Anthropic has committed to spending $1.25 billion per month through May 2029 to rent Colossus 1, xAI's supercomputer in Memphis, Tennessee, a deal that exposes the staggering compute costs driving the AI industry and raises questions about whether current business models can sustain this spending. The contract, totaling over $40 billion, gives Anthropic exclusive access to roughly 220,000 Nvidia GPUs (H100, H200, and GB200 processors) and 300 megawatts of power, according to filings disclosed in SpaceX's confidential S-1 document obtained by Business Insider on May 20, 2026.

Why Is Anthropic Spending Half Its Annual Revenue on a Single Compute Contract?

The scale of this commitment is staggering when placed in context. Anthropic's annualized revenue in April 2026 exceeded $30 billion, meaning this single infrastructure contract consumes roughly half of the company's annual earnings. To put this in perspective, OpenAI's five-year deal with CoreWeave totals $11.9 billion, or about $2.38 billion per year, making Anthropic's annual commitment roughly 6.3 times larger than OpenAI's public compute agreement. The contract reflects a fundamental shift in how AI companies view infrastructure: not as a variable cost to be minimized, but as a strategic asset to be controlled and scaled aggressively.

Anthropic is using Colossus primarily for Claude inference, the computational work required to answer user queries rather than train new models. This distinction matters because inference workloads are memory-bandwidth-bound, meaning they depend more on how quickly data moves through the system than on raw processing power. The company has already announced immediate benefits: doubled bandwidth limits for Claude Pro, Max, Team, and Enterprise users over five-hour windows, and removal of peak-hour caps for Pro and Max subscribers.

What Does SpaceX's Financial Disclosure Reveal About xAI's Viability?

SpaceX's S-1 filing, which the company confidentially submitted to the SEC on April 1, 2026, provides the first formal accounting of xAI's financial performance. The numbers are sobering. In 2025, xAI posted an operating loss of $6.4 billion against revenue of just $3.2 billion. In the first quarter of 2026 alone, the AI segment recorded an operational loss of approximately $2.5 billion on segment revenue of about $800 million. These losses dwarf what most analysts had previously estimated, suggesting the true cost of building and operating cutting-edge AI infrastructure has been significantly underestimated.

Capital expenditure tells an even more dramatic story. SpaceX allocated between $12.7 billion and $20 billion to AI infrastructure in 2025, with Q1 2026 alone consuming $7.7 billion, implying an annualized burn rate of $30.8 billion. This aggressive spending reflects Elon Musk's stated belief that owning physical compute assets will outcompete cloud rentals over the long term. However, the depreciation expenses from this hardware will drag on earnings for years, even after the equipment becomes obsolete.

How Does Colossus Compare to Other AI Infrastructure, and Why Is It Underutilized?

Colossus 1 represents one of the largest GPU clusters ever assembled, yet internal xAI documents reveal a troubling inefficiency problem. An internal memo signed by xAI president Michael Nicolls and obtained by Business Insider describes a Model FLOPs Utilization (MFU) rate of just 11 percent on the roughly 550,000 Nvidia GPUs operated by xAI, a figure the memo characterizes as "embarrassingly low". For context, production-grade large language model training typically achieves between 35 and 45 percent MFU. Meta reached 43 percent on Llama 3 training, Google achieved 46 percent on PaLM, and ByteDance posted 55.2 percent MFU on 12,288 GPUs during training of a 175 billion parameter model.

This utilization gap suggests xAI has struggled to keep its massive compute cluster fully occupied with productive work. The Anthropic contract, in effect, monetizes idle capacity. By leasing Colossus to a competitor, SpaceX transforms underutilized infrastructure into a revenue stream, though the pricing reflects strategic underpricing rather than market equilibrium. The implicit cost per GPU per hour works out to approximately $7.78, compared to $3 to $4 per hour for on-demand H100 rentals at AWS and Google Cloud, or as low as $1.49 to $2.99 on newer cloud providers like Lambda Labs and RunPod.

Steps to Understand the Broader Implications of This Deal

• Monitor SpaceX's IPO Timeline: SpaceX is targeting an IPO valuation between $1.75 trillion and $2 trillion, with a roadshow planned for the week of June 8, 2026. The Anthropic contract provides predictable revenue that softens the narrative around xAI losses, making it a critical element of the company's IPO story.
• Track Grok's Product-Market Fit: xAI's losses will only become sustainable if Grok, the AI model trained on Colossus, develops meaningful commercial traction. Currently, Grok remains primarily accessible through the X platform, and no concrete guidance on profitability timelines has been provided.
• Watch for Orbital AI Infrastructure Announcements: Anthropic and SpaceX have expressed interest in developing several gigawatts of AI computing capacity in orbit, but no contract has been signed as of May 21, 2026. This remains a declarative goal rather than a funded project, with significant technical obstacles still unresolved.
• Assess Competitive Pressure on Pricing: If other AI companies follow Anthropic's lead and sign long-term capacity agreements, the market for GPU compute may bifurcate into long-term contracts at premium rates and spot-market pricing that continues to decline.

The Anthropic-SpaceX deal also raises a contractual question that remains unresolved. Elon Musk stated on X that he reserves the right to reclaim computing capacity if Anthropic's AI "engages in actions harmful to humanity," though the exact contractual status of this clause has not been jointly confirmed by both parties. This ambiguity introduces legal and operational risk into what is otherwise a straightforward infrastructure lease.

Elon Musk

What Does This Mean for the Future of AI Infrastructure Spending?

The Anthropic contract signals that AI companies have moved beyond cost optimization and into a phase of strategic infrastructure control. Rather than renting compute on an hourly basis, leading AI labs are now signing multiyear capacity commitments that guarantee access to cutting-edge hardware regardless of utilization rates. This shift reflects both the scarcity of advanced GPU supply and the belief that controlling infrastructure is essential to maintaining competitive advantage.

However, the underlying economics remain precarious. Even with Anthropic's $15 billion annual contribution, xAI's segment still faces negative operating leverage and sustained losses. SpaceX's S-1 offers no concrete guidance on when xAI will reach profitability, only a promise to "optimize GPUs" and streamline Grok models. Depreciation from rapid compute spending will continue to balloon cost of goods for years, and the company has not demonstrated meaningful pricing power in the market.

For investors and industry observers, the Anthropic deal provides a rare window into the true cost of frontier AI development. It confirms that building and operating world-class AI infrastructure requires tens of billions of dollars annually, that utilization rates remain far below theoretical maximums, and that even companies with access to massive capital and vertical integration advantages struggle to achieve profitability. As SpaceX prepares for its IPO roadshow, this contract will likely feature prominently in the company's narrative about AI's future, even as the underlying financial reality suggests the path to sustainable returns remains uncertain.

" }