SpaceX aims to be the world's largest AI cloud: ~13 GW in 2027 with Colossus 2, Starbase and TeraFab

Disclaimer: what could be downloaded is only the free portion of a paywalled SemiAnalysis article (November 3, 2026). Everything below comes from that free excerpt; the in-depth analysis of power solutions, another possible site and pricing of future contracts sits behind the paywall and is not covered here. The figures are the firm's own estimates, not data confirmed by SpaceX except where noted.

The core thesis is that SpaceX is building data center capacity at an unprecedented pace. According to the authors, two months ago they argued that the 10 GW plan for 2027 was feasible and drew heavy criticism from those who called it impossible. They now say they see 3.6 GW under construction across several sites, both new and existing, and note that SpaceX has revised its estimate upward to around 13 GW in 2027. The firm compares this to all of Google's data center capacity built between 1999 and 2026, delivered here in 1.5 years. They say they believe it is possible and have already identified more than 9 GW.

The authors argue that SpaceX could be the fastest data center builder in the world, even ahead of its Chinese peers. As an example, Colossus 2 reportedly added nearly two “Stargate Abilenes” (about 2 GW) in just under a year. They also note that the power under Elon Musk's management (more than 8 GW) is lower than the total data center power coming online (more than 9 GW), and that at these sites the power supply is often figured out on the fly.

On construction timelines, the article cites several milestones. The “MiniHard” building in Memphis was reportedly completed in about five months. The “Gun Range” site, recently acquired just south of the state line in Mississippi, would take about four months; according to SemiAnalysis, it would be the fastest greenfield data center built at that scale, compared with the 8-12 months typical in the industry. The first conversion at Colossus 1 reportedly took 122 days for 150 MW, and at Colossus 2 just 91 days for 210 MW. The “Project SPAM” building, at the center of Colossus 2, broke ground on October 5, 2026 and aims to bring its first phases online in December, which could beat the 91-day record. The authors also note that SpaceX is trying to buy hundreds of nearby homes, both to ease noise complaints and to house construction workers.

A notable development is the conversion of facilities belonging to other Musk companies. Starbase (Texas), traditionally dedicated to building and launching rockets, will host two data centers. According to the article, the land went from being cleared in late August to having full vertical construction by late September, and SemiAnalysis believes the data halls will be fitted out in the coming weeks. On a 2027 timeline (that is, with Vera Rubin chips), they estimate the site could exceed 1 GW. They observe several construction processes running in parallel, such as steel structure, roofing and site preparation.

Another point is TeraFab, SpaceX's chip fab project in Texas (described as 100 million square feet and more than 10 GW). SemiAnalysis has medium-to-high confidence that it will also include a data center. They say the project started as a data center and then quietly pivoted to TeraFab, and they support their hypothesis with filings submitted to local authorities and an August 2026 recording in which SpaceX representatives speak before Grimes County officials. In it, according to the article, there is no confirmation that the capacity is solely for internal use, leaving open the possibility of leasing it to third parties. The open question is whether the current works correspond to one or more data centers or to fab support facilities (on-site gas, central utility plant, etc.). The firm leans toward the former, and suspects vertical construction is advancing slowly to ensure compliance with its permits and JETI agreements and to avoid TCEQ shutdowns or other regulatory scrutiny. They acknowledge the work is so early that the pads could be repurposed.

How do they build so fast? The authors attribute part of the answer to Musk's companies being among the best hardware companies in the world. They cite a workforce of about 3,000 people per GW thanks to top-tier engineers, and believe Tesla manufactures modular mechanical, electrical and plumbing (MEP) equipment for SpaceX. They compare it, at large scale, to Crusoe building its own switchgear to avoid lead times. They think Tesla will end up being more than a battery (BESS) supplier to SpaceX.

On power, SemiAnalysis tracks more than 8 GW of likely “behind-the-meter” capacity for SpaceX, including deals with FTAI, Solaris and others it details only to subscribers. At Starbase, APR turbines are being deployed just one month after vertical construction began; the site could use more than 1 GW of APR, although given the size of the buildings they believe it will not be enough. Two gas pipelines are underway, so they expect several GW and more turbines. According to a TikTok featuring the mayor of Brownsville, a fiber-optic channel is being built through the Port of Brownsville and leasing waterfront space is being considered, perhaps for a dedicated port. The authors suspect most of Colossus 2's temporary turbines will be relocated there.

At Colossus 2 the opposite is happening: a grid connection. According to TVA, Macrohard (700 MW) and MiniHard (450 MW) have applied for service. SpaceX is building a 500 kV gas-insulated substation (GIS) next to MiniHard, scheduled for December 31, 2026 and connected to the Freeport substation. It arrives just in time because the turbines at the Southaven plant will have to shut down: the site can legally operate 2.7 GW of nameplate capacity now, but only 1.24 GW from July 2027, under a binding requirement signed on July 30, 2026.

The authors point to a possible shortfall: subtracting the roughly 1.2 GW grid connection, 1.3 GW of compute would remain against 1.24 GW of nameplate generation, which also suffers efficiency and derating losses, and IT power does not equal the facility's total load. There could be a gap of up to 300 MW. Their hypothesis is that SpaceX has built the largest battery farm in the U.S. for this purpose (close to 6 GWh, with 5.9 GWh already being installed), so that gas and batteries work together to raise baseload while the turbines charge the batteries. To make the numbers work, they point to an ultra-low PUE of about 1.10-1.15, the variety of data center types (inference versus training, customer-facing versus internal use), lower-power racks (Project SPAM would use half the fiber of others) and the oversized battery; they also mention possible availability SLA tiers.

Overall, this is an analysis by a firm that sells its data models and claims past successes, such as having called Colossus 2 the first gigawatt data center in September 2025. The timeline, power and capacity figures come from imagery and records interpreted by SemiAnalysis, not from official confirmations; they should be treated as estimates. Most relevant for readers is the hypothesis that SpaceX's edge lies in execution speed, vertical integration with Tesla and improvised power solutions, and the possibility that it will lease capacity to third parties, including the future TeraFab site.