While the world focuses on SpaceX's record-breaking $1.75 trillion valuation IPO, a subtle detail in the filing reveals a deeper signal: the bottleneck in the AI race is shifting from models to computing power. SpaceX's prospectus shows that Anthropic has committed to paying SpaceX $1.25 billion per month, equating to roughly $15 billion annually, to access GPU resources from the Colossus and Colossus II data centers. These two facilities, with a combined computing power exceeding 1 gigawatt, were originally built primarily to serve xAI and its Grok chatbot. This partnership indicates that SpaceX is executing a sophisticated capital balancing act: cash flow from Starlink is being consumed by xAI's computing power demands, while surplus GPU capacity is then leased to other AI competitors. SpaceX is effectively becoming an AI infrastructure provider. SpaceX's Computing Power Strategy: Internal Use and External Leasing The filing discloses that Anthropic will pay an undisclosed discounted fee in May and June before transitioning to the standard $1.25 billion monthly rate, which is set to continue until May 2029. As xAI's largest external customer currently, Anthropic's contract, valued at approximately $40 billion, forms the core revenue support for xAI's AI segment in 2026. Notably, an investigation by The Information highlighted a key risk: either party can terminate the agreement with 90 days' notice. This highly unusual clause means the $40 billion contract could dissolve within three months. The Colossus and Colossus II data centers, spanning Tennessee and Mississippi, are equipped with about 100,000 H100 chips and 220,000 GB200/GB300 chips, making them one of the world's largest single AI computing clusters. SpaceX initially rushed to build these facilities for its xAI division. Elon Musk later stated that SpaceX ultimately did not require the full computing capacity and revealed discussions with other companies for similar arrangements. In its S-1 filing, SpaceX stated it expects to sign more computing service contracts while continuing to use the data centers for its own operations. The company claims its computing capacity is sufficient to simultaneously meet the training and inference needs of its internal AI models and its external contractual obligations. In Q1 2026, $7.7 billion of SpaceX's $10.7 billion in capital expenditures, or 76%, was directed toward AI. This means most of the money earned by Starlink is being funneled into computing power. However, this $40 billion order demonstrates that SpaceX has reframed this model into a "dual monetization strategy," claiming it provides multiple paths for capital returns. In other words, SpaceX is transforming its AI infrastructure from an internal cost center into an asset that generates external revenue even before its IPO.
Anthropic's Computing Power Expenditure Challenge What does a $15 billion annual computing power outlay mean for Anthropic? Anthropic reported Q1 revenue of $4.8 billion, with Q2 projections exceeding $10 billion, a growth rate surpassing even Google and Facebook before their IPOs. Despite this, the $15 billion annual computing power expense represents a significant portion of its yearly revenue. Market concerns center on whether Anthropic is channeling its rapidly growing revenue into computing power without building sustainable profit margins. The simultaneous surge in revenue and computing costs presents a delicate growth paradox for contemporary AI companies. On one hand, exploding demand for applications like AI coding tools drives massive expansion in model training and inference, pushing up computing consumption. On the other hand, while Anthropic's quarterly leap from $4.8 billion to an estimated $10.9 billion is impressive, it cannot avoid the reality: the faster the growth, the greater the pressure to procure computing power. Whether this "growth for growth's sake" cycle is sustainable depends on the company's ability to find a genuine balance between computing investment and commercial returns. As computing power itself becomes a scarcer strategic resource than capital, the core competitive dimension of the future AI industry may subtly shift—it will no longer be a simple contest of model capabilities, but rather who can secure sufficient GPUs, power, and data center capacity in advance to navigate this high-consumption race more effectively.
Comments