SAN FRANCISCO — Google has spent years insisting it would never be caught short on computing power. It built its own custom chips, the Tensor Processing Units, specifically to sidestep dependence on Nvidia’s GPUs. It has committed to more than $180 billion in capital expenditure this year. By some estimates, it is the world’s largest single owner of AI compute. And yet, on June 5, the company agreed to pay a rocket builder nearly $1 billion a month for GPUs it apparently cannot get anywhere else fast enough.
The deal, disclosed in a regulatory filing tied to SpaceX’s historic initial public offering, commits Google to pay $920 million per month from October 2026 through June 2029 for access to approximately 110,000 Nvidia GPUs, plus CPUs, memory, and related components housed at data centers tied to xAI, now a division of SpaceX following their February merger. Over its full term, the contract is worth roughly $30 billion, making it one of the largest cloud-compute commitments ever disclosed publicly.
A Google spokesperson told CNBC the arrangement was driven by demand that caught the company off guard. Google Cloud and SpaceX are long-time partners, the spokesperson said, and the agreement is a short-term, timely arrangement to ensure bridge capacity to meet surging customer demand for its agent platform, Gemini Enterprise, which had been higher than expected. The framing was careful: a temporary bridge, not a structural dependency. Whether investors read it that way is another matter entirely.
The more revealing question is not why Google signed the deal but why SpaceX had the capacity to offer it. The answer traces back to a design problem Elon Musk’s AI operation has not publicly acknowledged. According to internal documentation that circulated in May, xAI’s Colossus 1 facility in Memphis was running at roughly 11% Model FLOPs Utilization, a fraction of the 35 to 45 percent industry standard for production-grade AI work. The reason: the data center was built with an eclectic mix of H100, H200, and GB200 Nvidia GPUs, and the heterogeneous architecture proved incapable of parallelizing the training workloads Grok required. xAI quietly shifted its training operations to the newer Colossus 2 facility, leaving Colossus 1’s 220,000 GPUs largely idle.
SpaceX, now responsible for xAI’s finances and its IPO prospectus, had an obvious interest in turning that idle capacity into revenue. First came Anthropic, which agreed in late May to pay $1.25 billion per month through 2029 for the full output of Colossus 1, on the same day Anthropic raised usage limits for Claude, implying it had been compute-constrained before the deal closed. Google’s agreement, disclosed the following week, covers roughly half that compute at a proportionally lower price. SpaceX did not specify which data center Google would use, and Musk has indicated Colossus 2 is reserved for xAI’s own work. The Anthropic fundraise that preceded the compute deal, which valued the company at $965 billion, provides context for how expensive the AI infrastructure race has become.
What neither company has explained is why a facility that xAI itself found too architecturally fragmented to train Grok on has suddenly become indispensable to two of the most sophisticated AI operations in the world. The answer, industry analysts suggest, is simpler than it sounds: when GPU supply is this constrained, imperfect hardware at scale beats perfect hardware that does not exist yet. Both Google and Anthropic have enormous inference workloads, running models in production, processing queries, generating responses, that are less sensitive to architectural coherence than training is. An inference cluster does not need to parallelize across thousands of chips in perfect synchrony. It just needs chips.
The timing of the disclosure was not accidental. SpaceX announced the Google deal less than a week before its stock is set to begin trading on the Nasdaq exchange. The company is seeking to raise around $75 billion at a valuation of approximately $1.75 trillion, which, if achieved, would be the largest IPO in history. The compute contracts with Google and Anthropic together represent more than $2 billion in monthly committed revenue, a figure that gives SpaceX’s AI infrastructure story the recurring-revenue texture that public market investors have learned to prize. Whether that revenue is as durable as it appears is less clear. Both contracts include 90-day termination clauses exercisable by either party after December 31, 2026. A subscription is not the same thing as a purchase order.
SpaceX’s own IPO filing is candid about the competition the company faces. In a section on compute service agreements, it names Google as a competitor in connectivity, in AI, in cloud services. The company is now simultaneously Google’s landlord and its rival, a structural ambiguity that would ordinarily draw regulatory scrutiny. Alphabet has been a SpaceX investor since 2015, when it participated in a funding round that valued the rocket company at $12 billion. That stake is now worth many multiples of what it paid, and the two companies are reported to be exploring orbital data centers together. Whether those intertwined financial relationships complicate any future antitrust review of the compute market is a question that regulators have not yet been asked to answer. The SpaceX IPO roadshow, led by Jamie Dimon at $135 a share, opens that question for public investors for the first time.
For Google, the deal sits uneasily alongside its stated infrastructure ambitions. The company’s Cloud backlog exceeded $460 billion at its last earnings report, and it has publicly committed to building enough capacity to meet that demand through its own facilities. The Gemini Enterprise explanation, surging agent demand, temporary bridge, is plausible but incomplete. Google has committed to more than $180 billion in capital expenditure this year, a figure that was supposed to reflect exactly this kind of demand anticipation. That it still found itself short enough to sign a near-$1 billion monthly contract with a competitor raises a question the company’s data center teams will need to answer: was the demand genuinely unforeseeable, or was the build-out simply too slow?
The agreement ramps up through September at a reduced fee. If SpaceX fails to deliver the committed GPU capacity by September 30, Google may terminate immediately after a one-month grace period, or accept whatever hardware is available at proportionally reduced rates. That provision gives Google a meaningful exit if the facility underperforms, but it also implicitly acknowledges that the question of whether Colossus 1’s architectural complexity will complicate inference workloads the way it complicated training has not been fully resolved. For now, both sides appear to have decided that the risk of running short on compute is larger than the risk of running on imperfect hardware. In the AI infrastructure race of 2026, that is likely the correct calculation. Whether it remains so depends on how quickly Nvidia, and the data centers it supplies, can catch up with the demand that no one, apparently, forecast correctly.
The full picture of where AI compute scarcity leads remains unwritten. The trillion-dollar AI spending frenzy has produced data centers faster than any prior infrastructure cycle in history, and still not fast enough.
