SpaceX called it a $6.3 billion deal. The 90 days' notice is doing the heavy lifting.

The number SpaceX wanted you to take away from this week was $6.3 billion. That is the figure attached to the compute deal it signed with Reflection AI, a two-year-old startup that has yet to ship a product and is already valued at $25 billion. The headline writes itself: the rocket company that went public a fortnight ago has landed a multibillion-dollar contract for the data-center business almost no one was pricing into the stock. It is a big, round, repeatable number, and like most big, round, repeatable numbers in this cycle, it is worth interrogating before you repeat it.

Start with the arithmetic, because the arithmetic is the story. Reflection will pay SpaceX $150 million a month, beginning 1 July 2026 and running through 2029. Forty-two months at $150 million is roughly $6.3 billion. But the contract, as reported, lets either party walk with 90 days' notice after an initial three-month period. Strip the headline down to what is actually committed and you are left with about a quarter of guaranteed revenue — call it $450 million — wrapped in a notional ceiling twelve times larger. The $6.3 billion is not a backlog. It is a run-rate multiplied by a duration that neither side has promised to honor.

This is not an accusation of bad faith. It is how compute contracts in 2026 are built, and the structure tells you something the press release does not: the people signing them know the ground can move. A 90-day exit is what you negotiate when you are not sure your counterparty will exist, or be solvent, or still want the chips, three quarters from now. The flexibility is mutual, and it is the most honest line in the document.

The cost center that became a landlord

To understand why this deal matters more than its dollar figure, you have to understand what SpaceX is now selling. Colossus, the cluster in Memphis at the center of all this, was not built to be a business. It was built by xAI to train and run Grok — a captive cost center, the most expensive kind of internal infrastructure, justified by the model it fed. xAI merged into SpaceX earlier this year. Somewhere in that consolidation, a cost center quietly became a product. The GPUs that were bought to make Musk's chatbot competitive are now being rented, by the rack, to anyone facing a shortage of Nvidia's latest silicon — which, at the moment, is everyone.

Reflection is not the first tenant. SpaceX has reportedly signed compute agreements with Anthropic, at around $1.25 billion a month through May 2029, and with Google, at roughly $920 million a month from October 2026. Cursor — itself now inside the SpaceX orbit after a separate acquisition — is in the mix too. Add the three disclosed monthly figures together and you get something close to $2.3 billion a month, or a run-rate north of $27 billion a year, from leasing compute. That is the number that should reframe how you think about SpaceX. The rockets and Starlink are the company everyone knows. Grafted onto them now is a neocloud — a landlord business with the gross margins of real estate and the customer base of a venture portfolio.

The GPUs bought to make Musk's chatbot competitive are now rented, by the rack, to the rivals trying to beat it.

There is a strange elegance to it. The chips depreciate whether or not Grok is using them. Renting the idle cycles to Anthropic and Google turns a sunk cost into recurring revenue. But notice who the customers are. Anthropic builds Claude. Google builds Gemini. Reflection is building open-weight models explicitly pitched as an American answer to DeepSeek. Every one of them competes, directly or eventually, with Grok. SpaceX is now in the position of selling fuel to the cars trying to beat its own car. That can be a perfectly good business — the arms dealer often outlives the war — but it is a different business from the one the IPO was sold on, with a different risk profile, and the market has not finished deciding what to pay for it.

Who is exposed if the thesis is wrong

Concentration is the risk that compounds quietly and arrives all at once, and the compute-leasing business has it in an unusual form. A normal landlord diversifies across tenants in different industries, so that a downturn in one does not empty the building. SpaceX's compute tenants are all in the same industry, at the same stage of the same cycle, funded by the same pool of capital, betting on the same proposition: that demand for frontier AI will keep growing fast enough to justify spending more on training runs than they earn in revenue. They are not independent bets. They are the same bet, wearing different logos.

Look at Reflection specifically. It was founded in 2024 by Misha Laskin, who led reward modeling for Google DeepMind's Gemini, and Ioannis Antonoglou, a co-creator of AlphaGo — a serious team, by any measure. It raised $2 billion last October at an $8 billion valuation, itself a fifteen-fold jump from the $545 million it was worth seven months earlier. By this spring the reported valuation was $25 billion. The company's first frontier model was expected early this year. The revenue line is, as far as anyone outside the company can tell, not yet a line. None of that makes Reflection a bad company. It makes it a young one, valued on a story, paying $150 million a month for compute out of capital it raised on the promise of models it has mostly not yet shipped.

So when you read that SpaceX has $27 billion a year of compute revenue coming, ask the question a credit analyst would: how good is the receivable? It is exactly as good as the tenants' ability to keep raising money. The compute lease is senior to nothing. If the AI-capex cycle cools — if the next funding round for one of these labs comes in flat, or doesn't come — the first thing that gets cut is the discretionary training run, and the 90-day notice clause stops being a footnote and becomes the headline. The revenue that looks like infrastructure is, underneath, a leveraged bet on the fundraising environment staying open. That is not what infrastructure is supposed to be.

The tell in the spokesperson's quote

Reflection's stated reason for the deal is worth reading closely, because it points at the one risk everyone in this trade is now hedging. "Recent events," a spokesperson said, "highlight how important open source is to the AI ecosystem, with more nations and enterprises recognizing the risks and costs associated with exclusively depending on closed models." The recent events are not named, but they do not need to be: less than two weeks ago, a US export-control directive forced Anthropic to switch off two frontier models worldwide for the better part of a week. The lesson the market drew was that a closed model is a dependency someone else can revoke, and that open weights you can run yourself are a kind of insurance.

Reflection is selling that insurance. But notice what it is buying to manufacture it: Nvidia GB300 chips, in a data center owned by Elon Musk, paid for in monthly installments. The independence Reflection offers its customers lives at the weights layer — the model you can download and inspect and run on your own metal. The dependency has not been removed. It has been moved one level down, to the compute. An open-weight lab that rents its training capacity from a single landlord on a 90-day lease has swapped the risk of a revoked model license for the risk of a revoked rack. The geopolitics of who controls the intelligence does not go away because the weights are public. It relocates to who controls the silicon, and the silicon is more concentrated than the software ever was.

The independence lives at the weights layer. The dependency just moved one level down, to the compute.

What the stock already knew

Here is the detail that should anchor the whole reading. SpaceX announced a $6.3 billion compute deal, and on the day, the stock — listed as SPCX after this month's debut at $135 — fell about 10 percent, one of its worst sessions since it started trading. It still sits well above the IPO price, around $165, so this is not a collapse. But it is a tell. When a company announces several billion dollars of new revenue and the market marks it down, the market is not pricing the revenue as a backlog. It is pricing the risk that came attached: the customer concentration, the cyclicality, the cannibalization of Grok's edge, and the simple fact that a rocket company's multiple is not the multiple you would put on a merchant compute business with venture-stage counterparties.

That gap — between the headline number and the market's reaction to it — is the most useful thing in this story. It is the difference between a narrative and a balance sheet. The narrative is that SpaceX has discovered a vast new profit pool in the AI buildout. The balance-sheet version is that it has taken the depreciation risk on tens of billions of dollars of GPUs and is offsetting it with monthly rent from tenants who can leave in a quarter and whose own survival depends on a fundraising cycle no one controls. Both can be true at once. A trade can be a great company and a dangerous position simultaneously, and the skill is in holding the two thoughts together.

The downside no one is pricing

So what breaks if the thesis is wrong? Not SpaceX, probably — the rockets and Starlink are real, profitable, and have nothing to do with any of this. The compute-leasing line is the part exposed, and it is exposed in a way that is easy to miss because it is dressed as infrastructure. If even one of the marquee tenants stumbles — a down round, a model that underwhelms, an export order that lands on the wrong cluster — the others are correlated, because they are funded by the same believers making the same bet. The revenue that today reads as $27 billion a year of contracted demand could compress toward the only number that was ever truly committed: the next 90 days.

This is the same mechanism that has shown up everywhere in the AI buildout, just wearing a new costume. Capital flows in a circle — chipmaker to cloud to lab and back — and at each turn someone records demand that is, one link away, someone else's borrowed money. SpaceX has found a clever place to stand in that circle: it owns the asset, collects the rent, and keeps the right to repoint the chips at Grok if the tenants disappear. That is a genuinely good position. It is just not the riskless, multibillion-dollar annuity the headline implied. The headline number was $6.3 billion. The number that tells you what the deal is actually worth is the one printed in the termination clause: ninety days.