What happened
In a report published November 3, 2026, SemiAnalysis says SpaceX has revised its estimate for AI datacenter capacity in 2027 from about 10 GW to about 13 GW. The authors say they see approximately 3.6 GW under construction across multiple existing and new sites, and have identified more than 9 GW of capacity. They characterize the proposed buildout as potentially the largest amount of datacenter capacity built by one company over that timeframe. These are SemiAnalysis estimates and targets, not a report of 13 GW already online.
The report describes expansions around Colossus 2, including a newly purchased site called Gun Range in Mississippi. SemiAnalysis estimates that construction there could take about four months. At the Colossus 2 site, a building referred to as Project SPAM began construction on October 5, 2026; the report says SpaceX targets bringing its first tranches online in December. For comparison, the authors say earlier Colossus retrofits took 122 days for 150 MW at Colossus 1 and 91 days for 210 MW at Colossus 2. The report presents these as unusually fast timelines, not guarantees that later projects will match them.
SemiAnalysis also describes datacenter construction at Starbase, Texas, a SpaceX rocket-building and launch campus. It says the site could exceed 1 GW, based on the apparent scale and a 2027 timeline, while noting that this is an estimate. Another possibility is a datacenter associated with TeraFab, a proposed 100-million-square-foot, 10-GW chip-fabrication project in Texas. The authors say they have medium-high confidence that TeraFab will have a datacenter, but acknowledge that current work could instead be supporting infrastructure and that plans could change.
The report discusses power provision and battery storage as parts of the buildout. It says a 5.9-GWh battery energy storage system is already being installed at Colossus 2, with more potentially to come, and describes the use or relocation of mobile turbines. It also argues that building and power capacity may be developing on different schedules. The supplied material does not provide a general customer access policy, pricing, or a confirmed date when the full 13 GW target would be available to outside users.
What the archive adds
Our earlier Datacenters 2027: The Year of Acceleration, +50GW in a Single Year noted SemiAnalysis’s forecast that more than 50 GW of data-center capacity could come online in 2027. This is a new fact on that broader infrastructure story: the report now assigns SpaceX/xAI a higher target and gives site-level estimates, rather than merely restating the industry-wide forecast.
Trend signal
This supports the agent economics moves to completed work pattern only indirectly: more compute could expand the supply available for AI workloads, but the report does not establish cost per completed task or compute utilization. The more direct signal is that power, sites and construction throughput are becoming central constraints on AI capacity; none of the supplied patterns specifically covers datacenter construction.
Why now
SemiAnalysis frames the revision as an update two months after its earlier discussion of a roughly 10-GW 2027 ambition. It says construction has advanced and identifies sites that were not previously included in the public picture. The timing matters because the target is tied to a 2027 horizon, but the report does not establish that the full capacity will be completed by then.
Connections to other events
An earlier Telegram post described the 2027 datacenter expansion as an infrastructure-financing and energy problem, not simply a matter of ordering GPUs. SemiAnalysis’s current report adds a company-level example: it links rapid construction ambitions to power equipment, battery storage and site reuse. The connection is about the same pressure on physical infrastructure, not evidence that every projected site will be built.
Why it matters
For teams building AI products, a large announced capacity target is not the same as usable compute. If the reported projects are completed and connected to power, they could widen the pool of infrastructure available to model developers and cloud providers. But teams planning workloads should distinguish between capacity under construction, sites identified, and capacity actually online and contractable.
The report also illustrates why infrastructure forecasts need more than a headline gigawatt figure. Construction pace, grid access or on-site generation, storage, permits and customer commitments all influence when capacity becomes useful. SemiAnalysis discusses several of these factors, but the public material here does not establish which portion of the proposed buildout will serve third-party customers or on what terms. For builders, the practical implication is to treat these figures as a signal of potential supply, not as a basis for assuming lower prices or guaranteed availability.
Limits and uncertainty
The figures come from SemiAnalysis’s analysis and include targets, estimates and identified sites; the supplied material does not independently verify them. “Under construction” and “identified” do not mean powered, commissioned or available to customers. The report itself is partly paywalled, and the excerpt provides no complete project-by-project schedule, customer contracts, pricing, or confirmed external access terms. It also says TeraFab’s early work could be redirected and that the possible datacenter there is an inference. No conclusion about actual market prices or delivered compute follows from the capacity target alone.
Reader takeaway
SemiAnalysis’s update raises the scale of SpaceX/xAI’s 2027 ambition and gives concrete examples of sites and construction schedules. For AI teams, the key distinction is between projected, building and usable capacity: follow commissioning and power milestones before treating the reported gigawatts as supply you can buy.