Off the Grid on Purpose

The hyperscalers did not solve the power problem. They quit the grid. When Microsoft signs a twenty-year deal for a 2.7-gigawatt gas plant that is engineered, on purpose, to never connect to a public wire, that's not procurement. It's secession. And the reason is a single number the AI-buildout story keeps stepping around.

Here's the deal, stripped to the bone. Chevron and Microsoft are building Project Kilby near Pecos, Texas — roughly 2.7 gigawatts of natural-gas generation, enough to run about two million homes, dedicated entirely to one data center. The load-bearing detail isn't the size. It's the wiring. The plant will not be connected to the electric grid. It's "behind the meter," which is the polite industry term for off the grid on purpose. Power's expected to flow in 2028. Chevron just spent the back half of the week talking up the arrangement as a growth vector, and it is one — for them. For everyone reading it as an energy story, it's the wrong frame.

The constraint was never generation

Strip this to first principles and ask what's actually scarce. It is not electrons. The US knows how to make power. What it can't do — not on anything close to AI's clock — is connect new power to the grid.

This is the number. The interconnection queue, the line every new generator has to stand in to plug into the public grid, has swelled to something like 2,600 gigawatts of backlog, with median wait times pushing five years. That's the median. Data-center-scale loads are looking at worse — up to twelve years by some counts. One widely cited figure puts a 100-megawatt connection at Dominion at roughly seven years. California stretches past nine. Sit with that against the other number that matters: a hyperscaler wants the facility live in twelve to twenty-four months, because the model it's racing to serve will be two generations old by the time a seven-year interconnect clears.

So the timelines don't just miss. They miss by an order of magnitude. A five-to-twelve-year wait to power a thing you needed running last quarter isn't a delay you optimize around. It's a wall. And when the front door takes a decade, you stop treating it as the door.

Behind-the-meter is the side entrance. Build your own generation on-site, wire it straight into the load, skip the queue entirely, and deployment collapses from a five-year median to eighteen-to-thirty-six months. That's the whole trade. You give up the grid's redundancy and its shared cost base, and in exchange you get the one thing the grid can't sell you at any price: time. For a business where the compute depreciates in thirty-six months, time is the only currency that clears.

This is vertical integration, and it's already a pattern

One deal is an anecdote. This isn't one deal. A February analysis counted 46 data-center projects planning to build their own generation behind the meter — 56 gigawatts of private power, and ninety percent of it announced in a single year. That's not a workaround anymore. That's a category. The most software-native companies on earth are quietly turning into power companies, because the alternative is waiting in a line that outlasts the hardware they're trying to feed.

Watch what that actually is: vertical integration forced by a constraint the incumbent can't fix. Compute got bottlenecked on power, power got bottlenecked on interconnection, and rather than wait for grid operators and regulators to reform a queue that's been broken for a decade, the buyers with the deepest pockets in the market are routing around the public system entirely. They're not lobbying to fix the grid. They looked at the timeline and left.

There's a fair counterargument, and the twenty-year term is where it lives. Maybe this is a bridge — private gas to get online now, reconnect to a modernized grid later once the queue clears and the renewables catch up. Maybe. But you don't sign two decades and pour concrete for a 2.7-gigawatt station as a stopgap. The length of the commitment is the tell. Behind-the-meter got underwritten as permanent infrastructure, not scaffolding, because nobody building it believes the queue gets fixed on a timeline that matters.

A second grid is forming, and it's private

Step to the system level and here's what actually changed this month. A parallel power system is being built alongside the public one — dedicated, gas-fired, off-grid, and owned by the largest compute buyers in the world. The public grid keeps its residential and industrial load and its shared costs. The AI economy quietly decouples from it and builds its own.

That has second-order consequences nobody's pricing yet. Every gigawatt that goes behind the meter is a gigawatt of the biggest new customers on earth not helping carry the grid's fixed costs — the transmission, the reserve margin, the upgrades everyone else's rates were supposed to fund together. When the anchor tenants build their own utility and walk, the people left on the public system inherit a bill that was supposed to be shared. The grid didn't lose a fight. It just stopped being invited to the biggest buildout of the decade.

The power problem was never really about power. It was about a queue nobody would fix, and a set of buyers rich enough and impatient enough to stop asking. They didn't wait for permission. They built their own door.

When your biggest customers start generating their own electricity rather than stand in your line, the line was the product you failed to ship.

— Dustin