Cold air, shale gas, and US backlash: why OpenAI wants a $25B AI data center in Patagonia

OpenAI's $25 billion pitch to put an AI data center in Argentine Patagonia is not really about Argentina. It is the clearest test yet of whether US backlash against data centers can be hedged by moving AI compute offshore, and whether a country that actively wants the racks can deliver cold air, gas, wind, and fiber at the scale the largest AI labs need.

In October 2025, OpenAI and Sur Energy signed a non-binding letter of intent for Stargate Argentina, a planned 500-megawatt site framed at up to $25 billion in capital cost (TechRadar, DataCenterDynamics). Stargate is OpenAI's branded AI-infrastructure program for the large-scale facilities that train and run its models. The Argentine side is now anchored by talks with Pampa Energia, the country's largest integrated energy company, to host the build near the Loma de la Lata gas-fired plant, which draws from the Vaca Muerta shale formation (boe Report).

That stack is the reason Patagonia is the test case. The southern tip of South America offers three things the US no longer reliably offers at the scale the largest AI labs need: cold air that cuts cooling load, cheap gas and steady wind for power, and a national government that wants the project. Vaca Muerta, Argentina's main shale gas formation, sits under Neuquén province and feeds the Loma de la Lata plant; the winds along the Chubut coast have already driven wind farms for two decades.

The bet is a hedge, not a relocation. The October 2025 letter of intent is non-binding, no site has been selected, and the source reporting flags "little movement toward selecting a site, let alone breaking ground" (boe Report). OpenAI's other Stargate sites in the US, including flagship builds in Texas and New Mexico, are not being abandoned. Patagonia is an option being priced.

Two other Patagonia pitches sit in the same headlines but are not the same deal. FlexDomes, represented locally by Gabriel Obrador, is reported to be seeking investors for a 120MW, $1.4 billion site also powered by Vaca Muerta gas; there is no signed letter of intent (boe Report). Green Capital, led by Marta Złotkowska, is pitching a 300MW first phase at roughly $3 billion on about 501 square miles south of Trelew in Chubut province, with ambitions to scale to 3,000MW. Reading the three together as a single announcement would conflate a signed LOI with two investor pitches.

The same geography that makes the plan attractive also defines its failure modes. Three variables will determine whether Patagonia absorbs real AI compute or stays a colorful headline.

The first is fiber. Argentina's data backbone to the rest of the world is thin, and Patagonian sites sit far from Buenos Aires and the existing fiber routes. Training a frontier model on remote compute needs steady, high-capacity links, and the source reporting explicitly flags communications infrastructure as a constraint (boe Report).

The second is political durability. The Milei government's openness to foreign capital is part of the draw, but Argentine coverage treats President Javier Milei's long-term political future as a real risk (boe Report). A change in administration, or a balance-of-payments crisis, would reset the welcome mat.

The third is US backlash itself. If state-level opposition to large data centers softens, or if US power grids loosen, the hedge becomes optional. Google, Apple, and Meta are reported to be scouting Patagonia sites too (OpenTools), which suggests the geographic hedge is a sector-wide move, not an OpenAI-only bet.

The Patagonia story is not a wire blip. It is the first named, large-scale test of a new operating principle for AI infrastructure: where racks physically sit is now a permission-and-energy problem, not a chip problem. The next Patagonia-style headline will be readable as a forecast: if the OpenAI-Sur Energy LOI converts to a binding agreement, the bet becomes a template; if Milei's political coalition fractures, or the fiber does not get built, the bet stays a letter of intent.