SpaceX Just Rented Its Supercomputer to an Open-Source Lab

A rocket company is now a cloud provider, and the market is not sure how to price it. On June 22, 2026, SpaceX signed a computing deal with Reflection AI worth up to $6.3 billion, handing the open-source startup access to Nvidia GB300 capacity inside the Colossus 2 data center. Shares of SPCX fell roughly 10 percent on the news — the stock's worst day since its Nasdaq debut. A multibillion-dollar contract that the market greets with a selloff is exactly the kind of contradiction worth sitting with.

The deal

The structure is straightforward. Reflection AI pays $150 million per month, starting July 1, 2026, for compute capacity at Colossus 2. Run the contract through 2029 and it totals about $6.3 billion. Either side can walk with 90 days' notice after an initial three-month period, which makes the headline figure a ceiling rather than a guarantee — closer to a usage commitment than a locked four-year obligation.

What it represents is bigger than the number. Colossus — the GPU supercluster originally built to train frontier models in-house — is being turned outward and sold as a commercial platform. Reflection joins a roster that already includes Anthropic, Google, and Cursor. In other words, SpaceX is no longer just a launch provider with a side in satellite internet. It is renting raw compute to AI labs, including labs that compete with each other, the same way a hyperscaler does.

A strategically different customer

Reflection is not a generic tenant. It builds open-source AI models, and it frames the deal as fuel for what it calls "American open intelligence" — capable models whose weights are available rather than locked inside a closed API. The startup has been accumulating exactly the kind of customer base that makes that framing more than marketing: it works with the Department of Energy's Genesis Mission and participates in broader Pentagon AI efforts.

That matters because of when it is happening. Governments and large enterprises are actively reassessing how much they want to depend on a handful of closed, proprietary frontier models from a few US labs. An open-weight alternative with national-security customers and a guaranteed slab of GB300 compute is a credible hedge against that concentration — and SpaceX is now the landlord underwriting it. A rocket company supplying the compute for the government's preferred open-model lab is a strange and quietly significant alignment of infrastructure, policy, and capital.

Why the stock fell

Here is the second read. A $6.3 billion contract should, on its face, be unambiguously good news for the company selling the compute. The 10 percent drop says the market sees something more complicated.

The most likely reading is about what the deal reveals rather than what it promises. Selling compute to AI labs is a capital-intensive, lower-margin business than launch or Starlink subscriptions — every dollar of revenue is backed by enormous up-front spending on GPUs, power, and cooling, and the GB300s inside Colossus depreciate fast. A 90-day exit clause means the revenue is real but not bankable in the way a long contract would be. And there is the obvious question of capacity: every GB300-hour rented to Reflection is an hour not spent training SpaceX-aligned models, which forces a judgment about whether the company is monetizing spare capacity or diverting strategic capacity to chase cash.

There is also the read that the deal underlines how much of the AI economy now runs on circular, debt-and-compute financing — labs leasing chips they cannot afford to buy, from infrastructure owners spending faster than they earn, on contracts thin enough to unwind in a quarter. A selloff on a "good" deal is the market pricing in that fragility rather than ignoring it.

The bigger pattern

Strip away the ticker reaction and the structural story is the one that lasts. The companies that own large blocks of Nvidia's newest silicon are discovering that the compute itself is a product — arguably a better business than whatever they originally built the cluster to do. SpaceX joins a short list of players large enough to turn a training supercluster into a rentable utility, and it is doing so at the exact moment demand for open-model compute, government and otherwise, is inflecting upward.

For Reflection, the win is simpler and cleaner: a guaranteed, scalable pool of frontier-class compute, at a moment when access to GB300s is the binding constraint on training competitive open models. Money buys models only if it can be converted into GPU-hours, and this contract converts $150 million a month into exactly that.

What to watch

Three things. First, whether the $150 million-a-month run rate holds past the three-month trial or quietly gets renegotiated — the 90-day clause makes this a recurring decision, not a one-time signature. Second, whether more open-source and sovereign-leaning labs follow Reflection to Colossus, which would confirm SpaceX as the default compute home for the open-model camp. Third, whether SPCX's selloff was a one-day reflex or the start of investors repricing the company as a lower-margin infrastructure business wearing a rocket company's multiple.

The launch company became a cloud company this week. The market's job now is to figure out which one it's paying for.