Author: Dong Jing; Source: Wall Street News

SpaceX is transforming its vast computing infrastructure into a high-speed cash machine. On the eve of its upcoming launch on the U.S. stock market, SpaceX, owned by Musk, has signed huge computing power leasing agreements with Anthropic and Google. The two contracts have a total value of more than 70 billion US dollars and an annualized revenue of 26 billion US dollars, injecting strong fuel into its IPO narrative.

On June 6, according to Reuters, SpaceX announced on June 5 that it had signed a multi-year cloud services agreement with Alphabet’s Google. According to regulatory documents, Google will pay SpaceX US$920 million per month from October this year to June 2029 to obtain approximately 110,000 Nvidia GPUs and related computing resources; during the ramp-up phase before that, Google will pay at a lower rate.

Previously, Anthropic announced in May that it would lease most of the computing power at SpaceX’s Colossus 1 data center in Memphis, Tennessee, for US$1.25 billion per month, with an initial commitment period of six months. When the two agreements are combined, SpaceX can earn approximately US$2.17 billion in monthly revenue from the computing power leasing business, with an annualized scale of approximately US$26 billion. If both contracts are executed until their expiration date, the total contract value will exceed US$70 billion.

The above transactions directly strengthen SpaceX’s AI growth story for its IPO. SpaceX plans to complete its listing next week, with a target financing scale of US$75 billion, which is expected to be the largest IPO in the history of the US stock market. The explosive growth of the computing power rental business has provided it with quantifiable and high-visibility income, but at the same time, external doubts about SpaceX's AI laboratory xAI's own research and development capabilities are also quietly rising.

Google’s big order: US$920 million/month, locked until 2029

According to documents submitted by SpaceX to regulatory agencies, the Google Cloud Service Agreement was officially signed on June 5, 2026. The agreement stipulates that Google will use computing power at a lower rate before September this year and enter the ramp-up stage; from October to June 2029, Google will pay a fixed monthly payment of US$920 million.

The computing resources provided include approximately 110,000 NVIDIA GPUs, CPUs, memories and other related components.

The terms of the agreement also have protective clauses: If SpaceX fails to provide the agreed number of GPUs before September 30, after the expiration of the one-month grace period, Google can immediately terminate the agreement, or choose to accept the actual number of GPUs provided and reduce the monthly fee in proportion.

In addition, after December 31, 2026, either party may terminate the agreement by giving 90 days' notice to the other party. Google will retain ownership and all intellectual property rights to its content, AI models and related data.

Analysis said that this means that the total contract value of US$70 billion disclosed by SpaceX depends to a certain extent on whether both parties choose to fulfill the agreement until the expiration date.

The computing power agreement between Anthropic and SpaceX was announced in May this year. Anthropic will rent most of the computing power of SpaceX’s Colossus 1 facility at a price of US$1.25 billion per month.

The facility, located in Memphis, Tennessee, has more than 220,000 Nvidia processors and will provide 300 megawatts of additional computing power for the Claude chatbot and will be delivered within a month.

There is a hidden secret behind Anthropic’s big deal

According to technology media The Information, there is a little-known competitive dispute behind this transaction. xAI has long tracked and tried to use Anthropic's technology to develop its own products - including continuing to use its model output through various channels after Anthropic cut off access rights.

According to reports, six people familiar with the matter revealed that xAI had carried out a months-long distillation project and directly used Claude’s response results to train its own programming model. In January this year, xAI co-founder Tony Wu announced to employees that Anthropic had cut off access rights, but some xAI engineers continued to access Claude through personal accounts until Anthropic closed these accounts one after another. Two people familiar with the matter revealed that xAI has also used the intermediary Blackbox AI to access the Anthropic model, and the most recent usage record can be traced back to mid-May this year.

Musk himself publicly admitted in a court testimony in May that xAI had "partially" used the OpenAI model to train Grok, and said that such practices are common in the industry. But neither he nor the company disclosed the specific extent of xAI's use of Anthropic technology or the various workarounds employees adopted to gain direct access.

xAI hardware planning error gave rise to SpaceX computing power leasing business

The Information’s report revealed that xAI is facing a series of severe challenges at the AI R&D level, which is in tension with the AI growth story SpaceX told Wall Street.

At the personnel level, xAI has experienced large-scale turmoil: a large number of core personnel, including eight co-founders, have resigned one after another, and the engineering team has been laid off.

According to reports, a person familiar with the matter revealed that as of mid-May this year, the xAI pre-training team had fewer than five people, compared with more than 20 people a year ago. The head of the Grok Code project has changed at least four times in the past few months, with the most recent one, Beibin Li, leaving in May. In addition, an employee accidentally deleted the core training data of the AI ​​programming product during the data migration process, resulting in a loss equivalent to two to three weeks of work results.

At the strategic level, Musk stated at an internal meeting earlier this year that he plans to introduce Cursor and Mistral AI to fill key technology gaps - Mistral is responsible for pre-training, and Cursor is responsible for post-training. However, early contacts with Mistral did not translate into formal cooperation: Mistral executives visited xAI’s headquarters in mid-April, but high-level talks between the two parties were interrupted after a few weeks; Mistral co-founder engineer Devendra Chaplot left in May only about a month after joining xAI.

The cooperation with Cursor is deepening. In April this year, SpaceX announced that it would open its infrastructure to Cursor to train its own model Composer, and it had the option to acquire Cursor for $60 billion later this year. If it gives up, it will have to pay a breakup fee of $10 billion. Cursor CEO Michael Truell has become a regular visitor to the xAI offices.

According to The Information, a series of setbacks experienced by xAI recently have raised questions about whether SpaceX will focus on independent AI research and development or computing power leasing business. However, Anthropic's payment to SpaceX has, to a certain extent, made up for the financial gap in xAI's deeply loss-making unit, which led to a large loss for the overall company in the first quarter.

However, SpaceX’s leasing of computing power reflects xAI’s own technical difficulties. According to reports, sources revealed that xAI encountered serious obstacles when training the Grok model in the Colossus 1 data center - the facility mixedly deployed GPUs of various architectures such as H100, H200 and GB200. The heterogeneous environment made it difficult to effectively promote the training work, and xAI had to migrate the model training tasks to the Colossus 2 data center.

It was this mistake in design planning that provided an opportunity for SpaceX’s commercial layout. SpaceX is trying to turn xAI’s legacy computing power assets into a source of revenue. It has signed cloud service agreements with a number of AI-related companies to monetize Colossus 1’s idle computing power.

However, SpaceX’s upcoming IPO will focus on computing power rental revenue as a core growth highlight, but the rapid expansion of this business model has also raised questions about its AI strategic positioning.

Musk has repeatedly characterized the Anthropic protocol as a "short-term arrangement" on the X platform and said, "I think we may need to get the computing power back at some point." SpaceX also stated in the IPO document that the company is continuing to train the next generation of major Grok models, but did not disclose any release schedule.