SpaceX's Billion-Dollar-a-Month AI Hosting Deal Is Real. The Locked-In Revenue Isn't.

This week, from a quiet corner of the Goldman Sachs Communacopia conference, SpaceX's chief financial officer gave the headline its numbers: the company had signed its fourth major AI-hosting contract in under a year — an unnamed customer paying about $1.11 billion a month, starting December 1, for GPU capacity, power, and cooling. The stock barely moved. That reaction is the first clue to the real story. For a company that went public in June in the largest IPO in history and now trades near $2 trillion, another $13 billion a year in bookings is expected, not a surprise.

The part retail investors need to separate is which of these numbers is delivered and which is promised. The answer turns on a single oddity: the largest named customer in this business is paying for what it itself calls temporary capacity — and it can walk away on 90 days' notice.

Why a rocket company is becoming your new AI landlord

The reason hyperscalers sign leases with a rocket company is not rocket science. It's that AI's bottleneck has moved. For three years the constraint was the chip. Now it's everything around the chip: the power to run it, the permit to build the building, the months of construction before it hums. SpaceX took a shuttered appliance factory in Memphis and, in 122 days, filled it with 100,000 Nvidia H100s, then doubled it to 200,000 GPUs — a pace Nvidia's Jensen Huang called "superhuman", built by treating electricity like a launch-pad problem and staking portable gas turbines to get watts years ahead of the utilities.

That speed is the product, and the customers are not small players. Anthropic — a rival lab whose models compete directly with SpaceX's Grok — pays about $1.25 billion a month through 2029 for access to the Colossus campus. Google pays roughly $920 million a month for about 110,000 Nvidia chips. Startup Reflection pays $150 million a month. Add the new unnamed customer and the four contracts annualize to more than $40 billion a year — before a single Starlink satellite or rocket launch is counted.

The run-rate and the reality

Read the way SpaceX frames it and the number gets even bigger. The CFO tied the new deal to a path toward a $100 billion annual recurring-revenue run rate by year-end. Here is where the headline and the contract book part company, and it's the part a holder has to see. The four disclosed AI contracts alone annualize to roughly $40 billion, not $100 billion. The larger figure leans on the rest of the company — Starlink, launch — and on further bookings against a buildout that hasn't happened yet: 1.4 gigawatts of capacity installed as of June 30, a target of more than 2 gigawatts by year-end, and roughly 10 gigawatts by the end of 2027. That's a roughly sevenfold expansion in eighteen months, which management says pays back in under a year.

Now the crux. "Annual recurring revenue" is a booking, not a lock. Look closest at the largest named customer. Google, which pays three-quarters of a billion a month, describes the arrangement itself as "bridge capacity" to cover surging demand for its Gemini agents — in plain terms, peak-demand rental, not a structural commitment. It can cancel with 90 days' notice starting in 2027, and as early as October if the chips aren't delivered. Google is also an early SpaceX investor, with an executive on the SpaceX board — which raises the uncomfortable possibility that some of this "demand" reflects a strategic partner supporting a valuation as much as true scarcity. Anthropic's deal, per Musk, carries a similar 90-day termination.

So the signal cuts two ways at once. Every one of these leases reads as demand strength; every one of them is also leverage and delivery risk. The revenue is real if the cycle holds. It evaporates on 90 days' notice if it turns.

What the market is already paying for

The valuation tells you the market has already banked the buildout, not the contracts. SpaceX trades at roughly 85 times trailing sales with negative earnings, burning about $33 billion a year in free cash flow while spending $43 billion a year on capital — funded by the $75 billion raised at the IPO and the roughly $93 billion of cash it holds. Nvidia itself trades at about 17 times sales; Microsoft and Google, which own their clouds and their chips, at 9 to 11 times. That premium says investors already treat this rocket company as the clearest remaining play on AI-compute scarcity.

It is a coherent long-term thesis, and it's exactly why the near-term question is the one that matters. The contracts that support the buildout mature in 2027–2029, which is also when the hyperscalers' own capacity, and their growing reliance on their own silicon, comes online. Add the friction already visible: community lawsuits over noise and power, a NAACP suit over the gas turbines, a pause on new data centers in Memphis, more than a dozen states weighing restrictions. None of that breaks the model today. All of it raises the odds that the fastest path to watts gets slower precisely when the cancellable contracts are least forgiving.

There's no doubt SpaceX is winning this stage of the cycle — we're watching a rocket company out-build Microsoft and Google at their own constraint, because the fight for AI compute has moved from the chip to the plug. The reason to hesitate is not demand. It's that the market is already paying for the 2027 buildout, and the revenue that justifies it is written on 90-day paper. SpaceX doesn't need the AI cycle to end to become a bad trade; it needs only the timing to slip. For a leveraged bet on scarcity priced at 85 times sales, that's the whole question — and it's a question about the cycle's next phase, not about whether the deal was real.