Elon Musk's rocket company is now, functionally, a data center landlord. SpaceX has signed a computing power agreement worth up to 6.3 billion dollars with Reflection AI, an open-source AI startup founded by two former Google DeepMind researchers, according to CNBC. Reflection will pay 150 million dollars a month starting July 1 for access to Nvidia's top-tier GB300 chips housed at Colossus 2, the supercomputer complex SpaceX built near Memphis, Tennessee. A rocket and satellite company is now one of the largest sellers of AI computing capacity on the planet, and it happened faster than almost anyone in the industry expected.

A Rocket Company Just Became an AI Chip Landlord

Colossus was never built for this. Musk's AI venture xAI, now folded into SpaceX, constructed the facility to train its own chatbot, Grok, calling it a "gigafactory of compute." Somewhere along the way, SpaceX realized it was sitting on more Nvidia GPU capacity than xAI alone could use, and started renting the surplus to outside companies. Reflection is the fourth publicly known tenant, following deals with Anthropic, Google, and the AI coding startup Cursor, which SpaceX is separately in the process of acquiring. Combined, these external compute agreements now total more than 80 billion dollars in committed revenue through 2029, according to reporting from MLQ News, a figure SpaceX assembled in roughly two months.

That speed is the real story here. Traditional cloud providers like Amazon, Microsoft, and Google spent years building the customer relationships and infrastructure trust needed to sell compute at hyperscale. SpaceX backed into the same position almost by accident, simply because it had excess capacity sitting inside a facility built for an entirely different purpose.

The Deal Itself, and What Makes It Strange

The terms are straightforward on paper: Reflection gets immediate access to GB300 chips and supporting hardware at Colossus 2, pays 150 million dollars monthly, and either side can walk away with 90 days' notice once the first three months pass. What makes the deal genuinely unusual is the buyer. Reflection AI has no public product. It was founded in 2024, carries a 25 billion dollar valuation, and counts Nvidia among its own investors. A company with nothing on the market yet is committing to a payment schedule that could reach 6.3 billion dollars over the life of the contract, purely to secure chips before it has anything to run on them.

That is not how most companies buy infrastructure. Normally, revenue or at least a working product comes first, and the infrastructure spend follows to meet proven demand. Reflection is doing it backward: securing the scarce physical resource first, on the belief that chips, not capital, are the actual bottleneck standing between the company and a competitive frontier model.

Why a Company With No Product Needs a Trillion-Dollar Supercomputer

Reflection's bet only makes sense once you understand where it is trying to compete. The company describes its mission as building an "open frontier lab" for national security customers, and it already has real government relationships to back that framing, including work tied to the Department of Energy's Genesis Mission and multiple Pentagon AI programs. Its plan is to release open-weight models, meaning the underlying parameters are publicly available for governments and enterprises to inspect, modify, and run on their own infrastructure, rather than depending entirely on a closed system controlled by one vendor.

That pitch has gained real urgency recently. The most capable open-weight models developers have been reaching for lately have increasingly come from Chinese labs, a trend that has become a genuine source of strategic concern in Washington. Reflection's timing also benefits from a specific American cautionary tale: when the U.S. government temporarily restricted access to Anthropic's Claude Fable and Claude Mythos models earlier this year, it demonstrated to enterprise and government customers exactly what can happen when you depend on a closed model that a single Friday-afternoon directive can switch off. An open-weight American alternative, backed by guaranteed compute, is a direct answer to that anxiety, whether or not that was the explicit intent behind the deal's timing.

The Loop Nvidia Sits Inside

There is a closed circuit running through this entire arrangement that is worth pulling apart. Nvidia invested 800 million dollars in Reflection AI. Reflection will now spend that same money, and considerably more, running its models on Nvidia GB300 chips that SpaceX itself purchased for roughly 18 billion dollars. Nvidia sits on both sides of the transaction: as an investor benefiting from Reflection's success, and as the chipmaker profiting from the hardware Reflection is contractually obligated to rent. That structure is not evidence of wrongdoing, but it is a clear illustration of how concentrated the current AI infrastructure boom has become. A handful of companies are simultaneously funding, supplying, and profiting from the same small set of deals, and untangling genuine organic demand from circular capital flows is getting harder with every new agreement announced.

The Bigger Pattern: Colossus's Fast Pivot

Look at the sequence of deals and a clear pattern emerges. Anthropic signed on in May to lease the entirety of Colossus 1's original capacity, more than 220,000 Nvidia GPUs, in an agreement valued around 45 billion dollars through mid-2029. Google followed in early June with a roughly 30 billion dollar commitment at Colossus 2. Reflection is the newest, smallest, and structurally strangest of the four, a pre-revenue research lab writing nine-figure monthly checks. Each new deal adds credibility to SpaceX's pivot from aerospace company to AI infrastructure provider, and each one further tightens the industry-wide chip shortage that made this entire business model possible in the first place.

Why the Market Punished SpaceX for Getting Paid

Here is the part that should genuinely confuse anyone reading the news casually: SpaceX's stock reportedly dropped roughly 10 percent around the time this deal became public, despite the agreement adding real, recurring revenue to a company that priced its own IPO at a 1.77 trillion dollar valuation on June 12. That reaction says less about the Reflection deal itself and more about investor uncertainty over what kind of company SpaceX is becoming. A rocket and satellite business with a fast-growing side venture renting scarce AI chips to startups is a genuinely new category, and markets tend to punish businesses they cannot cleanly categorize, at least until the pattern repeats often enough to look normal. Given how quickly SpaceX has stacked up four separate multibillion-dollar compute tenants in a matter of months, normal may arrive faster than the stock price currently suggests.