How SpaceX’s AI Unit Turned Itself Into an AI Cloud Firm
The Takeaway
- SpaceX’s AI unit transformed into a lucrative AI cloud firm.
- Anthropic committed nearly $84.5 billion for SpaceXAI compute capacity.
- SpaceX plans to bring 420,000 Nvidia GPUs online in November.
SpaceX’s AI unit held talks over the summer about leasing computing capacity to Microsoft, say people familiar with the matter, a sign of how fully Elon Musk has embraced the neocloud business after his initial reluctance.
Selling computing power has quickly become one of SpaceX’s most lucrative businesses, generating billions of dollars a month in commitments from compute-constrained AI labs including Anthropic and Google, with more deals still to kick in. The status of the Microsoft talks isn’t clear, the people said. A spokesperson for Microsoft declined to comment.
Becoming a cloud computing firm was not what Musk originally set out to do with his AI startup, xAI, which he merged into SpaceX earlier this year. At one point last year, when xAI had more capacity than it was using, executives urged Musk to sell computing power to other AI companies. He rejected the notion, saying xAI would not be taken seriously as an AI lab if it became a cloud provider, according to a person familiar with the discussions.
Musk changed his mind earlier this year, starting initially with Anthropic. And now SpaceX has lined up so many outside customers that it urgently needs more capacity to serve them, according to people with knowledge of the project. (Since SpaceX acquired xAI, it has renamed the AI firm SpaceXAI).
The company is scheduled to bring around 420,000 Nvidia graphics processing units online in November, Musk wrote on X over the weekend. The capacity—twice the size of many data center campuses—is part of an expansion of SpaceX’s Memphis data center campus, Minihard. The company had an internal deadline of bringing Minihard online by Oct. 31, but some changes to the build-out strategy could slow the timeline, according to two people with knowledge of the project.
SpaceXAI needs the extra capacity to fulfill some incoming commitments, including a $1.1 billion a month deal slated to start in December that SpaceX Chief Financial Officer Bret Johnsen announced earlier this month, the people said. Anthropic may have also increased its spending. The company committed to spend nearly $45 billion on SpaceXAI’s computing capacity earlier this year but in a draft IPO prospectus it disclosed an additional nearly $40 billion spending commitment for SpaceXAI compute, according to a report from Reuters.
When reached for comment, a spokesperson for SpaceXAI referred back to a website detailing the company’s data center expansion. Anthropic didn’t respond to a request for comment.
Stockpiling Turbines
That Musk is in a position to rent out capacity reflects his willingness to invest aggressively in building data centers last year, even as xAI was struggling with its efforts to develop AI models.
The company stockpiled hundreds of millions of dollars’ worth of gas turbines, buying much of the equipment before it had even settled on a site to install them, according to people familiar with the effort.
Musk and his lieutenants reasoned that turbines would be in short supply in the years to come, and if xAI ended up with more equipment than it needed, the company could sell the excess later at a higher price, one of the people said.
XAI also explored ways to maximize the amount of power it could generate from its turbine equipment, including attaching additional units to primary gas turbines, according to the person.
Along with lining up power, xAI evaluated hundreds of potential data center sites and prioritized those with support from local government and tax abatements. As part of the effort, late last year Musk was meeting two to three times a week with a group of company executives and advisers to discuss potential data center sites, turbine suppliers and possible acquisitions, the person said.
Throughout the process, Musk emphasized speed. His approach worked—xAI was able to bring the company’s first data center—Colossus, located in Memphis—online with 200,000 GPUs in just 122 days. Typically it takes years to build a data center of that size.
But xAI wasn’t using all of that computing capacity. Last year, the utilization rate for Colossus was below 40% even as the company was building its second Memphis data center, Colossus II, according to two people familiar with the project. As of this spring, the company still had low utilization rates, one person familiar with the issue said.
That’s what prompted the discussions about renting out capacity. Musk eventually came around. When he outlined the idea to engineers at an all-hands meeting in April, he joked that Anthropic was practically begging for xAI’s compute, creating an opportunity for xAI to offload its stockpile of older-generation Nvidia H100 graphic processing units that were less efficient for training its Grok models, according to a person who attended the meeting. Musk told engineers it would be a temporary arrangement and said xAI was not entering the cloud computing business, the person said.
But persuading Musk was only part of the challenge. The xAI system wasn’t originally designed to let different customers or projects safely share the same computing infrastructure, two people said. That made it difficult to rent out unused capacity to outside companies without risking that one customer’s data or work could be exposed to another. xAI has since resolved this issue, one of the people said.
Wonder Valley
In its search for sites, xAI had discussions with Kevin O’Leary’s investment firm about working together on Wonder Valley, a data center development planned by O’Leary in Alberta, according to a person familiar with the discussions. XAI didn’t end up joining forces with O’Leary, who is developing a $70 billion, 7,000-acre data center park in Canada he has billed as the “largest AI computer data center park on Earth.”
In Saudi Arabia, xAI held discussions with developer Ruyah AI about a potential 1,000-acre site with access to about 2 gigawatts of power, according to a person familiar with the talks. (SpaceXAI has since partnered with Saudi Arabia’s state-backed AI firm, Humain, to build a 500 MW data center in the country.)
The company ultimately favored locations where local governments would be friendlier to data center developers, one person said. For example, last year xAI chose to build its power plant in Southaven, Miss., bordering its Memphis data centers. While Memphis enforces strict local controls capping xAI at 15 heavily restricted, low-emission turbines, Mississippi’s state regulators have permitted a far larger 41-turbine operation with less stringent environmental limits.
Robot Cabling
Musk didn’t just move fast to secure power. To build the Memphis data center campus, he also found ways to compress the construction timeline.
XAI brought in robots from ABB to automate data center cabling, according to two people familiar with the effort. Workers typically spent days organizing large numbers of cables, and the idea was that the robots could continue working after crews left and have the cables organized by the next morning, one person said.
The company also worked on modular data centers and electrical rooms, building portions of the facilities elsewhere and transporting them to the final site, according to two people familiar with the effort. That approach allowed crews to work on different components simultaneously rather than waiting for one phase of construction to finish before beginning the next. Contractors sometimes worked 14- to 16-hour days, and several workers were injured during construction, according to people familiar with the projects.
The race to build out computing capacity quickly has also left behind some unresolved infrastructure problems the company is now working to fix.
At Colossus II, a drainage problem has caused flooding and compromised the structural integrity of part of the building, according to two people familiar with the project. The drainage was improperly dug out, and the company now needs to pour additional concrete and reinforce the structure, the people said.
The work could require other construction activity to stop, delaying the project by at least two to three weeks, according to the people. The company brought in a civil engineer from Neuralink to oversee fixing the drainage concerns, but that person left before the issue was resolved, the people said.
As SpaceXAI has continued to sign new compute deals, the data centers and their reliability have become a top priority for the company. During an all-hands meeting over the summer, SpaceX leadership asked engineers across SpaceX to sign up for six- to eight-week rotations to assist with the centers, according to two people that attended the meeting. Musk has also become a frequent sight in Memphis as SpaceX attempts to change the company’s build strategy to focus on reliability.
In the weeks since SpaceX managers took over leadership of the project in June, the team experienced a high-profile outage caused by construction issues, as well as a second leadership shift after SpaceX’s senior director of Starship, Kevin Holland, took over leadership of the team and began reporting directly to SpaceXAI President Mike Nicolls. Previously, SpaceX’s Wes Salandro had overseen the team. Salandro now reports to Holland.
