Meta just signed the energy equivalent of annexing a small state — and Wall Street financed the whole occupation.
The Summary
- Meta is building a 7.5-gigawatt AI data center in rural Louisiana, powered by 10 new natural gas plants in one of the largest private capital deals ever assembled
- The facility will consume 5 gigawatts just for compute — more power than most American cities use for everything
- This isn't infrastructure. It's energy colonization disguised as economic development.
The Signal
Meta isn't building a data center. They're building a power plant that happens to run inference.
The Richland Parish project requires 7.5 gigawatts of total power capacity, with 5 gigawatts dedicated purely to computation. For context, the entire city of San Francisco uses about 1 gigawatt. Meta's AI facility will consume five times that just to run models. The remaining 2.5 gigawatts will handle cooling, networking, and keeping the lights on.
The capital structure is the real story. One of the largest private financing deals ever means Meta didn't just write a check. They assembled a consortium of infrastructure funds, energy investors, and probably sovereign wealth looking for 20-year returns on essential digital infrastructure. When you need that much money, you're not building a building. You're creating an asset class.
"When a tech company needs 10 new power plants, they've stopped being a tech company."
Louisiana offered what Silicon Valley couldn't: space, cheap natural gas access, and a state government willing to fast-track permitting for a region that's been economically hollowed out for decades. Richland Parish gets construction jobs and the promise of permanent technical employment. Meta gets uninterrupted power and a location far enough from major metros that nobody complains about the noise or the gas turbines.
But here's the math that should worry everyone:
- Training GPT-4 reportedly used 50 gigawatt-hours
- This facility can deliver 5 gigawatts continuously — 43,800 gigawatt-hours annually
- That's enough to train GPT-4 876 times per year, or run constant inference at incomprehensible scale
The agents economy doesn't run on code. It runs on electricity. Every autonomous AI doing work on your behalf, every real-time model generating responses, every agent negotiating with another agent — it all draws power. We've spent two years talking about the intelligence explosion. We should have been talking about the energy bill.
Meta's competitors are watching. Google, Microsoft, Amazon, and Anthropic are all scouting similar deals. The pattern is forming: find economically desperate regions with energy infrastructure, offer jobs and tax revenue, extract decades of subsidized power. It's the same playbook as manufacturing in the 20th century, except the factory runs 24/7 and the output is weightless.
The Implication
The AI buildout is entering its industrial phase. If you're investing, stop betting on pure software plays and start looking at power generation, grid infrastructure, and cooling technology. The returns are in the picks and shovels, not the prospectors.
If you're trying to understand where AI agents will actually run, follow the gigawatts. The companies that secure multi-gigawatt power deals in the next 24 months will control inference capacity for the next decade. Capacity is the new moat.
And if you're in state economic development, you now know the price of jobs: your energy grid, your environment, and a 30-year commitment to keeping turbines spinning for a company that could pivot to quantum computing or abandon the site the moment economics shift. Louisiana made a bet. We'll know if it paid off sometime around 2055.