SpaceX's AI division held discussions with Microsoft this summer about leasing computing capacity, according to people familiar with the matter, signaling Elon Musk's transformation from initial reluctance to full embrace of the next-generation cloud business.

Selling compute capacity has rapidly become one of SpaceX's most profitable businesses. AI labs facing compute shortages, including Anthropic and Google, have signed long-term procurement commitments worth billions of dollars per month, with more orders expected to follow.

The talks with Microsoft remain unclear in progress, and a Microsoft spokesperson declined to comment.

When Musk founded AI startup xAI, the goal was not to build a cloud computing company; earlier this year, xAI was merged into SpaceX. Last year, when xAI's compute scale exceeded its own needs, executives urged Musk to sell capacity externally. Musk rejected the proposal at the time, arguing that once it transformed into a cloud service provider, xAI would struggle to gain industry recognition as an AI lab.

Earlier this year, Musk changed his mind and first launched cooperation with Anthropic. Now SpaceX has secured a large number of external customers and urgently needs to expand compute capacity to fulfill obligations, according to people involved in the projects. After acquiring xAI, SpaceX renamed the AI company SpaceXAI.

Musk posted on X over the weekend that the company plans to bring approximately 420,000 Nvidia GPUs online in November. This compute capacity is located at SpaceX's Minihard data center campus in Memphis, equivalent in scale to multiple large data center clusters combined. The company's original internal target was to have Minihard operational by October 31, but two people involved in the project said adjustments to the data center construction plan may delay the launch.

SpaceXAI needs additional compute to fulfill signed orders, including an agreement announced earlier this month by SpaceX Chief Financial Officer Bret Johnsen for $1.1 billion in monthly revenue starting in December. Anthropic may also increase its investment further. Earlier this year, Anthropic committed to purchasing nearly $45 billion in compute from SpaceXAI; its draft IPO prospectus shows an additional commitment of nearly $40 billion, bringing the total to approximately $84.5 billion.

A reporter contacted a SpaceXAI spokesperson, who only pointed to the official website page introducing the data center expansion. Anthropic did not respond to a request for comment.

Stockpiling Gas Turbines

Musk's ability to rent out compute capacity now stems from his decisive, large-scale investment in data center construction last year, when xAI's large model development was still facing difficulties. The company stockpiled hundreds of millions of dollars worth of gas turbines, even purchasing large amounts of equipment before finalizing site locations, according to people familiar with the matter. One person said Musk and his core team judged that gas turbines would face supply shortages in the coming years; even if xAI could not use all the equipment itself, it could later resell surplus units at high prices.

xAI also studied various solutions to maximize gas turbine power generation, including adding auxiliary units to main gas turbines. While preparing power supply, xAI evaluated hundreds of potential sites, prioritizing areas with local government support and tax incentives. In late last year, Musk met with executives and consultants two to three times a week to discuss site selection, gas turbine suppliers, and potential acquisitions, according to people familiar with the matter. Throughout the project, Musk repeatedly emphasized speed.

This strategy worked: xAI built its first data center, Colossus, in Memphis in just 122 days, deploying 200,000 GPUs. Data centers of comparable scale typically take years to build. But xAI did not put all its compute capacity to internal use. Two people involved in the projects said that even as the company had already begun construction on its second Memphis data center, Colossus II, Colossus's compute utilization rate last year remained below 40%; other sources said utilization was still low this spring.

Idle compute capacity prompted discussions about renting it out. Musk eventually relented. An employee who attended an all-hands meeting said that when Musk introduced the plan to engineers in April, he joked that Anthropic was practically begging to buy xAI's compute. This created an opportunity for xAI to rent out a batch of older Nvidia H100 GPUs, which are relatively inefficient for training the Grok model. Musk told engineers this was only a temporary arrangement and that xAI would not formally enter the cloud computing industry.

However, persuading Musk was only one of the challenges. Two people familiar with the matter said xAI's system was not initially designed to support secure, isolated sharing of the same compute infrastructure among different customers and projects. Directly renting out idle compute carried the risk of Customer A's data and business being leaked to Customer B. People involved said xAI later resolved the isolation technical issue.

During the site selection phase for the Wonder Valley project, xAI held talks with Kevin O'Leary's investment company about jointly developing the Wonder Valley data center project in Alberta, Canada. The two sides ultimately did not reach a cooperation agreement. O'Leary is building a 7,000-acre, $70 billion data center campus in Canada, billed as the world's largest AI computer data center campus. In Saudi Arabia, xAI held talks with developer Ruyah AI about a 1,000-acre site with access to approximately 2 gigawatts of power. SpaceXAI has since partnered with Saudi state-owned AI company Humain to build a 500-megawatt data center there.

One person familiar with the matter said the company ultimately favored regions where local governments were more friendly to data center development. For example, last year xAI built a power plant in Southaven, Mississippi, adjacent to its Memphis data center. Memphis has strict local controls, allowing xAI to deploy only 15 tightly restricted low-emission gas turbines; Mississippi regulators approved a much larger 41-unit installation with relatively looser environmental constraints.

Robotic Wiring

Musk did not just compete for power resources; to complete the Memphis data center cluster, he tried every means to compress the construction timeline. Two people involved in the projects said xAI introduced ABB robots to automate server room wiring. Manually organizing large-scale cabling often takes days; robots can continue working after workers leave and complete wiring by the next morning. The company also advanced modular server rooms and electrical pods, prefabricating some facility modules off-site and then shipping them to the construction site for assembly. This allowed different modules to be built in parallel without waiting for one construction phase to finish before starting the next. Contractors sometimes worked 14 to 16 hours a day.

The rapid compute expansion also left some infrastructure hazards that the company is now addressing. Two people involved in the projects said the drainage system in Colossus II had defects that caused water accumulation, affecting the structural safety of part of the building. Drainage pipe trenches were excavated improperly, and the company now needs to pour additional concrete to reinforce the structure. Remediation work may force other projects to pause, delaying the project by at least 2 to 3 weeks. The company transferred a civil engineer from Neuralink to take charge of drainage remediation, but the engineer left before the problem was resolved.

As SpaceXAI continues to sign compute orders, data center stability has risen to the company's top priority. Two employees who attended meetings said that at an all-hands meeting this summer, SpaceX management called on engineers across the company to rotate in to support the data center in 6-to-8-week shifts. Musk also frequently appeared in Memphis, focusing on ensuring stable server room operations. After SpaceX management took over the project in June, the team suffered a major outage caused by construction problems; another management adjustment followed: Kevin Holland, senior director of Starship, took over the team and reports directly to SpaceXAI President Mike Nichols. Previously the team was led by Wes Sarando, who now reports to Holland.