The Story
Open your laptop and ask a frontier model a hard question. The answer takes a few seconds. The electricity to produce it does not come from a cloud. It comes from a turbine, somewhere on the ground, and increasingly that turbine sits behind a fence on a private campus that the public grid never touches.
A year ago this would have read as a niche infrastructure observation. Today it is the load-bearing fact of the AI buildout. The companies that wrote the loudest clean-energy pledges of the last decade are now the same companies financing gigawatt-scale natural gas plants in their own backyards. They are doing it not because the strategy changed. They are doing it because the calendar changed.
The chain runs like this. The Strait of Hormuz crisis put a hard premium on dispatchable, domestically-fueled power and rewrote what corporate boards mean when they say “energy security.” Federal permitting reform that was supposed to unblock the public grid stalled out in the Senate, taking the high-voltage transmission queue with it. The wait to interconnect a new gigawatt of load to the public system stretched into the back half of the decade. So the hyperscalers did the math and stopped waiting. They started procuring their own power on their own land.
That move solved the grid bottleneck and exposed the next one. The machines still have to come from somewhere. Heavy-duty gas turbines are not a commodity. They are some of the most precise pieces of industrial hardware ever built, and the three OEMs that make most of them are now booked toward the end of the decade. Walk in today with a checkbook and you join a queue that delivers in 2028 at the earliest. So the buildout adapts again. Where the big combined-cycle slot is unavailable, the project gets redesigned around what is available: hundreds of smaller reciprocating engines, modular fuel cells, midstream-built behind-the-meter plants. The constraint dictates the design.
This is the through-line of Episode 01. AI did not invent any of these constraints. It just made them load-bearing. The public energy story is now downstream of an industrial supply chain that most people in the AI world have never had to look at, and a permitting system that most people in the energy world have been complaining about for ten years. The two now sit on the same critical path.
What Made It Visible
The Turbine Is Not An Appliance
A modern heavy-duty gas turbine is closer to a permanent jet engine than to a piece of plant equipment. It runs continuously for thirty or forty years inside a fireball whose temperature exceeds the melting point of the metal that contains it. The trick is the blade. It is cast from a single-crystal nickel superalloy, then drilled with a maze of microscopic internal channels. Compressed air is bled through those channels and forced out across the surface. The blade survives by sweating a thin film of cooler air over itself, in flight, while spinning at thousands of RPM. Take a beat with that picture, because it explains everything that follows about lead times.
A combined-cycle plant pairs that turbine with a second one. The first turbine, on the Brayton cycle, burns gas and spins a generator. Its exhaust is hot enough to be the entire fuel source for a second machine, so it is routed into a heat-recovery boiler, makes steam, and drives a Rankine-cycle steam turbine. The same fuel does two jobs. Plant efficiency climbs from roughly 35% to 60% or more. This is the workhorse of new firm power in 2026. It is also why you cannot scale the supply chain by pouring more concrete. The factories that grow single-crystal blades, forge the rotors, and certify the welders take years to stand up.
GE Vernova ended 2025 sitting on an 80 GW backlog stretching to 2029 and expects to be sold out through 2030 by the end of this year. Siemens Energy is carrying a EUR136 billion backlog, the largest in its history. Mitsubishi Power is doubling capacity over two years, but the slot still lands in 2028 to 2030. Together those three serve more than 75% of projects under construction. Industry-wide lead times have stretched to roughly eight years from contract to commissioning. Slot reservations have started to behave like financial assets, because they are: a confirmed place in the queue is now worth real money to whoever did not get one.
If you are a hyperscaler watching that timeline, eight years is not a wait. Eight years is being out of the AI race entirely.
The Workaround Becomes The Plan
So they redesign around it. Behind-the-meter generation, where the data center owns and consumes its own power on its own footprint without ever touching the public grid, is now baked into roughly 48 GW of proposed US data center capacity. That is about a third of every project on the planning books. It used to be a workaround for one-off campuses. It is now the default architecture for the buildout.
The shape of those plants tells you everything. Williams Companies, a midstream operator that has spent a century moving natural gas through pipelines, has stood up a $5.1 billion behind-the-meter program with more than 6 GW in the pipeline across Texas, Tennessee, and Ohio. Their first project, a 750 MW build in New Albany for a Meta-affiliated facility, is expected online late 2026. The midstream company is becoming a private utility because the customer asked it to.
Meta’s El Paso campus shows what happens when even Williams-style speed is not fast enough. Meta needs roughly a gigawatt of continuous power and cannot wait for a single big combined-cycle slot. The solution: 813 individual reciprocating natural gas engines, totaling 366 MW, built around the data center for $473 million. That is not an elegant power plant. It is what you build when the supply chain forces you to assemble a gigawatt out of the parts that physically exist this year.
The other end of the same problem is the fuel cell. Bloom Energy’s solid oxide units skip combustion entirely. Natural gas enters the anode, oxygen enters the cathode, and at high temperature the oxygen ions migrate through a ceramic electrolyte and react with the fuel, stripping electrons that travel through an external circuit as direct current. No flame, no rotating mass, fewer local emissions, and 99.999% rated reliability. Bloom is guiding $3.1 to $3.3 billion of 2026 revenue on what they describe as a “very large product backlog” from AI data centers, and their installations now scale from 20 MW to 500+ MW. They are expensive per megawatt. They also exist today.
And then there is Goodnight. Google and Crusoe Energy are building a roughly $30 billion AI campus near the town of Claude in Armstrong County, Texas, anchored by a proposed 933 MW behind-the-meter natural gas plant with no carbon capture and no grid interconnection, alongside 265 MW of wind and up to 1 GW of battery storage. Approximately 4.5 million tons of CO2 per year, on a project from the company that built its brand on running entirely on clean energy. Stated flat, that is what the constraint set produced. Speed beat the pledge.
The Money Has Already Repriced
When a structural shift of this scale is real, capital arrives ahead of consensus. KKR and Energy Capital Partners committed $50 billion in mid-2025 to a strategic partnership built specifically around co-located power generation and data centers. Traditional project-finance lenders, the same banks that spent the last fifteen years underwriting renewables, are syndicating loans backed by the data center lease and the hyperscaler’s corporate credit rather than by a separate offtaker.
That last detail is the structural rewrite. The PPA, the long-dated power purchase agreement, has been the bedrock instrument of clean energy finance since the 2010s. It works because there is a generator and a separate buyer with strong credit. Behind-the-meter collapses those two parties into one. The hyperscaler is the developer, the offtaker, and the consumer. There is no merchant risk to hedge because there is no merchant. The financing model migrates from independent power project to integrated infrastructure on the data center’s own balance sheet.
Meanwhile the market is repricing the alternatives. Constellation Energy, the cleanest proxy for the nuclear-for-AI thesis, has retreated about 30% from its late-2025 peak. The carbon math still favors nuclear. The clock does not. A small modular reactor sited today does not produce an electron until 2035 at the earliest, and a hyperscaler that waits until 2035 has lost the race that started in 2024. The same market that was euphoric about nuclear-for-AI a year ago is now pricing in the primacy of speed.
The clearest signal of where the capital actually went is on the wrong side of the trade everyone expected. The best-performing equities in the AI complex through Q1 2026 are not the model labs and not the chip designers. They are the oilfield services companies. SLB and Baker Hughes are beating Big Tech by roughly 30% year-to-date. The picks and shovels of the AI gold rush turn out to be drill bits, hydraulic pumps, and seismic sensors used to extract the natural gas that ultimately ends up in those 813 reciprocating engines outside El Paso. That is not poetic. It is what the supply chain says the bottleneck actually is.
Conclusions
For energy people, this is the moment AI stops being a software story you read about and becomes a demand signal sitting on top of your order book. The customer is willing to write a ten-figure check, sign a fifteen-year contract, and tolerate gas combustion because the alternative, for them, is being late. That changes the politics of new generation, the price of slot reservations, and the deal structure of project finance.
For AI people, the scaling limit is no longer only inside the model layer. The binding constraint is now the federal infrastructure permit, the turbine slot, and the local air-quality decision in a county council you have never heard of. The cloud is downstream of an industrial supply chain that does not move at software speed and cannot be arbitraged with capital alone.
For the people standing between the two worlds, which is where this journal is written, the useful work is translation. The AI sector is now buying a kind of operational competence that the energy sector has been quietly accumulating for decades, mostly without the language to sell it. Closing that gap is the actual job.
What We Are Watching
- Whether behind-the-meter gas becomes the default hyperscaler strategy or runs into local air-permit resistance.
- Whether turbine reservation slots become strategic assets in their own right.
- Whether transformer shortages prove more constraining than gas turbines.
- Whether hyperscaler clean-energy commitments bend further around dispatchable gas.
- Whether hydrogen-ready hardware remains a transition story or becomes a procurement cover story for near-term methane combustion.
Field Notes
The source-layer research that backs this episode.