From rejection to increased investment: How Musk turned SpaceX's AI computing power into a multi-billion dollar "cloud business" every month.

Leasing out computing power is expected to bring SpaceX AI billions of dollars in revenue per month. This business, which Musk once rejected, has now become a scarce resource that AI companies are vying to acquire.

According to a recent report by The Information, citing sources familiar with the matter, SpaceX AI has been in talks with Microsoft this summer regarding computing power leasing; meanwhile, a $1.1 billion monthly computing power lease contract will begin in December. Reuters also reported that Anthropic had previously committed nearly $45 billion to SpaceX AI's computing power, and subsequently disclosed an additional commitment of nearly $40 billion.

Musk's change of attitude initially stemmed from the low utilization rate of xAI's computing power, while external AI companies were eager to acquire computing power. Faced with idle older-generation H100 GPUs, Musk ultimately decided to lease them to companies such as Anthropic first; as external demand continued to increase, this business, which was originally considered a temporary arrangement, gradually evolved into a considerable computing power order.

As orders increase, SpaceX AI is also accelerating the expansion of its infrastructure. Musk posted on the X platform last weekend that the company plans to deploy approximately 420,000 Nvidia GPUs in November, a scale equivalent to twice that of a typical data center campus.

Musk initially refused to "sell computing power," but later changed his mind.

Leasing computing power was not Musk's original plan. Earlier this year, Musk merged xAI into SpaceX and renamed it SpaceXAI. According to sources, when xAI had idle computing power last year, executives suggested selling it to external AI companies, but Musk explicitly refused, arguing that if it became a cloud service provider, xAI's role as an AI lab might no longer be taken seriously.

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 computing power. This conveniently provided xAI with an opportunity to utilize its inventory of older Nvidia H100 GPUs—GPUs that were less efficient at training Grok models. According to a report citing an attendee, Musk stated at the time that this was only a temporary arrangement and emphasized that xAI was not entering the cloud computing business.

However, convincing Musk was only the first step. According to reports citing two sources familiar with the matter, xAI's system was not initially designed to support the secure sharing of the same computing infrastructure among different clients or projects. Therefore, when leasing idle computing power to external companies, there was a risk of mutual leakage of client data or work content. xAI has now resolved this technical issue.

Aggressive stockpiling and rapid construction: Data center built in 122 days

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

According to reports citing sources familiar with the matter, xAI stockpiled gas turbines worth hundreds of millions of dollars even before the installation locations were finalized. Musk and his team anticipated a supply shortage of turbines in the coming years; even if they bought too many, the surplus equipment could be resold at a higher price later.

Regarding site selection, xAI evaluated hundreds of potential data center locations, prioritizing areas with friendly local governments that could offer tax breaks. Late last year, Musk met two to three times a week with a group of executives and advisors specifically to discuss site selection, engine suppliers, and potential acquisitions.

In terms of construction speed, xAI set a record by putting its first data center, "Colossus" (located in Memphis and equipped with 200,000 GPUs), into operation in just 122 days, while it usually takes several years to build a data center of the same size.

To shorten the construction period, xAI introduced ABB's robots to automate the data center cabling work and adopted a modular construction approach, prefabricating some facilities at other locations and then transporting them to the site for assembly. Contractors sometimes worked 14 to 16 hours a day.

Costs of Expansion: Low vacancy rates, flooding, and management turmoil

However, aggressive expansion has also left behind many hidden dangers.

Last year, Colossus's GPU utilization rate was less than 40%, while the company was simultaneously building its second Memphis data center, Colossus II. According to a report citing a source familiar with the matter, the company's utilization rate remained low this spring, which was the direct impetus that prompted Musk to change his stance and ultimately decide to lease out computing power.

Colossus II is now facing more pressing engineering problems. According to two sources familiar with the matter, a defective drainage system has caused flooding within the facility, affecting the structural integrity of parts of the building. The necessary repairs require additional concrete pouring and structural reinforcement, potentially forcing other construction work to halt and delaying the project by at least two to three weeks. The company seconded a civil engineer from Neuralink to oversee the repairs, but that engineer left before the problem was resolved.

The management team has also experienced turmoil. Since SpaceX management took over project leadership in June of this year, the team has encountered a large-scale failure caused by construction problems and a change in management – Kevin Holland, senior director of the SpaceX Starship project, took over the team and reports directly to Mike Nicolls, president of SpaceX AI; Wes Salandro, who was previously in charge of the project, now reports to Holland.

Global Expansion: Large-Scale Expansion in Saudi Arabia, Canada, and Memphis

To secure more space for data center construction, SpaceX AI had previously been actively seeking overseas expansion opportunities. Reportedly, xAI had spoken with Kevin O'Leary's investment firm to consider jointly developing the "Wonder Valley" data center project in Alberta, Canada. The project, planned to cover 7,000 acres and with a total investment of $70 billion, was described by O'Leary as "the largest AI data center campus on Earth," but the two parties ultimately failed to reach an agreement.

Saudi Arabia has become another key location for their strategic deployment. SpaceX AI had previously negotiated with local developer Ruyah AI for a potential data center site of approximately 1,000 acres, capable of receiving about 2 gigawatts of power. Now, SpaceX AI has partnered with Humain, a Saudi state-backed AI company, to plan the construction of a 500-megawatt data center there.

Meanwhile, SpaceX AI is accelerating its expansion in the United States. The 420,000 GPUs announced by Musk are part of the expansion of the Memphis Minihard campus, which the company originally planned to complete by October 31, but adjustments to the construction plan may cause delays. The additional computing power will be used to meet increasing external customer demand, including a $1.1 billion monthly computing power contract launching in December.

From sourcing land and electricity overseas to accelerating the expansion of its domestic data centers, SpaceX AI is shifting from "renting out idle computing power" to massively expanding its computing power supply to meet the rapidly growing demand from external customers.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.