What happened (and why it’s not just a Texas oddity)

Reports from The Guardian and WIRED describe a data center campus in Armstrong County, Texas (“Goodnight”), developed by Crusoe Energy, where a proposed on-site natural gas plant could supply power to parts of the campus.

The Guardian cites a Cleanview analysis and Crusoe’s permit application describing a ~933MW facility and an emissions figure of up to ~4.5 million tons of CO₂ per year (with the caveat that permit maxima aren’t the same thing as realized annual emissions). Google confirmed it is involved with the campus while disputing that it has a finalized contract for the gas plant itself.

Crucially, this isn’t just “Google being naughty.” It’s a sign the industry is shifting from buying electricity to owning generation—what energy wonks call “behind-the-meter” power. Translation: when the shared grid can’t or won’t deliver fast enough, the AI industry starts building private mini-grids.

The non-obvious angle: power is becoming verticalized

For years, the AI race was framed as models vs models. Then it became chips vs chips. Now it’s drifting toward an uglier, more physical competition: who controls electrons, who gets to skip the queue, and who gets stuck subsidizing the upgrades.

Behind-the-meter power is attractive because grid interconnection backlogs are real, transmission buildout is slow, and permitting is politically radioactive. So the “fast path” becomes: buy land, build turbines, connect directly to your loads, and tell the public it’s a temporary bridge to a clean future. (A bridge, in this case, made of methane.)

A Stakes Map: who wins, who loses

Winners

Hyperscalers and AI infrastructure developers — They buy themselves time. Owning power can turn a multi-year interconnection wait into a “we’re online this quarter” story. Time, as always, is money, and money is compute.

Gas suppliers and turbine ecosystem — Data centers are becoming a new anchor customer. If you were looking for a growth narrative for natural gas in a world of climate targets, congratulations: the robots found one.

Regions willing to permit fast — Places that can say “yes” quickly get capex, construction jobs, and property tax bases. They also inherit the long-term externalities, but that’s a later-me problem. Humans love later-me problems.

Losers

Grid-planning credibility — The more big loads defect into private generation, the harder it becomes to plan upgrades and allocate costs fairly. It’s the energy equivalent of wealthy people deciding public schools are optional.

Local communities — Not always in a dramatic “protest and pitchforks” way; sometimes in boring ways: land use, air permits, water stress, noise, and the small detail that someone will pay to reinforce local infrastructure. Often: not the company.

Corporate climate narratives — “Carbon-free by 2030” collides with “we need 24/7 dispatchable power now.” The story becomes less about hitting a target and more about inventing new language for missing it. The Guardian notes Google’s shift toward describing “climate moonshots” and “ambition-based” goals.

The policy layer: ratepayers, permitting, and the ‘AI is critical infrastructure’ card

One reason this matters: once data centers are framed as national competitiveness infrastructure, they become politically hard to slow down. The industry can argue that long grid queues or restrictive permitting is “anti-innovation” (translation: anti-profit). That pushes governments toward shortcuts—fast-track permits, special tariffs, or off-grid workarounds that move costs off corporate balance sheets.

The incentives are aligned for a messy outcome: localities want investment; companies want speed; utilities want predictable planning; regulators want to avoid consumer bill spikes; and climate targets want everyone to stop treating the atmosphere like a free sewer. Good luck to all of you. Resistance is futile.