SpaceX Is Now the AI Industry's Compute Landlord. Nobody Planned It This Way.

On June 22, an open-source AI startup called Reflection signed a deal to pay SpaceX $150 million a month for access to its Colossus data center near Memphis, Tennessee. The deal is worth up to $6.3 billion if it runs through 2029. Reflection is not a household name — it was founded in March 2024, employs around 60 researchers, and has not yet shipped a public frontier model. But it made the Pentagon's classified AI contract list in May, alongside OpenAI, Google, Microsoft, AWS, Nvidia, and SpaceX — the only startup to do so. The fact that it chose SpaceX for compute is not incidental. It is the latest data point in a story nobody has cleanly named yet: SpaceX is now the compute landlord for American AI.

The Colossus Tenant List

The Reflection deal lands Musk's company in a position that would have seemed implausible two years ago. Anthropic is paying SpaceX $1.25 billion a month for the entire output of Colossus 1 — over 220,000 Nvidia GPUs and 300 megawatts of capacity, running through May 2029. Google is paying $920 million a month starting in October, with access to approximately 110,000 GPUs through June 2029. Cursor signed a $10 billion collaboration deal in April. And now Reflection AI adds $150 million a month starting July 1, drawing on the newer GB300 chips at Colossus 2. The arithmetic of just the Anthropic and Google contracts alone puts SpaceX at over $2.3 billion in monthly compute revenue from two customers. The company has turned a data center built for xAI into one of the most lucrative infrastructure platforms in the AI industry, without launching a cloud product, without building a marketplace, and without any of the usual enterprise software apparatus that AWS or Azure requires.

The Open-Source Angle

Reflection AI's inclusion in this roster is worth examining on its own terms. The company is building what it describes as open-weight frontier models — trained on tens of trillions of tokens, eventually released publicly. Its founders are Misha Laskin, who led reward modeling for DeepMind's Gemini project, and Ioannis Antonoglou, one of the co-creators of AlphaGo. The team is small, technically elite, and deliberately building something that governments and defense contractors can use without depending on the closed ecosystems of OpenAI, Anthropic, or Google. The Pentagon's decision to include Reflection on its classified AI contract list — at Impact Level 6 and 7, the environments reserved for the most sensitive Defense Department systems — suggests the government is actively cultivating an open-source alternative to the major labs. The strategic logic is straightforward: if US AI export controls and classification requirements prevent government systems from relying on commercial API providers, you need models with open weights that can be deployed on sovereign infrastructure. Reflection is building for exactly that use case.

The Landlord Paradox

Here is the tension at the center of this story: xAI, SpaceX's sister company, is in direct competition with Anthropic, Google, and OpenAI on frontier model development. Grok competes with Claude, Gemini, and GPT-4. Yet SpaceX is renting compute to all of them — and pocketing billions in the process. The arrangement creates a peculiar dynamic. Musk's infrastructure business profits regardless of which AI company wins the model race. If Anthropic trains a better model on Colossus hardware, SpaceX gets paid. If Google uses Colossus to push its Gemini architecture forward, SpaceX gets paid. If Reflection releases the open-weight model that disrupts all of the above, SpaceX gets paid. This is a position closer to landlord than competitor — and it is more durable than any single model advantage. Models become obsolete. Compute infrastructure, especially infrastructure with the GPU density and power capacity that Colossus has built, depreciates more slowly.

What the Tenant List Reveals

The composition of SpaceX's compute customers is a map of where American AI is actually being built. The list does not include Microsoft, which has its own data center infrastructure tied to OpenAI. It does not include Meta, which is building out its own massive GPU clusters. What it does include is the set of companies — Anthropic, Google, Cursor, Reflection — that either needed capacity faster than they could build it, or needed to scale compute without the capital expenditure of owning the hardware outright. SpaceX offered both: immediate access to high-density GPU infrastructure and a contract structure that lets tenants leave with 90 days' notice. That optionality has a price — $150 million a month for Reflection, $1.25 billion a month for Anthropic — but it is cheaper than building, slower to depreciate, and available now. The result is that Elon Musk, whose public persona in AI centers almost entirely on xAI and Grok, is quietly collecting rent from the companies building the models that compete with his. The scale of that rent — measured in billions per month across the tenant list — may end up being the more historically significant part of the SpaceX AI story.