 2030, and expects more than half of its 2031 production capacity to be contracted by the end of 2026. It is expanding toward 30 GW of annual gas-turbine output in 2030 and says it is already in active customer discussions for 2032 and beyond.
Siemens Energy’s August 2026 investor presentation listed 69 GW of gas-turbine backlog and another 26 GW under slot-reservation agreements. Siemens said it had booked 15 GW and shipped 6 GW during the reported period, while continuing manufacturing expansion through fiscal 2030.
Mitsubishi Power describes the same market from the other side. Its head of EMEA told Platts that lead times have stretched to five years or more and that the company is signing contracts for deliveries between 2031 and 2034. He characterized the bottleneck as the capacity of the whole supply chain, rather than a lack of demand or capital.
That last qualification matters. The equipment queue is not a single assembly line inside a turbine factory. It runs through castings, forgings, specialized machinery, trained labor, electrical equipment, power-island integration, and EPC contractors capable of assembling the system. GE, for example, is adding machinery and labor while ramping casting and forging capacity for later-decade production.
Going off-grid was supposed to route around a publicly administered queue. At large scale, it creates another queue—privately controlled by manufacturers and their suppliers.
A 2031 reservation can therefore become a strategic asset. But it is not the equivalent of an operating plant, just as a signed PPA is not the equivalent of an energized data center. It is an option on one critical component of a much larger project.
The market is not converging on a single gas-plus-storage template.
GW Ranch represents a dedicated private-grid campus intended to supply its own load initially and connect to the public system later.
The Energy Vault agreement represents modular bridge infrastructure intended to energize compute before a permanent connection is available.
PPL and Blackstone’s Invitium Energy venture represents something different again: new front-of-the-meter generation connected through PJM and contracted to large loads under long-term energy supply services agreements. It is not a behind-the-meter project.
PPL and Blackstone formed the 51/49 venture in July 2025 to develop combined-cycle generation in Pennsylvania. In its second-quarter 2026 update, PPL said Invitium had secured sites capable of supporting 8 to 14 GW, obtained reservation agreements for more than 5 GW of combined-cycle turbines, and had more than 5 GW of generation-interconnection requests accepted by PJM. The turbine reservations represent a potential $12.5 billion to $15 billion of investment through 2032.
Those figures remain development milestones, not operating capacity. PPL says Invitium will not begin construction or make material financial commitments until it has signed contracts with appropriate risk protections or secured reimbursement for its costs. It expects the combined-cycle plants to contribute meaningful earnings only when they begin operating, potentially in 2031 or 2032.
Invitium is important because it shows that turbine scarcity is not merely a side effect of the off-grid turn. The same equipment queue constrains the utilities and independent generators trying to expand the public system. AI developers are competing for turbine capacity against the grid projects they are waiting for.
Equipment supply is only one critical path. Air permitting can be another.
GW Ranch is not an example of a minor-source permitting strategy. TCEQ’s determination says the project triggers federal Prevention of Significant Deterioration review for carbon monoxide, nitrogen oxides, particulate matter, sulfur dioxide, volatile organic compounds, sulfuric acid mist, and greenhouse gases. The permit materials calculate allowable greenhouse-gas emissions of approximately 33.2 million tons of carbon-dioxide equivalent per year. That is an upper limit based on the permitted configuration, not a forecast of actual annual emissions.
Other Texas projects have taken a different route. A Floodlight analysis found that at least 38 data centers had received minor permits for on-site power sources since 2024, collectively authorizing more than 2,100 backup diesel generators. More than half of the sites reviewed reported nitrogen-oxide estimates just below thresholds that would have triggered public participation and more extensive review.
That strategy is now being challenged. On July 22, the Environmental Integrity Project, Sierra Club, and Public Citizen served Vantage Data Centers and VoltaGrid with a 60-day notice of intent to sue over two San Antonio campuses. The groups allege that the data centers, gas plants, and backup generators were improperly divided into separate minor sources. The notice is a prerequisite to a Clean Air Act citizen suit; it is not yet a filed lawsuit.
David Spence, an energy-law professor at the University of Texas at Austin, told Inside Climate News that a gas-turbine data center would probably qualify as a major source unless it operated at a very low capacity factor, making a piecemealed permitting theory difficult to defend.
The xAI dispute in Mississippi involves another legal question: whether trailer-mounted turbines can remain outside stationary-source permitting requirements. The schedule logic is similar to the strategy SCN examined in xAI’s attempt to outrun grid and regulatory timelines, but the permitting posture is materially different from GW Ranch’s major-source review.
A proposed EPA rule could eventually give states more discretion over public participation in minor New Source Review programs. It would not remove public-participation requirements for major PSD permits, and it would not rewrite existing state programs immediately. States choosing to reduce minor-source participation requirements would still need to change their implementation plans and obtain EPA approval. The proposal is therefore a potential procedural tailwind, not an instant permit shortcut.
Texas has added another moving part. On August 3, Gov. Greg Abbott directed the Public Utility Commission of Texas and ERCOT to audit every data-center project advancing through ERCOT’s interconnection process before additional projects are approved. The directive calls for disclosures covering financing, ownership, projected power and water consumption, on-site generation, cooling technology, and community impacts. Abbott cited approximately 474 GW of requests to connect to ERCOT, about 90% of them associated with data centers.
The directive is aimed at projects seeking ERCOT interconnection, not necessarily campuses operating entirely on private generation. But a project such as GW Ranch, which anticipates a future grid connection, cannot treat that future connection as a fixed or optional administrative detail. The rules governing access are still changing.
Suppose the equipment arrives and the permits hold. The economics still have to survive.
Behind-the-meter power introduces two linked assets: a generation plant dependent on the data center as its customer, and a data center dependent on the plant for its operating schedule. A delay, cancellation, capacity reduction, fuel interruption, or early utility connection on either side can strand capital on the other.
A&O Shearman identifies stranded-asset exposure, power-supply interruption, construction delays, tariff arrangements, licensing uncertainty, and the interdependence of the generation and data-center assets as central lender concerns. That does not prove lenders are universally charging a measurable turbine premium. It does show what project-finance structures now have to protect against.
PPL’s approach is one institutional response: reserve sites, enter the interconnection queue, and secure turbine positions, but do not release material construction capital until a creditworthy customer contract or reimbursement arrangement is in place.
The systemic cost is more complicated. The Harvard Belfer Center warns that co-location can pull existing grid-facing generation behind a private fence, reducing capacity available to the wider system. That concern applies most directly when an existing plant is redirected from public service. A newly built on-site plant does not remove pre-existing grid capacity, but questions remain over backup service, transmission investment, standby charges, and whether tariffs allocate those costs to the customer that caused them.
Climate exposure is also material. Actual emissions depend on turbine design, operating hours, efficiency, and the mix of power displaced or supplemented. But a permitted upper limit of 33.2 million tons of carbon-dioxide equivalent per year makes emissions more than a reputational footnote at GW Ranch. It is a financing, permitting, and corporate-commitment risk.
Nuclear remains a long-term, firm, low-carbon option, but it is not the same near-term bypass. A new commercial nuclear plant in the United States requires Nuclear Regulatory Commission approval, environmental review, public processes, financing, and construction. It does not solve a first-power requirement scheduled for 2027.
That is why the likely outcome is not an industry-wide choice between grid power, gas, storage, renewables, or nuclear. The largest operators will pursue several in parallel and use each for a different time horizon.
On-site generation does not eliminate the power queue. It multiplies it.
Bridge assets can buy speed. Large-frame turbine reservations can buy future scale. Air permits determine whether either can run. Gas pipelines determine whether they can run reliably. Storage and controls determine whether the power system can support the dynamics of an AI cluster. Customer contracts and financing determine whether the plant gets built. Grid rights often remain part of the permanent architecture.
The next AI power moat is therefore not a turbine slot alone. It is a coordinated stack of deliverable assets:
A 2031 turbine reservation is strategically valuable because so few are available. But it is valuable only if every other part of the system arrives with it.
The operators that control the stack will energize first. The ones that control only a PPA—or only a turbine slot—may still be waiting.
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
