SpaceXAI's Colossus 2 data center in Memphis, Tennessee, has claimed the title of the world's largest AI data center, reaching approximately 946 MW of IT power capacity. The facility became operational on September 3, 2026, and houses an estimated 1,112,000 H100-equivalent compute units built on Nvidia's B200 and B300 chips — a metric that expresses total processing capacity in terms of Nvidia's widely deployed H100 GPU for comparability across facilities. The project required $35.8 billion in capital investment, a figure that puts a single site on par with the annual infrastructure budgets of the largest traditional cloud providers.

How SpaceXAI built the world's biggest AI factory

The Colossus 2 build-out reflects a strategy that prioritizes speed over convention. Rather than waiting years for grid connections and permanent power infrastructure, SpaceXAI deployed temporary natural gas turbines alongside limited grid power to bring the facility online quickly. That approach reflects a broader industry bottleneck: grid interconnection queues in the United States routinely take years, while AI developers measure competitive advantage in months.

The temporary turbines have been surrounded by legal and regulatory challenges, including questions about permits and environmental compliance. The plan is to eventually remove roughly 69 of these temporary units once a permanent 1.2 GW gas power plant comes online by mid-2027.

The broader Colossus cluster, which includes the original Colossus 1 facility launched in 2024, had already reached or exceeded 1 GW of total AI compute capacity by mid-2026. Compute capabilities at the cluster have roughly doubled every seven months since Colossus 1 first went live.

Projections for early 2027 show the facility scaling to 1,531 MW with 1,824,000 H100-equivalent compute.

The neocloud play: selling picks during the gold rush

SpaceXAI has contracted substantial compute capacity to third parties, with Anthropic, Cursor, and Google among the notable lessees. This positions SpaceXAI not merely as a consumer of AI compute for its own Grok training system, but as a merchant provider of AI infrastructure. Notably, Google — itself a hyperscaler with its own AI chips and data centers — leasing capacity from a rival's facility underscores how scarce top-tier compute remains relative to demand.

The power problem nobody can ignore

According to Epoch AI's tracking, the largest AI data centers have doubled in capacity every 10 months since mid-2024. That pace may slow after early 2027, as the easy wins in rapid deployment give way to harder infrastructure challenges.

The permanent 1.2 GW gas power plant expected by mid-2027 should resolve the immediate permitting issues, but it also locks in natural gas as the primary energy source for what may become the world's single largest concentration of AI compute — a choice that ties the facility's long-term economics and emissions profile to fossil fuel supply chains.

What this means for the AI compute market

SpaceXAI's emergence as a major merchant compute provider reshuffles the competitive landscape. Traditional hyperscalers — Amazon Web Services, Microsoft Azure, and Google Cloud — have long dominated the market for renting out AI compute, and Colossus 2 gives AI lab customers a major alternative supplier for the first time at this scale.

If SpaceXAI hits its early 2027 targets of 1,531 MW and 1.8 million H100-equivalent compute, the gap between Colossus 2 and the next largest facility will only widen.