Datacenter developers have announced about 90 GW of behind-the-meter generation across 59 US projects, and about 2 GW of it runs, which is 2.2%. Another 1.2% is under construction, 36% holds an air permit, and 60% exists as a press release. The same month EIA published its July 2026 planned-generator inventory: every gas turbine, combined-cycle block and engine plant any owner in the country has told EIA it intends to build comes to 68.1 GW, and 11.0 GW of that is under construction. The announced private fleet is bigger than the visible national one. This piece rebuilds the 90 GW from permits, filings and EIA’s unit list, prices the two gates that decide who energizes, and forecasts what runs by the end of 2028 using the Open Factory Lead-Time Monitor.
We cover the source of the 90 GW and what EIA sees of it, the nineteen campuses we could trace to a document, xAI’s 1,498 MW as the benchmark, the GSU on every frame-turbine critical path, the permits that stopped two Stargate sites, DOE 202(c) and FERC, and our forecast through 2028, with winners and payers behind the wall.
Source: Open Factory Lead-Time Monitor (compiled from EIA-860M July 2026, company filings and permits listed in data/btm-pipeline.csv; as of August 2026)
For 2028, EIA lists 2.9 GW of gas under construction, 6.9 GW approved or awaiting permits and 6.6 GW planned with no approvals, while the named behind-the-meter campuses we could document for that year and which are absent from EIA’s list add 10.1 GW on top; for 2027 the split is 3.7, 1.3, 6.4 and 8.3 GW.
Where 90 GW Comes From
The number is Cleanview’s. Its May 2026 report, built from air permits, SEC filings, utility dockets and satellite imagery, identified 59 datacenters with about 90 GW of planned on-site generation, 92% of it announced since the start of 2025. Its February 2026 edition had counted 46 projects and 56 GW, with equipment deals identified at about two thirds of them and 75% of the identified equipment (23 GW) gas-fired. The full dataset is a $4,900 report; we cite the free summary and newsletter posts. Global Energy Monitor arrives at the same order of magnitude from a different direction: the US had 252 GW of gas capacity in development at the start of 2026, more than a third of it slated to power datacenters on site, with Texas alone at 80.6 GW of which 40 GW is for datacenters. Two trackers, two methods, 84 to 90 GW.
Source: Cleanview, Bypassing the Grid (free summary), May 2026
Cleanview’s own split of the 90 GW is 54 GW announced or early-stage, 32.4 GW permitted, 1.1 GW under construction and 2.0 GW operating, and it expects 2.8 to 3.2 GW operating by the end of 2026.
EIA can see 16 GW of the 90. We went through all 487 planned natural-gas generator rows in the July 2026 EIA-860M workbook (turbines, combined-cycle parts and engines, 68.1 GW) and matched owners and plant names against the datacenter campuses in the press and the permit record. Two campuses are there: Fermi America’s Project Matador in Carson County, Texas, at 11,679 MW across 137 generator rows, 420 MW of it with approvals received and the rest at status P, planned with approvals not initiated, and Homer City in Pennsylvania at 4,396 MW across 14 rows, all status P, with online dates from July 2028 to November 2029. Nothing else in the file is a datacenter campus by name. Colossus, Stargate Abilene, Stargate Frontier, Project Kilby, GW Ranch, Monarch, the Meta plants in New Albany: none has a row. EIA-860M is a survey of generators 1 MW and larger and its planned sheet depends on the owner filing; mobile turbines that arrive on trucks, engine packs owned by a power-as-a-service provider and plants that have not passed a board’s final investment decision tend not to be filed. The gap is the measurement problem of this whole market, and it is why the Lead-Time Monitor carries a behind-the-meter pipeline table from permits rather than waiting for EIA.
The mix in EIA’s file also says something about who is actually building. Of the 68.1 GW, MISO holds 20.6 GW and ERCOT 16.7 GW, and the largest single owner after Fermi is Entergy Louisiana at 9.5 GW. Entergy’s Q2 2026 10-Q describes seven new combined-cycle units totaling 5,278 MW at about $12.9 billion, four of them near a Meta site in Richland Parish, with in-service dates in 2030 and 2031, on top of the earlier Franklin Farms and Waterford 5 units for the first Meta campus. That is dedicated datacenter generation built by a utility and rate-based, the front-of-meter cousin of everything in this piece, and it lands two to three years after the private plants promise to. The Large-Load Cost Table treats utility-built dedicated generation as its own category for exactly that reason.
The Nineteen We Could Document
We built the pipeline from documents, not announcements: an air permit, an SEC filing, a utility 10-Q, a state agency letter or a company release with equipment counts. Nineteen campuses met the bar. They add to 45 GW, half of Cleanview’s 90, and the other half is either in Cleanview’s paid rows or in announcements we could not tie to a filing. The full table with URLs is data/btm-pipeline.csv.
Source: Open Factory Lead-Time Monitor
Of the nineteen, three are operating (Colossus 1 and 2, Vantage’s Virginia site, and the southern half of Williams Socrates), two more are under construction for 2026 (Socrates North and the EdgeConneX New Albany plant), and the remaining thirteen are permits, orders and PPAs targeting 2027 to 2029, with 7.65 GW of the total in a single Amazon permit.
The operating benchmark is xAI. Cleanview puts Colossus 1 and 2 outside Memphis at 1,498 MW of operating gas-turbine capacity, with nearly 60 turbines visible on satellite and foundations laid for more. The Southaven, Mississippi half runs 27 portable turbines totaling 495 MW under the Mississippi Department of Environmental Quality’s position that portable units are mobile sources exempt from a stationary air permit; the NAACP sued under the Clean Air Act in April 2026 and the Department of Justice moved to intervene on xAI’s side in June. On the Tennessee side the Shelby County Health Department approved a permit for 15 turbines on July 2, 2025, which the NAACP and the Southern Environmental Law Center appealed. NGI reports that VoltaGrid’s engine packs and grid supply took Colossus from start to finish in 122 days. That is the speed everyone else is selling against: 1.5 GW in under two years, built from leased and mobile units, with the permits litigated afterwards.
The largest documented projects are frame-turbine plants that have not broken ground on generation. Chevron’s Project Kilby will supply Microsoft’s Pecos, Texas campus with about 2.67 GW under a 20-year PPA signed June 22, 2026, a majority from large GE Vernova turbines plus Solar Turbines units, with FID expected by the end of 2026 and first power in 2028; Chevron’s New Energies president told analysts the machines are seven GE Vernova 7HA.02s and twelve Caterpillar Titan 350s, that the PPA is take-or-pay, and that “only one” announced behind-the-meter project “is at multi-gigawatt scale” with a customer commitment. Thirty miles east, Amazon confirmed it acquired the GW Ranch site in Pecos County, where Pacifico Energy’s TCEQ prevention-of-significant-deterioration permit, issued January 21, 2026, covers 35 turbines at 7.65 GW nameplate and about 5 GW nominal delivered output, and where land clearing began in August 2026. Amazon also confirmed talks with Homer City’s developers the same week. Two Pecos County campuses and one Pennsylvania campus hold 14.7 GW of frame-turbine plans and 42 large turbines that must come out of the queues we measured in The $60 Million Deposit.
Fermi America is the only one of the giants that has filed with EIA, and its own filings are the best public record of what a behind-the-meter plant actually requires. The 10-K lists seven leased GE TM2500 mobile units rated about 132 MW, six Siemens SGT-800s, three Siemens SGT6-5000F heavy-duty turbines at about 240 MW each in simple cycle, three GE 6B frame units, “and certain substations, switchgear, transformers, and related balance-of-plant components”, with TCEQ’s approval of a roughly 6 GW air permit on February 25, 2026 and an application for 5 GW more filed March 27. By August 2026 the company had 1.5 GW of turbines landed, a 222 MW lease with TensorWave worth about $6.5 billion over 15 years, 86 MW of grid supply from Xcel’s SPS expected in the second half of 2026 and 114 MW more contracted by October 1, 2027, and a target of “about 200 megawatts of initial commercial power over the next six months, and about 1.5 GW over the next 18 to 24 months”. Hillcore will build, own and operate about 2.6 GW more under a ground sublease, with first power of about 350 MW targeted within 24 months of notice to proceed. Cleanview’s satellite series shows the site at least a year behind the schedule in the IPO filing, with no vertical construction between February and April 2026. One detail for anyone reading EIA’s file: 860M codes Matador to ERCOT, while Fermi’s 10-Q says the campus “resides in the Southwest Power Pool,” outside the reach of Governor Abbott’s August 2026 ERCOT audit; the Interconnection Cost tool treats the two as different regimes.
The GSU Nobody Prices
Every press release in the pipeline names the prime mover and none names the transformer. A frame turbine or an aeroderivative generates at 11.5 to 15 kV. To move hundreds of megawatts across a campus, or to a 138 kV or 345 kV collector bus where the grid will eventually be tied in, that output goes through a generator step-up transformer, one per unit or one per pair, sized at 100 to 500 MVA. Wood Mackenzie’s survey put the average GSU lead time at 143 weeks in Q2 2025, with GSU demand up 274% since 2019 and a projected shortfall of more than 700 units in 2025; Terrapin’s June 2026 buyer note has 50 to 150 MVA GSUs at 100 to 150-plus weeks and 15 kV metal-clad switchgear at 60 to 80; SemiAnalysis’s grid piece says gas turbine and GSU lead times “have each extended to three to four years, versus a historical norm of roughly 18 months”. We covered the transformer side of this in Transformer Lead Times Hit 143 Weeks; here the point is where the GSU sits relative to the machine it serves.
Source: Open Factory Lead-Time Monitor (compiled from Wood Mackenzie Oct 2025, Terrapin Jun 2026, GE Vernova and Siemens Energy statements, Caterpillar via Power Engineering, NGI, Bloom 10-K; as of August 2026)
A new frame turbine is 130 to 208 weeks, a GSU 100 to 156 with 143 typical, a new large reciprocating engine 78 to 130, MV switchgear 44 to 104, an aeroderivative or mobile fleet 17 to 78, and a fuel-cell block 8 to 78 weeks from order to delivery.
A recip plant does not need a GSU to energize, and that is why engines are winning the first two years. A Jenbacher J624 or a Caterpillar G3520 generates at 4.16 to 13.8 kV; the 25 MW packs VoltaGrid ships put that output straight onto medium-voltage switchgear and distribute across the campus at 13.8 or 34.5 kV through pad-mount unit transformers, the class the Transformer Sizing and Cost tool prices at 52 weeks or less. The step-up to transmission voltage happens only when the site chooses to tie into the grid, which most say they will do later. A 7HA.02 at 384 MW cannot do that; it needs a 400-plus MVA GSU on the day it fires, and that unit was ordered, if it was ordered at all, in the same quarter as the turbine. Fermi is the only developer in the pipeline whose filings say the substations and transformers were bought alongside the turbines, and it raised a dedicated equipment facility “intended to support the procurement of long-lead electrical infrastructure required for Project Matador, including substations, transformers, breakers, and related high-voltage switchgear”. Chevron’s Kilby release mentions “associated electrical infrastructure” and no date. For every frame-turbine campus targeting 2028, ask when the GSU purchase order was placed; if the answer is after mid-2025, 143 weeks puts delivery in 2028 at the earliest, before installation, oil fill and commissioning. The Lead-Time-Adjusted Schedule does that arithmetic for a bill of gear; the Datacenter MEP Cost Table carries the GSU as its own line for on-site generation because no vendor quotes it inside the turbine package.
The turbine side of the same gate is the one the market talks about. GE Vernova ended June 2026 with 116 GW of gas equipment under contract, 63 GW of it slot reservations, and booked 52 heavy-duty and 61 aeroderivative units in the quarter while shipping 3 GW; Siemens Energy’s CFO put lead times at “three years or more”. Jefferies’ February 2026 note, as reported by Latitude Media, has large turbines more than 90% booked for 2028 and 70% for 2029, while behind-the-meter systems are about 65% and 33% booked for the same years, with at least 19 GW of available equipment capacity by 2028. The same note flagged that “high-voltage equipment, such as transformers and switch gears, could see a supply crunch.” That is the GSU.
The Permit Is the Second Gate
The equipment queue is one gate; the air permit is the other, and in 2026 it stopped more megawatts than any turbine shortage. Cleanview’s report attributes the biggest delay to Project Jupiter in Doña Ana County, New Mexico, where the state blocked a planned gas pipeline that would have fueled the project’s 2.45 GW on-site plant, so the project “is likely to miss its target of coming online in 2027”, and to Nebius in Vineland, New Jersey, which is struggling to obtain an air permit to run a 400 MW on-site gas plant on which 7 of the 9 compute tranches of its $17.4 billion Microsoft deal depend. Jupiter is a Bloom fuel-cell site, so its bottleneck is neither a turbine nor a GSU. It is a right of way.
Texas shows the other pattern: small permits first, the big one later. Floodlight’s review of TCEQ records found that Stargate Abilene’s fleet of 10 turbines and 62 backup diesel generators was permitted under permits by rule and standard permits that require no public notice, that the developers then filed a first major permit for 41 more turbines and 18 more generators, and that at least 38 Texas datacenters have received minor permits for on-site power since 2024, covering more than 2,100 backup generators. A former EPA air enforcement chief told Floodlight the staggered approach “could violate” EPA aggregation policy. Cleanview’s timeline for the same site: Crusoe filed for a 360 MW plant five months before Stargate was announced, got approval six days later, ordered 10 GE Vernova turbines in December 2024 and 19 more in June 2025. SemiAnalysis reads the TCEQ filing as up to 933 MW of gross nameplate, about 665 MW of it the 19 LM2500s. Georgia went the other way: after VoltaGrid installed 8 of 33 planned engines at a Serverfarm site in Covington before the permit issued, Georgia EPD wrote on July 2, 2026 that the company had likely violated state air rules and must pause construction. The Large-Load Policy Tracker logs each state’s minor-permit threshold because that threshold, not the turbine catalog, sets how many megawatts a developer can energize before anyone can object.
What the Grid Side Is Doing Meanwhile
Three federal actions frame the behind-the-meter decision. The first is DOE’s use of Federal Power Act section 202(c) to keep retiring plants running. DOE’s own logs show the order numbers running from 202-25-3 for the J.H. Campbell coal plant on May 23, 2025 to 202-26-43 on September 3, 2026, with the plant-specific orders renewed every 90 days.
Source: DOE CESER 202(c) order logs 2025 and 2026; nameplate from EIA-860M July 2026; via Open Factory Large-Load Policy Tracker
Eight plants in seven states, 5,349 MW of nameplate by EIA’s unit list: Campbell 1,561 MW, Schahfer 847, Eddystone 782, Centralia 730, Stanton 465, Craig 446, Wagner 415 and Culley 104. The Sierra Club’s methodology, as reported by Utility Dive in March 2026, put the ratepayer cost of the first 13 orders at more than $235 million. For a developer the relevance is indirect but real: 5.3 GW of capacity that the RTOs had already planned around is being counted twice, once in the resource plans that retired it and once in the emergency orders that keep it, and none of it is available to a new 500 MW load in the queue.
The second is FERC. On June 18, 2026 the commission issued show-cause orders to all six RTOs and ISOs, finding their large-load interconnection rules appeared inadequate on five points: study process efficiency, cost-shift prevention, co-location and behind-the-meter generation, new transmission services for flexible loads, and a study process for generators serving “electrically proximate” loads, with 60 days to respond. The third is Texas, where ERCOT’s Batch Zero process, effective July 2026, adds two pathways for large users, the Withdrawal-Limited Private Use Network for customers who build their own on-site generation and the Provisional Controllable Load Resource for those who accept curtailment, against a queue of more than 438 GW. We covered that queue in The 2% Funnel. SemiAnalysis counts about 2,885 MW of announced ERCOT co-location with existing plants, led by AWS’s 1,200 MW at Comanche Peak and CyrusOne’s 400 MW at Thad Hill; those are net-metering deals against existing steel, not new iron. The Large-Load Cost Table carries the WLPUN and PCLR terms because a WLPUN is the legal form most Texas projects in our pipeline will take.
What Actually Energizes by 2028
The forecast is built block by block from data/btm-pipeline.csv and the lead times above, and it is an estimate, marked as such. The rule we applied: a megawatt counts in a year only if the equipment for it was ordered or leased early enough for the published lead time to land before that year ends, the air permit for it exists, and gas is contracted. Cleanview’s end-2026 range of 2.8 to 3.2 GW is the anchor; SemiAnalysis’s June 2026 title asks whether the market reaches “40GW+ of behind-the-meter datacenter by 2028” and forecasts that behind-the-meter will power “well over half of new US datacenters in 2028+”; Cleanview’s own high case is 13 GW by the end of 2027 if every project with a signed tenant hits its schedule, and as low as 5 GW if delays like Jupiter’s and Vineland’s spread.
Source: Open Factory Lead-Time Monitor
Base case: 3.0 GW at end-2026, 7.4 GW at end-2027 and 15.0 GW at end-2028, of which 2.0 GW is the existing mobile and aero fleet, 1.4 GW new aeroderivatives, 7.5 GW reciprocating engines, 3.1 GW frame turbines and 1.0 GW fuel cells, with a low case of 10 GW and a high case of 22 GW for 2028.
The 2027 block is engines and jet engines, because those are the only orders placed in 2025 and 2026 that deliver inside the window. Stargate Frontier in Shackelford County is the largest: Vantage ordered more than 670 reciprocating engines for 2.58 GW, and Texas filings show VoltaGrid approved to operate 210 Jenbacher units totaling 700 MW at that site. VoltaGrid says its Oracle program is 2,300 MW of modular engine packs on Energy Transfer gas, that it has ordered 1.5 GW, 300 Jenbacher J624 and J620 engines in 25 MW units, for delivery by 2028, and that its $5.0 billion financing funds a “4.3+ GW fully contracted power deployment plan through 2028” against about 1.4 GW fielded. We give Frontier 0.7 GW in 2027 and the rest in 2028, the Oracle program 1.5 GW in 2027 and 1.4 GW in 2028 net of the Frontier overlap. The aero block is Abilene’s 19 LM2500s, PROENERGY’s 13 PE6000 packages for Crusoe and Fermi’s 132 MW of TM2500s. We covered why those two technologies beat frames on a 100 MW bridge in Recips vs Turbines vs Fuel Cells; this is that finding at 10 GW scale.
The 2028 block is where the frame turbines and the GSUs start to bind. Kilby’s first power is “anticipated in 2028” after an FID at the end of 2026; seven 7HA.02s under a slot reservation and twelve Titan 350s do not all fire in one year, and NGI’s reading of Chevron’s own capacity chart has the 2.67 GW reached around 2031, so we give it 0.7 GW in 2028. Homer City’s 14 rows in 860M are all status P with the first 1,256 MW dated July 2028, and the campus is still negotiating its anchor tenant, so we give it 0.9 GW. Fermi’s simple-cycle fleet is on site; the gate is the tenant that triggers vertical build, then the substation gear, so 1.0 to 1.5 GW by end-2028 is plausible and we take 1.3 GW. The engine block keeps growing: Joule Capital Partners’ Utah campus expects a 1.5 GW first phase of 636 Caterpillar G3520K gensets to ramp by early 2028, fed by Kern River; Nscale’s Monarch campus in West Virginia has 2 GW of Caterpillar G3500-series engines tied to Microsoft; Wartsila’s Ohio project has 40 34SG engines with delivery aligned to a COD in early 2028. Against that, Evercore’s David Raso told NGI that new large engine orders are landing in late 2027 at the earliest, with most slotting into 2028, and Caterpillar’s own line is 24 to 30 months for orders placed in Q1 2026. Anything not ordered by mid-2026 is a 2029 megawatt. We wrote up the genset side of that allocation in Cummins and Caterpillar Are Sold Out Through 2028.
Fuel cells are the smallest block and the most permit-dependent. Bloom’s 10-K says it plans to double factory capacity from 1 GW to 2 GW by the end of 2026, and its Q2 2026 revenue was $1.065 billion, up 166%, with full-year guidance raised to $3.9 to 4.2 billion, so the factory is not the constraint. Jupiter’s 2.45 GW is: without the pipeline it is 0 in 2027 and we carry 0.5 GW in 2028 on the assumption that a rerouted line or trucked gas covers the first phase. The other fuel-cell megawatts sit in Tallgrass and Crusoe’s Cheyenne campus, which the county approved in January 2026 at 2.7 GW of “natural gas fired turbines, as well as fuel cells,” scalable to 10 GW, and which Crusoe paused in June 2026 at its customer’s request, per DCD and Bloomberg headlines we could not open. We carry 0.5 GW for it in 2028 and would cut that to zero if the pause holds into 2027. The 30% fuel-cell ITC is why Bloom keeps winning the sites that have gas.
The Vendors Are Building for the Engines
The supply side has already voted. Caterpillar, whose large-engine backlog grew more than 3.5x since January 2024, is on track to add 15 GW of annual large-engine capacity on top of the 50 GW it had targeted for 2030, from a 2024 base of about 25 GW; Cummins plans to add 20 GW to reach 55 GW of high-horsepower output by 2030. GE Vernova is going from 20 GW of annual heavy-duty output in Q3 2026 to 24 GW in 2028 and 30 GW in 2030. INNIO booked a 1.1 GW order for more than 200 J624s for a single campus in July 2026 and a 1.25 GW framework with Rehlko in April. Cleanview’s OEM count has Caterpillar at 33% of tracked behind-the-meter projects with more than 8.8 GW of permitted Caterpillar engines and Solar Turbines, and Bloom at 14%.
Source: Open Factory Book-to-Bill League Table (compiled from GE Vernova Q2 2026 8-K, Bloom Energy 10-K, Caterpillar and Cummins statements as reported by NGI, May 2026)
Caterpillar 25 to 65 GW a year, Cummins 35 to 55, GE Vernova 20 to 30, Siemens Energy at a 13 GW run rate with no stated 2030 figure, Bloom at 2 GW by end-2026: the two engine makers add 70 GW of annual capacity by 2030 and the heavy-duty leader adds 10.
That is the forecast in one chart. The 90 GW was announced as a turbine market and the first 15 GW will be built as an engine market, because engines are 78 to 130 weeks from order, ship in 2.5 to 25 MW increments a permit-by-rule can cover, and plug into 13.8 kV switchgear without a 143-week transformer. Kilby, Homer City, GW Ranch and the back half of Matador are 2029 to 2031 projects with a 2028 press release, and what decides each is the GSU delivery date they have not published. The Datacenter Electrical Roadmap tracks both dates per campus.
What to Do Monday
If you are a developer holding a 2028 energization promise on a frame-turbine site, get the GSU purchase order date and the transformer OEM’s slot in writing from your EPC this week; if the order was not placed by mid-2025, re-baseline to 2029 and tell your tenant before their tranche schedule does it for you. If you are choosing technology for a 100 to 500 MW bridge, price the engine plant at MV distribution against the turbine plant with its GSU in the Genset TCO and the Lead-Time-Adjusted Schedule, and read the Datacenter MEP Cost Table line for on-site generation with the transformer broken out. If you are a factory or colocation buyer under 75 MW in Texas, file for the WLPUN pathway rather than firm service and check the Large-Load Cost Table for the security and study fees that come with it. And if you are an investor reading 90 GW, read 15.
Behind the paywall: the twelve campuses ranked by what they can have running by end-2028, 8.8 to 14.2 GW against 30 GW announced, the binding gate on each one, and the vendor-by-vendor winners and payers, all of it maintained in the Lead-Time Monitor behind-the-meter table.