Google Owns the World’s Third-Largest Cloud. It’s Paying a Rocket Company $920 Million a Month for GPUs.
SpaceX disclosed two compute contracts — $920 million per month from Google and $1.25 billion per month from Anthropic — totaling $26 billion a year in annualized revenue, 40% more than the company earned from rockets and satellites combined in 2025. The infrastructure was originally built for xAI’s Grok, which an internal memo suggests used 11% of it.
Nine hundred and twenty million dollars a month.
That is what Google agreed to pay SpaceX — a rocket company — for access to roughly 110,000 Nvidia GPUs, according to a regulatory filing SpaceX submitted on June 5, four trading days before its expected Nasdaq debut, in a contract that runs from October 2026 through June 2029 and totals more than $30 billion if neither party exercises the termination clause that lets either side walk away with 90 days’ notice after December.
Google is the world’s third-largest cloud provider, with a $71 billion annual run rate, $460 billion in outstanding cloud contracts, and $190 billion in AI capex guidance for 2026. It just told a company whose primary business involves launching payloads into orbit that it cannot build GPU capacity fast enough to meet demand for Gemini Enterprise, calling the arrangement “a short-term, timely agreement to ensure we have bridge capacity to meet surging customer demand.”
Bridge capacity — from Elon Musk.
The Numbers Behind Elon Web Services
The Google contract is the second massive compute deal SpaceX has disclosed in its IPO filings. In May, Anthropic signed an agreement to rent the full capacity of SpaceX’s Colossus 1 data center near Memphis, Tennessee: more than 220,000 Nvidia GPUs across H100, H200, and GB200 architectures, delivering 300 megawatts of compute power, for $1.25 billion per month through 2029.
Combined, the two deals represent $26 billion in annualized revenue. To put that number in context:
| Cloud Provider | Annual Revenue / Run Rate | Source |
|---|---|---|
| AWS | $142B ARR | Q4 2025 earnings |
| Microsoft Azure | $131B ARR | Q4 2025 earnings |
| Google Cloud | $71B ARR | Q4 2025 earnings |
| SpaceX Compute | $26B (contracted) | SEC filings, June 2026 |
| Oracle Cloud | ~$24B ARR | FY2025 earnings |
SpaceX’s contracted compute revenue now exceeds Oracle Cloud’s annual run rate. Eighteen months ago, it sold zero GPU-hours. Now it has leapfrogged into the top tier of cloud infrastructure by contract value, behind only AWS, Microsoft, and Google itself — the same Google that just signed a $30 billion check to rent capacity from it. SpaceX’s entire 2025 revenue across all business lines — Starlink, launch services, government contracts, and xAI — was $18.7 billion. The compute contracts alone are 40% larger than that.
How a Chatbot Built a Cloud
The origin of SpaceX’s cloud business is not a strategic pivot. It is a salvage operation.
In February 2026, SpaceX merged with xAI, Musk’s artificial intelligence venture. xAI had spent 2024 and 2025 assembling what it claimed was the world’s largest GPU cluster at Colossus 1 in Memphis, at a cost the S-1 reveals as $12.7 billion in AI-directed capital expenditure in 2025 alone, with another $7.7 billion in Q1 2026.
Then came the utilization problem, and it was devastating. According to reporting that cites a reported internal xAI memo, Colossus 1 was running at approximately 11% Model FLOPs Utilization. That is the standard measure of how much of a GPU cluster’s theoretical computing power is actually doing useful training work; production-grade industry benchmarks sit between 35% and 45%. The culprit was a heterogeneous mix of H100, H200, and GB200 GPUs that could not efficiently parallelize Grok’s training workloads across the cluster. xAI moved its model training to the newer Colossus 2 facility, leaving 220,000 GPUs and 300 megawatts of power infrastructure sitting largely idle — not worthless, but deeply underutilized.
SpaceX’s response was to rent it out. Anthropic, which had been capacity-constrained and struggling to scale Claude, took the entire Colossus 1 facility and raised its API usage limits the same day the deal was announced. Google’s deal for a separate pool of 110,000 GPUs followed within weeks. Neither customer is running workloads on a purpose-built, reliability-engineered cloud platform with the service level guarantees that AWS or Azure provide. They are renting raw GPU capacity from a data center that was built for a chatbot and couldn’t train it effectively.
The Per-GPU Math
The per-GPU pricing tells its own story when you divide the monthly fees by the disclosed GPU counts:
| Deal | Monthly Fee | GPUs | Per-GPU/Month |
|---|---|---|---|
| Anthropic (Colossus 1) | $1.25B | 220,000 | $5,682 |
| $920M | 110,000 | $8,364 | |
| AWS H100 on-demand (benchmark) | — | — | ~$8,850 |
| AWS H100 1yr reserved (benchmark) | — | — | ~$3,500–5,300 |
Google is paying near on-demand cloud rates. Anthropic is paying closer to reserved-instance pricing — a discount that is reasonable for committing to an entire facility, mixed architecture and all. (AWS benchmark rates are derived from published p5.48xlarge on-demand pricing as of June 2026.) The difference matters: Colossus 1’s heterogeneous GPU mix depresses training utilization for a single large model but may impose less penalty on diverse inference and fine-tuning workloads. Neither price is a bargain. Both reflect the market clearing rate when demand for GPU capacity outstrips what a company spending $190 billion per year on AI infrastructure can build by itself.
What Google’s Desperation Reveals
The most significant detail in the filing is not the dollar amount. It is that Google needs it at all.
Google Cloud grew 50% year-over-year in Q4 2025, the fastest clip among the three hyperscalers, while its parent guided $190 billion in AI capex for 2026 and had already burned through $75 billion in the first quarter. It owns proprietary TPU hardware, designs custom chips, and operates one of the three largest computing networks ever built, backed by $460 billion in signed cloud contracts signaling that enterprise demand is not the constraint. Physical infrastructure is. For Google, paying SpaceX $920 million per month is apparently cheaper than telling customers to wait.
The IPO Calculus
SpaceX is pricing its Nasdaq debut at $135 per share, targeting a $75 billion raise at an implied $1.77 trillion valuation, with trading expected around June 12.
The compute deals are central to the narrative. SpaceX’s S-1 reveals three segments with radically different economics: Starlink ($11.4 billion in 2025 revenue, $4.4 billion operating profit, 63% EBITDA margin), Space ($4.1 billion revenue, $657 million loss), and AI ($3.2 billion revenue, $6.4 billion loss — a 2:1 loss-to-revenue ratio that in any other era would have killed the IPO). Without the compute contracts, the AI segment is a black hole funded by Starlink profits. With them, the story transforms into capital-light infrastructure arbitrage with two blue-chip anchor tenants — exactly the narrative IPO investors want to hear, even when the exit provisions suggest the revenue could vanish with 90 days’ notice.
Those exit provisions matter. After December 31, 2026, either party in the Google deal can walk. Google is spending $190 billion per year building its own clusters. When those come online, the monthly checks stop.
Environmental Overhead
The infrastructure supporting these deals carries liabilities that do not appear in the revenue line. SpaceX’s Colossus complex was brought online fast, which in Memphis meant trucking in natural gas turbines and bypassing standard air quality permitting entirely. In April 2026, the NAACP filed suit alleging Clean Air Act violations and claiming the turbines exposed predominantly Black neighborhoods to harmful pollutants, while SpaceX amended its S-1 to warn investors that access to water is “increasingly constraining” growth plans. The world’s newest cloud provider was also, briefly, the world’s least regulated power plant operator, and speed has costs that the revenue line does not capture.
The Strongest Case For SpaceX
The counterargument deserves its full weight: maybe this was the plan all along. Build massive GPU clusters, train frontier models in intense bursts, then rent the idle capacity at cloud-market rates while training migrates to newer, purpose-built hardware. If SpaceX earns $26 billion per year from rentals while xAI trains Grok on Colossus 2, the combined value proposition — frontier AI research subsidized by infrastructure arbitrage — is stronger than any pure-play cloud provider or any pure-play AI lab can match independently. The 11% utilization figure, if accurate, may reflect the natural lifecycle of a training cluster rather than a design failure, and the speed at which SpaceX converted idle capacity into revenue suggests this was at minimum a contingency they had thought through. Whether it was genius or improvisation depends on information SpaceX has not disclosed.
Limitations
The 11% Model FLOPs Utilization figure originates from social media posts citing an internal xAI memo; SpaceX has not confirmed it in regulatory filings. We do not know the exact GPU model breakdown for the Google deal or whether it draws from the same Memphis infrastructure as Anthropic. Both contracts include early termination provisions, so total value assumes full execution through 2029. Per-GPU pricing comparisons use published AWS rates as of June 2026; actual negotiated enterprise rates at this scale are not public.
The Bottom Line
The AI compute supply crisis has reached the point where Google — a company that builds cloud infrastructure as a core business — is renting GPUs from a company whose primary expertise is launching payloads into orbit. SpaceX poured $12.7 billion into GPU infrastructure for a chatbot that achieved 11% utilization, then converted the stranded asset into $26 billion per year of recurring revenue from two of the most sophisticated compute buyers on Earth. That is either a masterclass in infrastructure arbitrage or a monument to an industry that cannot build fast enough. Probably both.
If you are an investor evaluating the SpaceX IPO, know that the compute revenue is real but comes with 90-day termination clauses and a customer spending $190 billion per year to replace you. If you run infrastructure procurement, the world’s largest cloud builders cannot keep pace, and the gap is wide enough for a rocket company to park a data center in it. And if you need large-scale GPU access for a startup, the window is now — in eighteen months, Google’s own clusters will be online, and the price of compute will have either collapsed or gone somewhere nobody has a model for.