SpaceX’s 10-GW data-center build faces a power test

By Rebecca Stone

SpaceX’s 2027 target calls for 10 GW of additional data-center capacity against 3 GW already in place, a build rate Semianalysis says could match Google’s entire lifetime deployment in 1.5 years.

Two weeks before the analysts’ assessment, Elon Musk said he was “highly confident” SpaceX could build 10 GW of data-center capacity in 2027 alone. Semianalysis called the goal achievable if conditions aligned and said delivery would make SpaceX the fastest-building hyperscaler. The 10 GW figure remains a target. Not operating capacity.

How fast are the sites moving?

Semianalysis used aerial photographs and available information to assess the rapid buildouts of Colossus 1 and Colossus 2. It estimated Colossus 2’s power-generation capacity at 0.4 GW in January and projected 2.4 GW by year-end as hardware appeared on site.

The wider tally found 9 GW of potential data-center capacity across the Colossus and Starbase facilities the analysts track. Neither figure represents completed, operating compute.

The hardware already at Colossus 2 includes 89 gas turbines with about 2.5 GW of combined output and 1,377 Tesla battery Megapacks on campus. Emissions rules will eventually allow only about 1.24 GW of those turbines to operate.

SpaceX has an agreement with Mississippi allowing temporary operation of the excess turbines, with a promised phase-out by July 2027. Some mobile units sit on truck flatbeds, an arrangement Semianalysis says appears intended to skirt operating restrictions while the temporary approval lasts.

Another 1.2 GW is expected eventually from the grid. That leaves a 300 MW gap, which Semianalysis expects the Megapacks to fill through compute-load variation and possibly external power.

Are the GPUs already working?

On September 25, Musk said Colossus 2 had 550,000 GPUs working, presumably inside the existing Macrohard and Macroharder facility. SpaceX was also reported to be pressing to ready the Gun Range and Project SPAM data centers in quick succession.

Aerial images show new buildings and graded ground where little more than loose dirt stood a few months ago. They do not show evidence that GPU racks have been installed at Gun Range or Project SPAM, and Musk’s 550,000 figure remains his claim rather than an independently reported operating total.

The 10 GW target is a facility-capacity ambition, not a disclosed accelerator count or confirmed workload. The available GPU number is the 550,000 Musk cited for Colossus 2.

What could slow the rollout?

Semianalysis says SpaceX routinely begins a build before settling how the facility will receive power, then brings generation in behind it. The model works only if SpaceX sustains the flow of gas turbines and other required equipment. Keeping that stream arriving is itself an achievement, the analysts argue.

Tesla may be producing mechanical and electrical systems in modules ready for installation directly from the factory or on trucks. That in-house approach could shorten lead times that hyperscalers face from outside suppliers, though the mobile-turbine workaround also invites regulatory scrutiny.

The analysts also point to GPU and RAM scarcity, but argue that obtaining usable power is the more immediate data-center problem. AI customers are willing to pay premium prices because current demand far outstrips available compute.

Speed matters in a market where Semianalysis also notes mounting doubts that every hyperscaler will survive the buildout. A finished shell, however, does not guarantee that servers will operate at the intended rate.

Zoning, permits, grid connections, generation limits, water supply and public opposition have repeatedly delayed data centers, independently of whether computing hardware is available. Microsoft CEO Satya Nadella described the bottleneck plainly: “You may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today.”

Other reports have likewise put a significant percentage of AI accelerators in warehouses while they await power and rack readiness. Colossus 2’s apparent speed therefore does not guarantee the same result at SpaceX’s other sites.

Hardware and materials can still run short, legal conditions can change, and local opposition can harden. Those risks apply even to a company with Elon Musk’s resources, while the comparison does not assume uninterrupted turbine deliveries, grid access or every necessary permit.

Can SpaceX scale beyond Colossus?

Two data centers at Starbase are undergoing rapid construction, which Semianalysis reads as evidence that SpaceX can reproduce the Colossus approach. The analysts credit Tesla’s manufacturing role for keeping much of the supporting infrastructure and power-generation hardware in-house, potentially bypassing supplier delays.

Even if SpaceX misses the “10 GW in a year” target, the visible pace of Colossus could still put it ahead of rivals. The next test is whether the projected 2.4 GW power ramp and September 25 GPU claim hold at usable scale before the July 2027 turbine phase-out; if they do, scarce AI compute could reach customers sooner and at premium prices, sharpening the race against hyperscalers still waiting for grid connections.