For two years, the “bring your own power” debate has run on adjectives: faster, costlier, cleaner, dirtier, inevitable, unsustainable. What it has lacked is a price. On October 15, Venture Global supplied one, or something close to it. The liquefied natural gas exporter announced frame agreements signed October 9 with GE Vernova and Baker Hughes for up to 16 gas-fired power islands of 720 megawatts each, plus a grid-equipment agreement with GE Grid, and disclosed in a Form 8-K filed the same day that if every contemplated purchase order were executed, aggregate payments would total approximately $17.2 billion, including reservation fees and subject to adjustment. Divide that headline figure by the 11.52 gigawatts of island capacity and you get roughly $1,490 per kilowatt. That is not a quoted price for a power island — the company disclosed no unit prices, and the $17.2 billion also covers grid equipment for up to 21 projects, not just the 16 islands — but it is the first public arithmetic of its kind at this scale, and it gives the rest of the industry something concrete to argue with.
The announcement matters because of who is making it. Venture Global is not a data center developer dabbling in generation. It already operates what chief executive Mike Sabel has called one of the world’s largest private industrial electric grids: about 4.5 gigawatts of modular, gas-fired generation across six power islands ordered, under construction, or operating at its Louisiana LNG terminals, running in island mode — entirely behind the meter, independent of the public grid. The new agreements extend a machine the company already knows how to build, in the same combined-cycle configuration, with deliveries contemplated from 2028 through 2034. The company says the expansion could support up to 10 gigawatts of new compute capacity over time at sites it owns or controls, by the end of 2034, subject to customer commitments and market demand. No customers were named, no sites were named, and no data center has yet been contracted. What was signed is a framework for future orders, not an order book.
What was actually signed
Precision about the paperwork matters here, because the headlines have been looser than the filing. The Power Frame Agreement with GE Vernova and Baker Hughes Energy Services covers turbines, steam turbines, associated equipment and services, plus Baker Hughes controls and electrical integration, for up to 16 new power island units. A separate Grid Frame Agreement with GE Grid covers transformers and gas-insulated switchgear for up to 21 power island projects. Alongside those, Venture Global executed a manufacturing reservation agreement with GE Vernova for eight additional power islands and five long-lead equipment reservation agreements with GE Grid. The company expects to place its first power island purchase order by the end of 2026, with orders for eight of the remaining 15 islands anticipated beginning in early 2027. Reservation rights on the rest lapse in stages between October 2027 and October 2029, according to the filing’s summary in trade coverage.
That structure — frame agreements plus staged reservations, not a single bulk purchase — is itself the story. Venture Global has bought optionality in a market where the scarcest commodity is no longer gas or land but a manufacturing slot. Wood Mackenzie’s April 2026 report on the U.S. gas turbine market found global orders of 110 gigawatts at the end of 2025 against global manufacturing capacity of only 60 to 70 gigawatts, with order books sold through 2027 and lead times of six years. The consultancy expects U.S. gas turbine prices to reach $600 per kilowatt by the end of 2027, up 195 percent since 2019, and notes that turbines now make up an estimated 20 to 30 percent of combined-cycle project costs. In that market, a frame agreement is not a procurement detail. It is the product.
It is also worth noting what the framework does not commit. The equipment can serve LNG export projects, bolt-on expansions, or LNG-adjacent data centers; the filing keeps the destination flexible. If data center customers do not materialize on acceptable terms, the islands can go to liquefaction. If they do, Venture Global becomes, in Sabel’s phrase, a “vertically integrated AI utility” — offering island-mode power, gas supply, closed-loop water systems, and land as a bundle. That dual-use hedge is the difference between this announcement and a speculative data center power play: the fallback customer is the company’s own core business.
The $1,490 question
So what does the price tag actually tell us? Start with the arithmetic and its limits. Seventeen point two billion dollars divided by 11,520 megawatts is about $1,493 per kilowatt. Call it roughly $1,490. But the numerator includes reservation fees and grid equipment for as many as 21 projects, while the denominator counts only 16 islands, so the figure overstates the equipment cost per island kilowatt and understates nothing else — it is a ceiling on the equipment math, not a quote, and it is certainly not an installed cost. Land, site work, gas interconnection, permitting, construction labor, financing, and the data center itself are all outside it. Anyone citing $1,490 per kilowatt as the cost of a data center power island is misreading the filing.
The right comparison is against installed costs for new grid-connected combined-cycle plants, and there the picture is sobering for everyone. The Institute for Energy Economics and Financial Analysis, in an April 2026 report, put recent U.S. gas plant costs at roughly $2,500 per kilowatt installed, about triple the cost of similar projects earlier in the decade, and cited a Dominion Energy and Santee Cooper proposal for a 2,180-megawatt combined-cycle plant in South Carolina at $5 billion, or $2,293 per kilowatt — a figure the utilities said had doubled in a couple of years, before counting three new or expanded pipelines to feed it.
Set Venture Global’s roughly $1,490 per kilowatt of equipment and grid hardware against $2,293 to $2,500 per kilowatt installed, and the gap is the island model’s argument: a repeatable power island from a builder with six already built may land below a bespoke utility plant procured in the same inflated market.
May. The comparison is directional, not decisive. The IEEFA figures describe utility plants with different configurations, interconnection costs, and regulatory burdens, and Venture Global’s installed cost per kilowatt will not be known until it actually builds one of these islands for a data center and says what it spent. The honest reading is narrower: the equipment bill for bring-your-own-power at gigawatt scale is now denominated, it is large but not obviously prohibitive relative to the grid-connected alternative, and the binding constraint has shifted from “can we afford the turbines” to “can we get the turbines, and on what schedule.”
Speed is still the product
That shift is why the timeline in the announcement deserves as much attention as the money. Deliveries are contemplated from 2028 through 2034, the first purchase order is expected by the end of 2026, and trade coverage of the announcement put first power for data centers around late 2028 or early 2029. That is not the 18-to-30-month turbine deployment this desk described in the bring-your-own-power explainer — that window assumed available aeroderivative units, not heavy-frame combined-cycle islands in a six-year-lead-time market. It is, however, faster than the three-to-seven-year grid interconnection queues that drove operators to self-generation in the first place, and it comes with a certainty the queue cannot offer: a reserved manufacturing slot and a repeatable design.
The schedule also explains the reservation structure. Venture Global is not buying 16 islands; it is buying the right to buy them, in tranches, as customers commit, while the reservation clock runs to 2029. That is a rational response to a market in which Wood Mackenzie expects turbine orders to peak in 2026 as developers lock in equipment for 63 gigawatts of gas capacity additions through 2030. Reserve early, decide late, and let the customer commitments — which the company stresses the 10-gigawatt compute figure depends on — determine how much of the framework converts into steel.
There is a second, quieter timeline: permitting. Venture Global planned to file its first permit applications in the week of the announcement, and an island that never touches the public grid still needs air permits. The island model removes the interconnection queue; it does not remove the permitting queue.
The 10-gigawatt gap
One discrepancy in the announcement deserves a plain note, in the spirit of this desk’s earlier flagging of Google’s “1 gigawatt” virtual power plant that was actually 1.17 gigawatts. Sixteen islands at 720 megawatts equal 11.52 gigawatts of generation, but the compute capacity on offer is “up to 10 gigawatts.” The company has not publicly reconciled the two figures. Some of the gap may reflect reserve margin, parasitic loads, or islands earmarked for LNG rather than compute — but that is speculation, and the filing does not say. What can be said is that 10 gigawatts of compute is a development ambition stretched over eight years and conditioned on customers who have not yet signed, not a contracted pipeline. Treat it as a ceiling, not a forecast.
The same discipline applies to the company’s claim that the expansion would place no additional demand on the public grid or public water utilities and therefore impose no added costs on local consumers. That follows logically from the island-mode design — generation behind the meter, closed-loop cooling, private distribution — but it is the company’s claim about its own future projects, not an independently verified finding, and it sidesteps the questions island-mode projects do raise: air emissions from 11.52 gigawatts of gas combustion, local gas pipeline capacity, and the long-term carbon lock-in of infrastructure built to run into the 2050s. A data center that never touches the grid still touches the atmosphere.
What this means for the grid — and for the operators who stay on it
Step back, and the announcement clarifies the fork in data center power strategy that this desk has been tracking all month. Last week, Google’s expanded agreement with Voltus offered one answer to the interconnection crunch: buy flexibility across the grid, 1,070 additional megawatts of contracted demand-side capacity through 2040, and let the demand side function as supply. Venture Global offers the opposite answer: leave the grid out of it entirely, generate behind the fence at industrial scale, and sell power as a bundled utility service to customers who would rather not wait. One strategy makes the grid more flexible; the other makes it less necessary, at least for the largest customers.
For grid planners, the Venture Global model is the harder one to plan around. A gigawatt of flexible demand in PJM can be accredited, dispatched, and counted, however imperfectly. Ten gigawatts of island-mode compute simply never appears in a load forecast, never pays a transmission charge, and never shows up in a capacity auction. If it scales, utilities forecasting load must guess how much of the data center pipeline will self-supply, and guessing wrong in either direction means overbuilding or underbuilding the shared system.
For data center operators, the price tag reframes the choice. Bring-your-own-power was long discussed as a premium product — faster, but expensive, and therefore reserved for operators who could not wait. Venture Global’s arithmetic suggests the premium may be smaller than assumed, at least for the equipment, and that the real cost is the commitment: reserving manufacturing slots years ahead, accepting gas-price and emissions exposure for decades, and becoming, in effect, a power developer. That is a different business, and not every hyperscaler wants it. Some will pay the island-model price; others will keep buying flexibility, signing grid PPAs, or waiting in the queue. The market is segmenting, and the segmentation is now priced.
What to watch
Four things will determine whether this framework becomes a fleet or a footnote. The first is the year-end purchase order: Venture Global expects to place one by the end of 2026, and a missed or slipped first order would say more than any announcement. The second is customer disclosure: the first named data center customer, and the commercial terms — which the company has not disclosed and this desk will not infer — will reveal whether the “vertically integrated AI utility” has buyers or only a brochure. The third is permitting: the pace of those first applications, filed the week of the announcement, is the earliest schedule signal. The fourth is cost transparency: the first island built for compute, with a disclosed installed cost, will let the industry replace the $1,490 equipment arithmetic with a real number.
Until then, the honest summary is this: onsite gas power now has a price tag. Roughly $1,490 per kilowatt of equipment and grid hardware, against $2,293 to $2,500 per kilowatt installed for a new grid-connected combined-cycle plant, with six-year turbine lead times and a reservation clock running to 2029. Whether that arithmetic scales bring-your-own-power from a workaround into a business model depends on customers, permits, and the first real build — the same three things it has always depended on. The difference is that now the argument can be had in numbers.