SpaceX AI Cloud

Can SpaceX Build a 13GW AI Cloud Before Regulators Stop It?

SpaceX may soon run the largest AI cloud on Earth. SemiAnalysis made that case in a November 3 report, and the numbers are hard to ignore.

Two months earlier, the firm called a 10GW SpaceX buildout for 2027 feasible. Now it says SpaceX has raised its own target to about 13GW. The analysts count 3.6GW under construction and more than 9GW of identified capacity.

For scale, they compare 13GW to everything Google built from 1999 to 2026. SpaceX would match that in about 1.5 years.

These figures are analyst estimates. They rely on satellite imagery, local filings, and supply chain tracking. SpaceX has not confirmed the site details.

Three clocks decide the race

Every AI buildout runs on three clocks:

  • Build clock: how fast crews raise buildings.
  • Power clock: how fast electricity reaches the racks.
  • Permission clock: how fast regulators, utilities, and neighbors say yes.

That framing is mine, not SemiAnalysis’s. SpaceX leads on the first clock and improvises on the second. The third will likely decide how this ends.

The build clock

SemiAnalysis says most developers fight for 8 to 12 months per datacenter. SpaceX appears to have cut that in half.

Its own records keep falling:

  • The Colossus 1 retrofit took 122 days for 150MW.
  • A Colossus 2 build took 91 days for 210MW.
  • Project SPAM, also at Colossus 2, broke ground October 5 and targets first capacity in December.
  • A Mississippi plot nicknamed “Gun Range” should finish in about four months.

The analysts say Colossus 2 added nearly two Stargate Abilenes (2GW) in just under a year. Rivals chase the same goal. OpenAI’s $600B compute plan rests on the same scaling bet.

Why is SpaceX faster? The analysts credit its engineers, with labor running near 3,000 workers per gigawatt. They also suspect Tesla makes modular electrical and mechanical equipment for SpaceX sites. That would skip the lead times that stall other builders. They present it as a belief, not a confirmed fact.

Rocket pads and chip fabs become datacenters

The surprise is where capacity appears.

Starbase, Texas. Crews cleared ground in late August. By late September, vertical construction was under way. The site will house two data centers, and SemiAnalysis thinks it could run well over 1GW.

TeraFab. This is SpaceX’s planned 10GW+, 100-million-square-foot chip fab. The analysts have medium-high confidence it will also host a data center. Local filings point that way. SpaceX representatives did not say only SpaceX would use the capacity, which leaves room for outside tenants.

One caveat applies. The pads could still turn out to be fab support, such as onsite gas or a central utility plant.

The power clock

SemiAnalysis tracks more than 8GW of likely behind-the-meter power for SpaceX. The company seems to pick power sources mid-build instead of waiting on utilities.

At Starbase, turbines arrived one month after vertical construction began. The site may use more than 1GW of them, with two pipelines under way. The analysts suspect SpaceX will move most of Colossus 2’s temporary turbines there.

Colossus 2 is heading the other way, toward the grid. TVA filings show service requests of 700MW (Macrohard) and 450MW (MiniHard). SpaceX is also building a 500kV substation, due by December 31, 2026.

The timing matters. A binding requirement signed July 30, 2026 cuts Southaven’s legal turbine capacity from 2.7GW today to 1.24GW by July 2027.

Here is the math. A roughly 1.2GW grid hookup covers part of the load. That leaves about 1.3GW of compute to feed from 1.24GW of nameplate generation. Nameplate output loses some to efficiency and derating, and IT load understates total facility load. SemiAnalysis sees a gap of 300MW or more.

Its answer is storage. Colossus 2 is installing 5.9GWh of batteries, which could become the largest deployment in the world. Turbines would charge the batteries while the batteries lift baseload. The analysts add three helpers:

  • A very low PUE of about 1.10 to 1.15.
  • A mix of customer-facing and internal workloads, which allows uptime tiers.
  • Lighter racks. Project SPAM uses half the fiber of others.

Gigawatt-scale AI training swings create real grid risk, so this design carries engineering risk too.

Epoch AI projects that power will bite first as AI scales. SpaceX is trying to solve that problem on site.

The permission clock

SpaceX can pour concrete in four months. It cannot control what regulators, courts, and neighbors do.

The warning signs are already visible:

  • The Southaven cap shows regulators can force SpaceX to cut generation.
  • SpaceX is trying to buy hundreds of homes near Colossus. The Guardian reports the aim is to ease noise complaints and house construction workers.
  • SemiAnalysis thinks TeraFab’s vertical work moves slowly on purpose. The goal is to keep its permits clean and avoid pauses from Texas regulators.

Others face the same wall. Alberta’s off-grid gas and GPU campus now faces judicial review of its environmental exemption.

Voters have a say too. One count found 21 datacenter ballot measures across 18 Ohio jurisdictions.

What it means for customers

SemiAnalysis argues that deals with Anthropic and Google already make SpaceX a very profitable neocloud. Fortune reports a catch. Those contracts include a 90-day cancellation provision. It lets SpaceX reclaim compute if Grok’s needs jump.

Renters get speed, and landlord risk comes with it.

SiteWhat analysts seeMain uncertainty
Colossus 2 (Memphis and Mississippi)Record builds, grid hookup, 5.9GWh of batteriesTurbine cap of 1.24GW from July 2027
Starbase, TexasTwo data centers, well over 1GW possibleFinal power mix and size
TeraFab, TexasLikely datacenter beside a 10GW+ chip fabDatacenter or fab support

The takeaway

Speed is now a product. A cloud that delivers capacity in four months wins buyers who cannot wait a year.

Power strategy is the moat. Gas, batteries, and a selective grid hookup cut dependence on utility queues.

Size multiplies exposure. Every added gigawatt brings another permit, another neighbor, and another possible pause.

FAQs

Q. How big is 13GW?

SemiAnalysis says it matches Google’s total data center capacity built from 1999 to 2026. SpaceX would add it in about 1.5 years.

Q. Has SpaceX confirmed these numbers?

The 13GW figure comes from SemiAnalysis, which says SpaceX raised its own estimate. The site-level details come from the analysts’ own research.

Q. Why does the Southaven cap matter?

It forces SpaceX to lean on the grid and batteries at Colossus 2. That shift creates the gap of 300MW or more that the analysts model.

Q. Will SpaceX rent compute to outsiders?

The analysts think so. The TeraFab filings and the lack of any exclusivity language support that view.

Related: Google Put AI Chips in Space. The Real Problem Isn’t Power

Tags: