• SpaceX aims to assemble a 20 GW project portfolio, with Musk’s base expectation closer to 15 GW after potential delays
• Compute capacity reached 1.4 GW in 2Q26 and is expected to exceed 2 GW by year-end
• SpaceX will build exclusively on Nvidia Vera Rubin, while reserving most future capacity for external customers

SpaceX CEO Elon Musk outlined a plan to assemble a portfolio of projects capable of delivering 20 GW of power and cooling infrastructure for artificial intelligence compute by the end of 2027.

Musk acknowledged that 20 GW is an aspirational target rather than his base forecast. Assuming about one-quarter of the projects take longer than planned, he expects SpaceX could have close to 15 GW of power infrastructure online by the deadline.

“Our tentative target is to actually have 20 GW at the power and cooling level online by the end of next year,” Musk said on the company’s 2Q26 call, its first since going public in June. “Now, I don’t think we’re going to achieve 20 GW, but we want to have a series of projects that cumulatively come to 20 GW.”

SpaceX ended the second quarter with 1.4 GW of nameplate compute, up from 1 GW in the first quarter and 400 MW a year earlier. The company expects to surpass 2 GW by year-end. Musk initially described year-end 2027 compute as likely “closer to 10 GW than 5 GW,” before distinguishing installed GPU capacity from the larger power-and-cooling target.

SpaceX plans to deliberately build more power, cooling and electrical infrastructure than it initially has GPUs available to occupy. Musk said the approach is logical because GPUs are substantially more expensive than the balance-of-system infrastructure, allowing SpaceX to install chips as supply becomes available.

The company has decided to build exclusively around NVIDIA technology, citing its Vera Rubin architecture as the strongest available AI platform. Musk said SpaceX expects to receive a “very significant percentage” of Nvidia’s GPU output in 2027.

Trivial problem

SpaceX also plans to begin launching its Starmind AI satellite next year. Musk described Starmind as an optimized Vera Rubin NVL72 computer designed for orbit. The company intends to deploy the same form factor terrestrially, arguing that it will be simpler, less expensive and more effective than conventional rack configurations.

Musk said SpaceX’s experience engineering reusable rockets and Starlink satellites gives it an advantage in data center construction.

“The terrestrial data centers are a trivial problem compared to making gigantic reusable rockets,” he said, arguing that applying even a small portion of SpaceX’s aerospace expertise produces faster and more efficient deployment.

CFO Bret Johnsen said current compute contracts are producing payback periods of less than one year. SpaceX signed cloud services agreements totaling USD 14.1bn and spent USD 15.8bn on AI infrastructure during the quarter.

Over time, Musk expects only about 10% of SpaceX’s compute to support internal Grok training. The majority would be used for inference or leased to third parties for training and inference workloads.