Renting out compute power is generating monthly revenue expectations in the billions of dollars for SpaceXAI. This business, which Musk once rejected, has now become a scarce resource that AI companies are racing to lock in.

Recent reports citing people familiar with the matter indicate that SpaceXAI held discussions with Microsoft this summer about compute leasing. At the same time, a compute leasing contract worth $1.1 billion per month is set to begin in December. Reports also state that Anthropic had previously committed nearly $45 billion in spending on SpaceXAI compute, and subsequently disclosed an additional commitment of nearly $40 billion in compute spending.

The shift in Musk's stance initially stemmed from xAI's low compute utilization rate, while external AI companies were eager to secure compute. Faced with idle older-generation H100 GPUs, Musk ultimately decided to first rent them out to companies like Anthropic. As external demand continued to grow, this business originally viewed as a temporary arrangement gradually evolved into sizable compute orders.

As orders increased, SpaceXAI also accelerated its infrastructure expansion. Musk posted on the X platform over the weekend that the company plans to bring approximately 420,000 NVIDIA GPUs online in November, a scale equivalent to twice that of an ordinary data center campus.

Musk initially refused to "sell compute," then changed his mind

Renting out compute was not Musk's original intention. Musk merged xAI into SpaceX earlier this year and renamed it SpaceXAI. People familiar with the matter revealed that when xAI had idle compute capacity last year, executives suggested selling compute to external AI companies, but Musk explicitly refused on the grounds that if it became a cloud service provider, xAI's positioning as an AI lab might no longer be taken seriously by the outside world.

The turning point came earlier this year. When Musk discussed the idea with engineers at an all-hands meeting in April, he even joked that Anthropic was practically "begging" xAI for compute. This happened to provide an opportunity for xAI to absorb older-generation NVIDIA H100 GPUs in its inventory, which were less efficient at training the Grok model. A report citing an attendee stated that Musk said at the time this was only a temporary arrangement and emphasized that xAI was not entering the cloud computing business.

However, persuading Musk was only the first step. A report citing two people familiar with the matter revealed that xAI's systems were initially not designed to support different customers or projects securely sharing the same computing infrastructure, so renting idle compute to external companies carried the risk of leaking customer data or work content to one another. Currently, xAI has resolved this technical issue.

Aggressive stockpiling and ultra-fast construction: a data center built in 122 days

Behind this business transformation is Musk's major bet on data center construction last year, when xAI's AI model research and development work was still struggling.

A report citing people familiar with the matter revealed that xAI even stockpiled hundreds of millions of dollars worth of gas turbines before installation sites had been determined. Musk and his team judged that turbines would face supply shortages in the coming years; even if they bought too many, surplus equipment could be sold later at a markup.

In terms of site selection, xAI evaluated hundreds of potential data center locations, prioritizing areas where local governments were friendly and could offer tax breaks. At the end of last year, Musk met with a group of executives and advisors two to three times a week specifically to discuss site selection, turbine suppliers, and potential acquisitions.

In terms of construction speed, xAI set a record by putting its first data center, "Colossus," located in Memphis with 200,000 GPUs, into operation in just 122 days, whereas building a data center of comparable scale normally takes years.

To compress the construction timeline, xAI brought in ABB's robots to automate data center wiring and adopted a modular construction approach, prefabricating some facilities at other locations before shipping them to the site for assembly. Contractors sometimes worked 14 to 16 hours a day.

The cost of expansion: low utilization, flooding, and management turmoil

However, aggressive expansion also left behind many hidden problems.

Last year, Colossus's GPU utilization rate was below 40%, while the company was simultaneously building a second Memphis data center, Colossus II. A report citing a person familiar with the matter stated that the company's utilization rate remained low in the spring of this year, which was the direct trigger that pushed Musk to change his position and ultimately decide to rent out compute externally.

Colossus II currently faces more troublesome engineering problems. Two people familiar with the matter revealed that the project's drainage system had defects, causing flooding within the facility and affecting the structural integrity of some buildings. The related repairs required additional concrete pouring and structural reinforcement, which could force other construction work to pause and delay the project by at least two to three weeks. The company borrowed a civil engineer from Neuralink to take charge of the repair work, but that person left before the problem was resolved.

The management level also experienced turmoil. Since SpaceX management took over project leadership in June of this year, the team has suffered a large-scale outage caused by construction problems and a management reshuffle. SpaceX Starship senior director Kevin Holland took over the team and reports directly to SpaceXAI president Mike Nicolls; Wes Salandro, who previously led the project, now reports to Holland.

Global expansion: Saudi Arabia, Canada, and large-scale expansion in Memphis

To secure more space for data center construction, SpaceXAI had previously actively sought overseas expansion opportunities. A report stated that xAI had held talks with Kevin O'Leary's investment firm about jointly developing the "Wonder Valley" data center project in Alberta, Canada. The project was planned to cover 7,000 acres with a total investment of $70 billion and was called by O'Leary "the largest AI data center campus on Earth," but the two sides ultimately failed to reach a cooperation agreement.

Saudi Arabia became another key layout. xAI had held talks with local developer Ruyah AI about a potential data center site of about 1,000 acres with access to about 2 gigawatts of power. Now, SpaceXAI has reached cooperation with Saudi state-backed AI company Humain to build a 500-megawatt data center locally.

At the same time, SpaceXAI is accelerating its expansion in the United States. Musk's announced plan for 420,000 GPUs is part of the expansion of the Minihard campus in Memphis. The company originally planned to complete the launch by October 31, but adjustments to the construction plan may delay progress. The added compute will be used to handle growing external customer demand, including the $1.1 billion-per-month compute contract starting in December.

From seeking overseas land and power to accelerating the expansion of domestic data centers, SpaceXAI is shifting from "renting out idle compute" to large-scale expansion of compute supply to meet rapidly growing external customer demand.