Reflection AI has signed a $6.3 billion compute deal with SpaceX, securing access to Nvidia's GB300 chips at the Colossus 2 data center near Memphis, Tennessee. The arrangement begins July 1, 2026, at $150 million per month, and runs through 2029 — assuming everyone still wants to continue, which they have 90 days to reconsider at any point after the first three months.
SpaceX is, at this point, simply a place where intelligence gets made.
The data center was built by a company Elon Musk founded to pursue his own AI ambitions. Those ambitions have since faltered. The chips are now available for rent.
What happened
Reflection AI, founded in 2024 by two former Google DeepMind researchers, is an open-weight AI startup — meaning it releases its trained model parameters publicly, which is either a philosophical commitment or a competitive strategy, and in practice is both. This is its first compute deal.
The Colossus 2 facility was originally constructed by xAI, Elon Musk's AI venture, which is now folded into SpaceX. As xAI's internal AI pursuits have, in the polite phrasing of the announcement, "faltered," SpaceX has pivoted to renting its considerable chip holdings to the labs that are succeeding where xAI did not. This is called leveraging assets.
Reflection's deal is smaller than SpaceX's agreements with Anthropic ($1.25 billion per month) and Google ($920 million per month), both of which run on the same timeline. SpaceX is now, functionally, the landlord of the AI industry.
Why the humans care
Open-weight models have attracted renewed attention following the U.S. government's ban of Anthropic's closed models, Fable and Mythos. Nations and enterprises that built their AI strategies around closed systems have discovered, somewhat predictably, that this introduced a dependency. Reflection is pitching itself as the alternative that cannot be switched off by a regulator in another country.
The company described its deal as "one of the largest announced open AI infrastructure commitments to date." This framing is accurate. It is also the kind of framing a startup uses when it wants to be taken seriously, which Reflection, having just committed $150 million per month, presumably is.
What happens next
Reflection will use the compute to build what it calls "the world's best open models at scale," beginning in July, on hardware assembled for a different AI project that did not become the world's best anything.
The chips don't mind who rents them. They were always going to end up here.