BYOB
How AI’s energy thirst is rebuilding the gas demand curve
There was a time — within living memory, but only just — when the abbreviation BYOB belonged to college parties, suburban housewarmings, and the sort of dinner invitations where the wine was bad and the conversation only slightly better. The host, having either declined or failed to stock the bar, asked the guests to provision themselves. The implicit confession was that the establishment could no longer be trusted to keep its visitors watered.
It now belongs to data centres.
When Elon Musk wanted a hundred thousand graphics cards humming in Memphis last year, he did not wait for the local utility to find him a glass. He trucked in turbines. The xAI campus at Colossus is, today, powered by a fleet of gas-fired engines that arrived on flatbeds — equipment that, six years ago, would have been mistaken for a stopgap. It is no longer. Memphis Light, Gas & Water has since been informed of what the next phase requires; it has nodded politely; it has joined the queue. Where the host cannot supply, the guests must.
The parallel is not, as it happens, flippant. Alcohol and electricity occupy a curiously similar place in their respective gatherings. Both are, by volume, the cheapest part of the evening; both are also, when supply is interrupted, the part the participants notice first. A wedding without champagne is a tense affair. A data centre without electricity is no data centre at all. The difference is that the wedding planners of the AI era — Meta, Microsoft, Oracle, xAI — have decided the host’s cellar is no longer to be relied on.
The pattern, once one begins to look, is everywhere. Meta’s seven-gigawatt Hyperion campus in Louisiana is now contracted with Entergy for seven new gas plants — built, in essence, to serve a single guest. Microsoft has announced 1.4 gigawatts of behind-the-meter gas in West Virginia and is in late-stage talks with Chevron and Engine No. 1 for another 2.5 gigawatts in West Texas. Oracle’s Project Jupiter in New Mexico is being powered by gas turbines supplied by Siemens and GE. Amazon, the last holdout, settled for two hundred megawatts of solar around the edges of a $12 billion Louisiana data centre — the polite gesture of a guest who has at least brought a bottle of cava. Even Google, whose decade-long campaign for clean-energy purity made it the moral lodestar of its peers, is now exploring on-site gas with Crusoe Energy at its 933-megawatt Goodnight campus in Texas. The bridge fuel, it turns out, has reached the other side of the river without finding anywhere to disembark.
Goldman Sachs reckons up to a third of incremental American data-centre capacity over the next five years — perhaps twenty-five gigawatts — will be built off-grid. McKinsey puts the figure at 25–33%. Either way, the implication is the same. A new and structurally durable consumer of natural gas is being hard-wired into the energy system, on a scale large enough to redraw demand curves that had only just begun to look settled.
There are, broadly, two ways to read this. One is the polite version: gas is a temporary expedient while small modular reactors and long-duration storage catch up. The other is the version the data-centre developers, by their actions if not their press releases, are quietly endorsing — that gas is the answer because nothing else can be in the field by the end of the decade. The grid interconnection queues now stretch to five or seven years across most American ISOs. ERCOT, the Texas operator, has fielded 226 gigawatts of demand requests for 2030, almost a hundred times the volume it approved in 2022. Hyperscalers building campuses with eighteen- to thirty-six-month construction windows cannot wait for the grid to find its slack. They have therefore decided to build the grid themselves.
Once the conceit takes hold, it does not stop with the bar. The bouncer at the door — the interconnection queue — has become the gatekeeper that will keep the guests waiting until 2031 if they are patient enough to apply through normal channels. The gas-turbine manufacturers, meanwhile, are the new sommeliers: GE Vernova, Siemens Energy, Mitsubishi Heavy Industries, with order books that now stretch into the early 2030s and pricing 195% higher than their 2019 baseline. Reciprocating-engine makers — Wärtsilä, Caterpillar, Cummins, INNIO’s Jenbacher — are filling the gaps that the turbine giants cannot, the way decent house wines fill the gap when the cellar runs out of vintage. Their manufacturing capacity, on Thunder Said Energy’s reckoning, can double from 15 GW per annum to over 60 GW per annum by 2035. The orders are justifying the build-out.
Behind the engines stand the pipes that feed them. East Daley Analytics now tracks more than four hundred data-centre projects with an aggregate load of 183 gigawatts. Even on its base-case assumptions, that translates to four-to-six bcfd of incremental American gas demand by 2030; bull-case scenarios run as high as twenty bcfd. Energy Transfer alone is fielding inquiries for pipeline projects to serve forty-five power plants across eleven states, plus over forty prospective data-centre offtakers. MPLX is doing similar. Williams’ Socrates project in Ohio — a two-hundred-megawatt facility hardwired to a Meta campus in New Albany, never connected to the grid — is the new template. Behind the meter, meet behind the pipeline.
And behind the pipelines, finally, the molecules themselves. Enverus and EQT both think the call on US gas could rise by 10–18 bcfd by 2030 once data centres, LNG exports and coal-to-gas switching are stacked together. That number does not appear in the consensus model of any major equity strategist. It does, however, appear in the consensus model of the people actually building the data centres.
Three observations, before the bar closes. The first is that a sector whose demand was, until very recently, modelled as flat-to-declining is in fact looking at asymmetric demand growth that almost no climate model has yet incorporated. Global gas-fired power generation rises from 144 bcfd in 2023 to 315 bcfd by 2050, more than doubling. Data-centre demand sits on top of that as a kicker. The asymmetry is heightened by the fact that the relevant infrastructure cannot quickly be rebuilt. Permitting cycles in North America now run five to seven years for greenfield interstate pipelines; in Europe, closer to eight. Whatever investments are sanctioned now will be needed; whatever is not, will probably remain not.
The second is that value sits with operators whose contracts are explicitly structured to absorb the new demand. A take-or-pay agreement with Microsoft is, in revenue terms, very nearly indistinguishable from a Treasury bill with a 25-year tail and a hyperscaler counterparty. The only meaningful risk is regulatory; the only meaningful upside is escalator-driven. The market continues to treat such cash flows as oil and gas equity. That seems wrong.
The third concerns the politics. President Trump’s promise on the 13th of January — that ordinary Americans would not “pick up the tab” for AI’s electricity binge, but Microsoft would — has the unintended effect of locking the hyperscalers further into self-supply. If you must build your own bar, you stock it with the cheapest, most reliable, most permissible spirit available. In 2026, in America, that is gas. The same logic is now operating, with a lag, in Europe and Asia.
A small irony to close. The data-centre boom was supposed to be the great accelerator of decarbonisation: AI for grid optimisation, AI for materials science, AI for fusion. It may yet be all those things. In the meantime, it has built natural gas a multi-decade demand runway whose existence almost no climate model has yet incorporated. The reluctant kingmaker of the green grid, it turns out, is also the reluctant kingmaker of artificial intelligence. The molecule that everyone keeps planning to retire is the one that everyone keeps quietly ordering — a round, as it were, on the house. Or rather, on the houses, plural, that the guests are now obliged to build for themselves.
EDITORIAL NOTE
These three essays form part of a series on the structural case for natural gas as an asset class. They are intended to underpin the investment thesis behind the Navigate Global Energy High Income Fund, a sub-fund of Navigate Funds SICAV plc (subject to final MFSA approval). The Quality of Income Score, Strategic Asset Value, Asset–Liability Duration Index and Demand-Indexed Income Score are proprietary analytical constructs developed by Navigate’s investment team and form part of the Fund’s QuadLogic-overlaid security selection process. The framework is illustrative; numerical estimates are directional and based on Navigate’s internal modelling unless otherwise stated.
Marketing communication. For professional and institutional investors only. Not a recommendation to buy or sell any security. Capital at risk.
