Inside Stargate: OpenAI's megafactory bet on the compute century

01 What Stargate actually is: the largest compute buildout ever attempted

Stargate is the infrastructure venture that OpenAI announced in January 2025 together with SoftBank, Oracle and the Abu Dhabi-backed investor MGX. The headline numbers were simple and enormous: a stated $500 billion of AI infrastructure in the United States over four years, with $100 billion deployed immediately, starting with a campus in Abilene, Texas. Unlike a product launch, there is nothing for a user to click. Stargate is a bet on buildings, transformers and power contracts.

The Bloomberg Originals segment that anchors this piece gives the project a face. Sam Altman walked The Circuit through a Stargate site in Abu Dhabi, describing the construction effort as a megafactory, a word that until recently belonged to battery plants like Tesla's Gigafactory rather than anything in computing. The tour is worth watching precisely because these sites are usually off limits: security, competitive secrecy and sheer construction hazard keep AI campuses among the least-documented industrial projects of the decade.

A note on evidence is warranted. What is verifiable today is a stack of announcements: the January 2025 unveiling, the May 2025 Stargate UAE agreement with Abu Dhabi's G42, and partner disclosures from Oracle, NVIDIA, AMD and Cisco. What is not yet verifiable is the end state. The gigawatt figures are committed targets rather than measured output, and the $500 billion total is a stated ambition spread over years. Everything in between is interpretation.

05 The megafactory model: why AI campuses are built like chip fabs

The term megafactory is doing specific work. A semiconductor fab, think of the plants TSMC operates in Taiwan and Arizona, is defined by extreme capital intensity, multi-year construction schedules and designs that are replicated across sites so that lessons learned in one location transfer to the next. An AI campus behaves the same way: tens of billions of dollars of accelerators, networking and power equipment, arranged in a layout closer to an industrial facility than to the office-park data centers of the 2010s.

The reason is the physics of frontier training. A large model is not trained on one machine but on tens of thousands of GPUs that must behave as a single computer, which pushes engineering toward standardized halls, high-density racks and liquid cooling loops rather than bespoke buildings. Standardization is also a speed strategy: the binding constraint on AI capacity in 2026 is less the silicon supply than the time it takes to energize a site, so repeating a proven design across Texas, Abu Dhabi and future locations is itself a competitive advantage.

The fab analogy carries a warning, too. Chip manufacturing consolidated into a handful of players because the capital costs crushed everyone else, and a similar dynamic may await AI infrastructure. That reading is interpretation rather than measurement, but it explains why investors, and not just engineers, watch Stargate so closely.

09 Abu Dhabi and the geopolitics of compute alliances

The Abu Dhabi site Altman toured is Stargate UAE, announced in May 2025 as the first international Stargate campus and an early flagship of OpenAI's OpenAI for Countries program. The structure is a joint effort with G42, the Abu Dhabi technology group, with Oracle, NVIDIA, AMD, Cisco and SoftBank named as partners. The stated plan is a one-gigawatt campus, with a first tranche of capacity intended to come online in 2026. The exact split of capital and ownership across the partners has not been fully disclosed, which is worth remembering when reading coverage of the project.

Why Abu Dhabi? The emirate brings three things AI builders struggle to find elsewhere: patient sovereign capital, a serious energy portfolio that spans natural gas, solar and nuclear, and a geographic position that lets a single campus serve Europe, the Middle East and Asia. The UAE has also spent a decade trying to convert oil wealth into a technology sector, and hosting a Stargate campus is the strongest signal yet of that ambition. That is a measured fact about the announcement, whatever one thinks of the strategy.

The deeper story is what analysts have started calling compute diplomacy. Export controls already govern which accelerators can ship where, and a US-linked AI campus on foreign soil raises new questions about who governs the models that run on it. The Bloomberg interview touches on this tension obliquely. Our read is that Stargate UAE is as much an alliance instrument as a commercial data center, but that is analysis, not a claim anyone at OpenAI has made on the record.

13 The numbers: gigawatts, gigaflops, and the economics of scale

Start with the unit. A gigawatt is a billion watts, roughly the output of a single large nuclear reactor unit, or of several hundred utility-scale wind turbines running at full tilt. Conventional hyperscale data centers are measured in the tens of megawatts. Stargate's sites are specified in fractions and multiples of a full gigawatt, which is why grid engineers, not just software architects, are now central to AI planning. A gigaflop, by contrast, is a billion floating-point operations per second, and modern accelerators deliver thousands of them per chip; the campus exists to aggregate that arithmetic at scale.

On the demand side, the International Energy Agency's Energy and AI report estimated that data centers consumed roughly 415 terawatt-hours of electricity in 2024 and projected about 945 terawatt-hours by 2030 in its base case, the equivalent of average power draw rising from roughly 45 gigawatts to over 100 gigawatts. Those are institutional projections with stated uncertainty, not measurements of the future, and the chart below rounds them further. The honest summary is that AI is pushing data center demand onto a growth curve with few precedents in the electricity sector.

The economics follow from the units. Industry estimates put the capital cost of AI-ready capacity at tens of millions of dollars per megawatt before a single model is trained, and most of the operating cost is electricity. Stargate's scale is therefore a financing project as much as an engineering one, which is why a sovereign investor like MGX and a balance sheet like SoftBank's appear in the partner list. Whether the resulting compute earns back its capital depends on demand for AI services that still has to materialize. That is the central unmeasured variable in the whole program.

18 Why compute scale became OpenAI's core strategy

OpenAI is unusual among frontier labs in that, for most of its history, it owned essentially no computers. Google trains on its own TPUs, and Amazon and Microsoft run the largest cloud fleets on earth, while OpenAI rented capacity: first from Microsoft Azure, then from an expanding roster that came to include Oracle and CoreWeave, along with direct NVIDIA and AMD partnerships. Stargate is the logical endpoint of that trajectory. Stop renting, and help finance the factories themselves.

The strategic logic is straightforward even if the execution is not. Training frontier models is compute-hungry, inference at consumer scale is compute-hungry, and capacity sold out faster than it could be built through 2024 and 2025. Owning influence over the supply chain converts capital into a durable constraint on competitors, at least while compute remains the binding limit on model progress.

The honest counterargument is that scale has not always been destiny. Algorithmic efficiency gains, the kind that made smaller training budgets more capable than anyone projected in 2023, could erode the value of any fixed stock of compute, and nobody has measured where the returns to additional scale currently sit. Altman's wager, made explicit throughout the interview, is that compute remains the century's defining resource. That is a thesis, and it is falsifiable.

22 The constraints: power, land, cooling, and talent

Power is the binding constraint, and the industry's response has been improvisational. Grid interconnection queues in the United States can stretch for years, so operators have turned to on-site gas turbines, behind-the-meter generation and, in reported cases such as the agreement to restart a reactor unit at Three Mile Island, direct nuclear contracts. Land in the right places, near fiber routes, substations and water, is scarcer than the empty-desert imagery of AI campuses suggests.

Cooling is the second hard limit. Racks that once drew 10 kilowatts now draw well over 100 at AI density, past the point where air cooling works, which forces liquid cooling loops into the design and pulls water consumption into the public debate. The chart below puts the announced campus sizes side by side. The gap between a flagship Stargate target and a conventional hyperscale campus is roughly an order of magnitude.

The least discussed constraint is people. Building at this pace requires high-voltage electricians, transformer manufacturing slots, turbine supply and site engineers, all of whom are in global shortage. Gigawatts of ambition are only as real as the construction crews that can pour the foundations. That is a measured, practical limit that no amount of capital relaxes quickly.

27 What it means for the AI race and who pays

If Stargate lands even roughly on schedule, the United States and its partners will hold a compute capacity that no other bloc has organized at comparable scale, with Abu Dhabi as the first export of the model. China is building its own AI infrastructure and has emphasized efficiency per chip, partly because export controls cap the accelerators it can buy. That is a measured policy fact, and it shapes the entire competitive map.

Who ultimately pays is the question the interview answers least. The capital comes from investors and sovereign funds today, but the returns are expected from customers: enterprises and governments buying AI at prices that must cover the electricity, the depreciation and the cost of capital. If AI demand keeps compounding, the largest infrastructure buildout since the railway age is underway. If it stalls, Stargate becomes a case study in overbuilding. Both outcomes remain live.

For anyone trying to track this from the outside, the practical advice is boring and useful: follow the permits, the interconnection agreements and the partner earnings calls rather than the headline totals. Buildings, unlike announcements, cannot be revised.