Imagine a pop-up chef who’s famous for a single, brilliant sauce suddenly renting out a fully staffed, Michelin-starred kitchen in the middle of Manhattan for the next three years. The chef doesn’t own the stoves, the pans, or the building, but they have the keys to the most expensive equipment on the planet. That’s essentially what we’re looking at with the deal between SpaceX and Reflection AI.
The numbers are genuinely absurd. According to the report, Reflection AI is committing to $150 million a month starting July 1, 2026, running through 2029. That is a staggering $1.8 billion a year just to keep the lights on and the GPUs humming. They aren’t buying the hardware; they’re paying for immediate access to Nvidia’s GB300 AI chips and the supporting infrastructure inside SpaceX’s Colossus 2 data center in Memphis. For those who haven’t tracked the hardware sprawl, Memphis is becoming an unlikely hub for compute, largely because the power grid can actually handle the load (and probably because the tax breaks are too good to pass up). But paying $150 million a month is a level of burn that would make even the most aggressive Silicon Valley VCs break into a cold sweat.
Here is where the math stops making sense: Reflection AI calls itself an “open-source AI lab.” We’ve seen this movie before. Usually, “open source” in the LLM space is a convenient label used by labs that can’t afford their own clusters or want to outsource the fine-tuning to a million unpaid developers on Hugging Face. Who actually signs a check for $1.8 billion a year in the name of “open source”? Does anyone actually believe that a lab spending this much on private compute is doing it for the public good? If you’re spending that kind of money on private compute, you aren’t building a community resource; you’re building a moat. You’re just letting the public see the weights so they can help you find the bugs for free, while you keep the real proprietary sauce locked in the Colossus 2 vault.
Then there is the hardware friction. The GB300 is a beast, but the logistics of a Memphis-based cluster mean Reflection is completely tethered to SpaceX’s infrastructure. They aren’t just renting FLOPS; they’re renting an ecosystem. If SpaceX decides to prioritize their own internal Starship-related AI or Tesla-adjacent projects, Reflection is just another tenant in a very expensive building. It’s a precarious position for a lab that claims to be leading the charge in open research. It’s like renting a Ferrari but having to let the owner decide when you can take it out of the garage, or worse, having the owner decide how much fuel you’re allowed to burn during peak hours. (Or maybe I’m being too cynical and SpaceX is just playing the role of the ultimate landlord).
The financial backing required for this is the real mystery. You don’t get a $5.4 billion compute contract by selling “openness” to a board of directors. You get it by promising a return that justifies the spend. The tension between the “open” label and the “private cluster” reality is a gap wide enough to drive a Falcon 9 through, and the smell of corporate capture is already wafting from Memphis. We’ve seen this pattern before with other labs that started with a mission of transparency only to pivot the moment the compute bill arrived. The reality is that open source is a great marketing strategy, but it’s a terrible way to pay for GB300s.
I’ll bet my last H100 that this “open” philosophy lasts exactly as long as it takes to reach the first major commercial milestone. By Q4 2026, Reflection AI will pivot their license to a “weighted open” model—essentially a “free for researchers, pay us for production” scheme—to start clawing back that monthly $150 million. The industry has a habit of starting with a handshake and ending with a restrictive EULA that makes you wish you’d just used a smaller, truly open model from the start.
This is just a glorified rental agreement.
