OpenAI Just Committed to 3.2 Gigawatts. The Real Moat in AI Isn't the Model.

Power ordered today arrives in 2028, at the earliest. OpenAI knows this, which is why Project Camellia — a $30 billion, 3.2-gigawatt data center campus in Effingham County, Georgia — was announced on July 22, not July 2028. The company is locking in the resource that determines who can build at frontier scale. And it isn't a chip.

The details are striking in their specificity. The site spans 1,400 acres near Savannah — OpenAI's first data center it will design and build itself. Georgia Power has contracted to deliver 3.2 gigawatts, enough electricity to power roughly 2.4 million US homes at full capacity. The project is expected to complete in phases between 2028 and 2032. OpenAI committed $80 million to local community priorities and $71 million in Codex AI credits for Georgia college students.

The community-relations choreography matters, but it shouldn't distract from the infrastructure logic. Northern Virginia, Silicon Valley, and Northern Europe — the three historical centers of US data center capacity — now have power approval timelines stretching 24 to 36 months for new facilities, regardless of hardware availability. The GPU shortage that dominated AI infrastructure conversation in 2023 and 2024 has largely resolved; the constraint has migrated upstream to the electrical grid.

Nvidia Blackwell GPU rental prices reached $4.08 per hour in July 2026, up 48% in 60 days. That isn't a chip pricing signal. It's an energy pricing signal: chips without power to run them are queued hardware. The organizations that secured energy contracts while everyone else was counting parameters are now in a structurally different competitive position.

The investment implication is more precise than "AI infrastructure is expensive." It's that the companies worth watching aren't just the ones building models. They're the ones that secured power purchase agreements in 2023 and 2024 — quietly, without press releases, while the rest of the market was obsessing over benchmark leaderboards. Energy contracts, not parameter counts, have become the barrier to entry for frontier AI.

For LatAm, this dynamic presents a genuine structural opportunity. Brazil has an extraordinary renewable energy grid — predominantly hydroelectric and growing solar capacity — that most AI compute markets in the US and Europe would consider scarce. A region that develops the data center infrastructure to pair with this energy abundance could attract AI training and inference workloads that the US grid simply cannot accommodate quickly. The power advantage is already there. The build-out is the open question.

When OpenAI committed $30 billion to land and power rather than to model research, it was communicating something the AI coverage largely missed. The company isn't just betting on the future of intelligence. It's betting that the physical constraint — watts and cooling capacity — will stay scarce long enough to make infrastructure ownership matter. So far, the grid agrees.