$920 million per month. That is what Google agreed to pay SpaceX on June 5, 2026, for access to 110,000 NVIDIA GPUs housed in the same Memphis data centers that Elon Musk built to train Grok, his direct rival to Google’s own Gemini.
The contract runs from October 2026 through June 2029. At full rate, the total value reaches approximately $30 billion. Combined with Anthropic’s $1.25 billion monthly deal signed weeks earlier, SpaceX now collects $2.17 billion per month renting AI compute to two of its biggest competitors.
Five years ago, Google was the cloud provider and SpaceX was the customer. The roles have completely reversed.
Colossus Was Built for Grok. Grok Couldn’t Fill It.
SpaceX’s Colossus complex in Memphis, Tennessee, represents $18 billion in hardware: 555,000 NVIDIA GPUs across H100, H200, and GB200 generations, drawing 2 gigawatts of power. It is the largest single-site AI training installation on the planet.
The facility exists because Musk wanted to train Grok after SpaceX acquired xAI in February 2026. But the execution stumbled. xAI’s model utilization hit roughly 11%, far below the 40% that rival labs achieve on their own infrastructure. Colossus 1’s mixed GPU architecture (combining H100s and H200s in the same clusters) made efficient large-scale training difficult, and xAI eventually moved its primary workloads to Colossus 2.
That left Colossus 1 sitting mostly idle with 222,000 GPUs consuming power and generating no revenue. Within 90 days, SpaceX signed contracts that converted the entire underutilized complex into the AI industry’s most expensive rental property.
Two Deals, $26 Billion a Year, Zero Days of Vacancy
In May, Anthropic locked down every megawatt of Colossus 1’s capacity: 222,000 GPUs, 300+ megawatts of power, $1.25 billion per month through May 2029. At full run rate, Anthropic’s payments alone ($15 billion per year) represent 80% of what SpaceX earned across all its businesses in the prior fiscal year.
One month later, Google signed for 110,000 GPUs at Colossus 2 at $920 million per month through June 2029. A Google Cloud spokesperson told TechCrunch the deal was “a short-term, timely agreement to ensure we have bridge capacity to meet surging customer demand for our agent platform, Gemini Enterprise, which has been even higher than we expected.”
Between the two contracts, SpaceX now generates approximately $26 billion in annualized recurring revenue from AI compute rentals. To put that in perspective: NVIDIA, the company that manufactured every GPU in those data centers, reported $44.1 billion in quarterly revenue in its most recent earnings. SpaceX is generating NVIDIA-scale revenue simply by renting NVIDIA’s chips to companies that could not build data centers fast enough.
“Bridge Capacity” and What It Reveals About AI Demand
Google’s characterization of the deal as bridge capacity deserves scrutiny. This is Alphabet, a company that committed more than $180 billion in capital expenditure for 2026, most of it allocated to AI data centers and model training. Google designs its own TPU chips. It operates one of the three largest cloud platforms on Earth. And it still cannot build capacity fast enough to serve enterprise demand for Gemini.
The bottleneck is not money. Alphabet has more cash on hand than most countries’ GDP. The bottleneck is time. New data centers take 18 to 24 months from groundbreaking to GPU installation. Chip foundries operate on 12 to 18 month lead times. Even with $180 billion in annual capex, the physical construction of AI infrastructure cannot keep pace with the rate at which enterprise customers are signing Gemini Enterprise contracts.
Anthropic tells a similar story. It has a $25 billion relationship with Amazon Web Services. It operates primarily on AWS infrastructure. Its revenue recently surpassed OpenAI’s. And it still needs to rent 222,000 GPUs from a rocket company’s data center in Memphis to keep up with demand for Claude.
These are not startups scrambling for server time. These are trillion-dollar enterprises with dedicated chip programs and multi-year datacenter construction pipelines, and they are both renting GPUs from a facility that was underutilized six months ago.
The signal is clear: enterprise AI demand has crossed from “compute is expensive” to “compute is scarce regardless of price.” When two of the three largest AI labs outside of OpenAI need to lease capacity from a competitor’s idle hardware, the supply deficit is structural, not temporary. It also explains why SoftBank is building an $87 billion data center campus in France and why NextEra acquired Dominion for $67 billion to lock up the power grid capacity that AI workloads require.
Both contracts include flexibility. Either party can terminate with 90 days’ notice after December 31, 2026. If SpaceX cannot deliver the committed GPU capacity by September 30, Google gets a one-month grace period followed by the option to terminate or accept reduced capacity at a lower monthly rate. The ramp-up period through September carries reduced fees.
That flexibility lets Google frame the arrangement as temporary. Whether it exercises the termination clause is another question entirely.
The IPO Prices Today. The Revenue Appeared Last Month.
SpaceX prices its IPO on Thursday, June 11, at $135 per share. Trading opens on Nasdaq under the ticker SPCX on Friday. The company targets $75 billion in proceeds at a valuation of $1.75 trillion, which would make it the largest IPO in recorded history.
The timing of these compute contracts is not subtle. Epistrophy Capital Research chief market strategist Cory Johnson told Yahoo Finance: “I can safely say SpaceX is the first company to ever add $26 billion in ARR between the date of its IPO filing with the SEC and the first trade.”
Goldman Sachs, which is marketing the IPO to institutional investors, now has a revenue story that did not exist 60 days ago. SpaceX went from training Grok at 11% utilization to collecting $26 billion a year from the two largest non-OpenAI AI labs. Morningstar, applying more conservative assumptions, contends that SpaceX is worth about half its implied market cap.
The 90-day cancellation clauses sit at the center of this debate. If both Google and Anthropic exercise their termination options after December 31, SpaceX loses $26 billion in ARR just four months after its first earnings report as a public company. If they stay, the revenue justifies the valuation. Investors are pricing a bet on which outcome is more likely.
Capex That Funds Itself on Paper
There is a circularity embedded in this transaction that few headlines mention. Alphabet owns approximately 6% of SpaceX, a stake worth roughly $100 billion at the $1.75 trillion IPO valuation. Google also invested $40 billion in Anthropic, which is itself paying SpaceX $15 billion per year.
Every dollar Google pays SpaceX in compute fees flows into SpaceX’s revenue, which supports the valuation, which inflates the value of Alphabet’s own equity position. Every dollar Anthropic pays SpaceX does the same, and Alphabet profits from Anthropic’s growth as both an investor and a cloud competitor.
This is not fraud or even unusual by 2026 AI industry standards. It is the same circular capital architecture that defines the entire AI investment landscape: invest in everything, rent from everyone, monetize the portfolio appreciation as much as the products. The AI industry’s largest companies are simultaneously competitors, customers, investors, and landlords to one another.
From Rocket Company to AI Landlord
SpaceX’s pivot tells a broader story about where value accrues in the AI stack. The company spent $18 billion building Colossus. Its own AI model could not utilize the capacity. Within one quarter, it converted idle infrastructure into $26 billion in annual revenue by renting to the very labs it was supposed to compete against.
NVIDIA still sits at the top: every GPU in those contracts was purchased from Jensen Huang’s company. But SpaceX has inserted itself into a layer between chip manufacturing and model training that did not exist a year ago. It is not a cloud provider in the traditional sense. It does not offer a managed platform, an API, or developer tooling. It rents raw compute capacity at scale, and it turns out that is the one thing the AI industry cannot get enough of.
For enterprise AI buyers, the deal also reshapes the competitive map. Companies evaluating Gemini Enterprise, Claude, or GPT-5.5 should understand that two of those three products now run, at least partially, on infrastructure controlled by the company that builds Grok. Musk has publicly attacked both Anthropic and Google; his company now has contractual visibility into their compute consumption patterns. The confidentiality provisions in these contracts will get more attention as the competitive dynamics sharpen.
SpaceX also disclosed in its IPO filing that it plans to deploy “as many as one million orbital data center spacecraft” using Starship and Starlink infrastructure. That vision, building AI data centers in orbit to avoid terrestrial power constraints, sounds like science fiction until you realize the same company just closed $26 billion a year in terrestrial compute contracts that its own AI model could not justify. If SpaceX can monetize idle capacity this effectively on the ground, the business case for orbital compute changes from speculative to inevitable.
The $1.75 trillion question is whether SpaceX’s position is durable or temporary. If Google and Anthropic are genuinely capacity-constrained and these contracts represent real demand, SpaceX has stumbled into one of the most profitable business pivots in corporate history. If the contracts are primarily structured to support an IPO narrative, the 90-day termination clauses will do the talking by early 2027.
Either way, the image is hard to forget: Google, the company that invented the Transformer architecture, paying Elon Musk’s rocket company for the GPUs to run its own AI models. The AI compute market in 2026 is stranger than anyone predicted.