Renting From the Enemy

The most revealing numbers in SpaceX's IPO paperwork weren't about rockets. They were about rent.

Buried in the regulatory filings that accompanied SpaceX's public debut are two leases that quietly redraw the map of frontier compute. Anthropic has taken exclusive capacity at the Colossus 1 data center for inference, agreeing to pay roughly $1.25 billion per month for three years. And Google has agreed to lease about 110,000 GPUs and related infrastructure at Colossus 2 for approximately $920 million per month — around $30 billion in total, running from October 2026 through June 2029.

Add them up and two of the most formidable AI operations on earth are committing something in the neighborhood of $2.2 billion a month to a data center owned by xAI, the maker of Grok and now a subsidiary of SpaceX. Musk built Colossus to train models that compete with Gemini and Claude. He is now collecting rent from both.

The landlord economics

The Colossus campus is the largest single-site AI installation in the world: roughly 2 gigawatts of capacity and about 555,000 NVIDIA GPUs, bought for approximately $18 billion. That is an enormous fixed asset, and fixed assets want to be filled. The IPO disclosures suggest xAI has worked out that the fastest way to monetize a hyperscale build is not only to run your own training jobs on it — it's to lease the silicon to whoever will pay.

The timing matters because xAI is bleeding. The company lost $2.4 billion in the quarter ending March 2026, up from $936 million a year earlier. Against that, an annualized lease book approaching $26 billion from Google and Anthropic alone is not a rounding error — it is the difference between a cash furnace and a business. The merger that folded xAI into SpaceX valued the combined entity at $1.25 trillion, and the leases help explain how that number was defended: Colossus is no longer just a cost center for training runs. It is a compute REIT with a rocket company attached.

Musk built the world's biggest GPU cluster to beat Google and Anthropic. Now they're his two biggest tenants — and the rent may be more reliable than the models.

Note the division of labor. Anthropic's Colossus 1 capacity is earmarked for inference — serving paying customers, not training the next model. Google's Colossus 2 lease is for a fresh block of accelerators. Neither company is short on ambition or capital; Google builds its own TPUs, and Anthropic has its pick of clouds. That they are still writing nine- and ten-figure monthly checks to a rival tells you how tight the market for ready, powered, networked GPUs has become.

Everyone is a tenant now

The scramble doesn't stop at Colossus. According to CNBC, Anthropic is in preliminary talks to also lease computing power from Meta, in a deal reportedly worth about $10 billion — this on top of its xAI arrangement. Meta, for its part, has reportedly been weighing entry into the cloud-computing business outright; Mark Zuckerberg said in May the company was considering it, and CNBC reports Meta could spend as much as $145 billion on capital expenditure in 2026. Zuckerberg has said before that companies regularly ask to buy compute from Meta at a premium to what Meta paid. All of these figures are unverified and should be read as reported intent, not signed terms — but the direction is unmistakable: the labs are becoming each other's customers.

The pricing environment corroborates the squeeze. On Polymarket, bettors put the odds that B200 rental prices rose in July at 90%, with H200 at 86% (thin volume, point-in-time sentiment). When rental rates climb and the newest silicon is still scarce, owning 555,000 installed GPUs starts to look less like a liability and more like beachfront property.

There is a strategic risk here that the filings don't spell out. A frontier lab that rents its inference capacity from a competitor is, in some sense, handing that competitor a live read on its demand curve. Anthropic running Claude inference on Musk's hardware is a trust exercise as much as a lease. But when the alternative is not shipping — because the electrons, the interconnects, and the racks simply aren't available anywhere else — the calculus changes. In a compute famine, you rent from the enemy.