Inside Stargate: the AI megafactory buildout reshaping computing

01Why AI broke the data center: compute demand outpacing supply

For most of computing history, the industry assumed hardware would keep up because it always had. The current AI cycle broke that assumption. Frontier models are not trained on ordinary servers; they are trained on clusters of tens of thousands of specialized accelerators that draw more electricity than small cities. Each successive model generation has required roughly an order of magnitude more training compute than the last, a trend documented in detail by research groups such as Epoch AI. Once you see the curve, the building boom makes sense: a training run that fit in one building in 2020 needs a campus today. The bottleneck has shifted from writing algorithms to manufacturing silicon, wiring power, and cooling it — industrial problems with industrial timelines. Data centers already consumed about 4.4 percent of US electricity in 2023, according to the Department of Energy's Lawrence Berkeley National Laboratory, which projects demand could roughly double again by the end of the decade. That projection, more than any single model release, is the backdrop against which Stargate was conceived.

02What Stargate actually is: campus scale, power, and capital

Stargate LLC is a joint venture formed by OpenAI, SoftBank, Oracle, and the Abu Dhabi investor MGX, announced in January 2025 with plans to invest up to $500 billion in US AI infrastructure over four years. Its first physical site is a data-center campus in Abilene, Texas, originally developed with the data-center builder Crusoe. 'Megafactory' is the right mental model: not a server room but a manufacturing plant whose product is computation. Reporting on the project describes buildings measured in hundreds of thousands of square feet, on-site electrical substations, and dedicated cooling capacity, with Oracle committing to lease the completed capacity to OpenAI under multi-year contracts. Bloomberg Originals visited the site with Sam Altman for The Circuit, offering one of the first close looks at the scale involved. The venture's structure matters as much as its headline number: rather than one company building one data center, Stargate distributes capital, land, construction, and operations across partners whose incentives only loosely align — a design choice that speeds construction and complicates accountability at the same time.

Approximate orders of magnitude of total training compute for flagship models; Epoch AI-style estimates, rounded. Values are estimates, not measurements.

03The physics bottleneck: gigawatts, cooling, and grid interconnects

Electricity is the constraint that decides where these campuses go. A gigawatt is the output of a large nuclear reactor, and the Abilene campus and its successors are being planned in precisely those units. Getting a gigawatt to a field in Texas means years of lead time: transmission lines, substation transformers (which themselves have multi-year backlogs), gas turbines or on-site generation, and interconnection agreements with a grid operator. Then the heat has to leave the building. AI accelerators run hot enough that liquid cooling — pumping coolant directly to racks rather than chilling air — is becoming standard for frontier clusters, and water availability has become a genuine site-selection factor. The IEA's 2025 Energy and AI report emphasizes that data-center growth is colliding with grid planners who historically added capacity in single-digit percentage points per decade. In effect, the AI industry is asking the North American grid to do something unprecedented: add city-scale load in months, at rural sites, with power-purchase contracts signed before the buildings exist.

04The money: who pays, who leases, and what the contracts say

The financial architecture is as novel as the engineering. In September 2025, OpenAI and Oracle disclosed cloud contracts reported at roughly $300 billion over about five years — a sum larger than Oracle's entire existing cloud business, payable against capacity that largely did not yet exist. Oracle, in turn, finances construction of the campuses and leases them to OpenAI; SoftBank anchors the equity; Nvidia and other vendors have been reported as taking equity or supplier positions in related deals. Critics point out the circularity risk: an AI startup promising to pay hundreds of billions of dollars, funded in part by the chip vendors who benefit from the purchases, backed by projected AI revenue that is itself growing fast but from a smaller base. Defenders answer that the same structure — specialized builder, long-term take-or-pay lease — is how railroads, telecoms, and hyperscale cloud have always been financed. Both things can be true. The contracts are real liabilities, the buildings are real assets, and whether the demand curve justifies either is the central unresolved question of the buildout.

05Evidence from the field: what Bloomberg saw inside the Abilene site

Field reporting matters here because infrastructure claims are unusually easy to inflate on paper. Bloomberg's visit to Abilene documented the physical reality behind the press releases: shell buildings in various stages of fit-out, rows of racks awaiting accelerators, and the electrical infrastructure that gives the site its rating. The visible sequencing is the story. Power equipment arrives before servers; substations and switchgear are installed years ahead of the chips they will feed; construction crews, not engineers, are the critical path. The site also illustrates why these projects cluster in specific places — cheap land, available grid capacity, wind resources nearby, and state permitting that moves quickly. Abilene's transformation from rail town to AI logistics hub is the clearest public demonstration that the industry's constraint is concrete, copper, and kilowatts. It also grounds the depreciation debate: buildings last decades, but the accelerators inside them turn over every three to four years, which means the money at risk is concentrated in the fastest-aging part of the machine.

Press-reported power commitments, approximate. A gigawatt is roughly the output of a large nuclear reactor; sources: Bloomberg, CNBC, company announcements.

06The risks: depreciation, demand curves, and the bubble question

Every infrastructure boom ends with the same question: did anyone overestimate demand? The bear case for Stargate-class projects has three prongs. First, depreciation: if GPU generations turn over every two to three years, capacity bought today competes against cheaper, faster hardware tomorrow, pressuring prices on the older fleet. Second, demand composition: a large share of current AI revenue comes from inference on chatbots and coding assistants whose willingness-to-pay is still being discovered; enterprise adoption has been real but uneven. Third, concentration: the same handful of labs anchors most of the commitments, so a single stumble in their funding or product cycle ripples through every lease. The bull case replies that inference demand is compounding, that model capability gains keep converting skeptics, and that under-building is the historically more expensive mistake — cloud providers that capacity-starved in the 2010s lost workloads they never recovered. Both sides agree on one thing: the contracts signed in 2025-2026 will not be re-negotiated on 2030's information.

07What it means for the rest of the industry

Stargate's scale changes the operating environment for everyone else. Smaller labs and enterprises will increasingly rent frontier capacity rather than build it, because competing with gigawatt campuses is not a realistic option for most balance sheets. Energy-rich regions — Texas, the upper Midwest, the Gulf states, and internationally, the Gulf states and Nordics — are becoming the compute geography of the 2030s, the way port cities shaped trade geography. Chipmakers gain guaranteed multi-year demand, which stabilizes their capital expenditure and, in turn, the whole semiconductor supply chain. And the grid becomes a strategic asset in a way it has not been since rural electrification: utilities with fast interconnection processes are suddenly indispensable partners. The open question is whether the megafactory model concentrates AI's future into a few linked ecosystems — or, as its architects argue, industrializes compute enough that abundant capacity becomes as boring and as consequential as bandwidth did in the 2000s. Either way, the era of AI as software alone is over; it is now an industry of buildings, contracts, and megawatts.