Reflection AI Rents $6.3 Billion of Colossus, and the Landlord Era Arrives

Reflection AI, the two-year-old lab valued at $25 billion on roughly $2 billion raised, has signed a deal worth up to $6.3 billion for compute capacity at SpaceX's Colossus 2 site in Memphis — Nvidia GB300 clusters, at a reported $150 million per month, running from July 1 through 2029. The arrangement carries a notable asymmetry: after an initial three-month commitment, Reflection can walk away on 90 days' notice. It becomes the third external tenant on the Colossus campus, after Anthropic and Alphabet.

Read that tenant list again, because it is the strangest sentence in AI infrastructure this year. The cluster built to train Grok — the flagship asset of a company whose founder is in open competition with every one of those tenants — is now a multi-tenant facility renting frontier capacity to three rivals. The explanation is not diplomacy. It is economics: a campus of that scale is a capital commitment so large that filling it matters more than who fills it. Compute has crossed the line from weapon to real estate.

The tenant's math

For Reflection, the logic of renting is straightforward and brutal. A lab with $2 billion raised cannot build a GB300 campus — the buildout it is renting a slice of costs multiples of its entire funding history. What it can do is convert investor capital directly into training runs, starting July 1, without waiting three years for its own concrete to cure. The $150 million monthly rate annualizes to $1.8 billion — a burn that consumes its entire raise in just over a year if revenue does not arrive — but the 90-day termination clause converts what looks like a $6.3 billion commitment into something closer to a rolling option on frontier capacity. Reflection is not betting $6.3 billion. It is betting $450 million at a time, quarterly, on continuing to believe its next model justifies the rent.

That option structure is the detail worth generalizing. The first wave of compute deals this cycle were take-or-pay commitments measured in years — the model I covered when OpenAI and Oracle turned procurement into universal credit lines. A 90-day out clause prices in something the industry did not admit a year ago: that a lab's demand for compute is only as durable as its latest training result, and both landlord and tenant now know it.

The landlord's math

For SpaceX, the deal answers a different question: how do you underwrite a campus whose cost runs to tens of billions when your own AI subsidiary cannot consume it all? The answer every hyperscaler discovered a decade ago — tenancy — is now being rediscovered at the frontier. Three external tenants turn Colossus from a cost center attached to one lab's ambitions into an asset with a rent roll, which is to say: into something a lender can value. This is the same structural move as Anthropic's special-purpose-vehicle chip financing earlier this month — the industry converting compute from an expense into a financeable asset class — executed from the supply side rather than the demand side.

The pattern now has three legs in a single month: chips financed through SPVs, sovereign campuses underwritten by states as national infrastructure, and now frontier clusters operated as rental property with a diversified tenant base. Different instruments, one direction: the capital structure of AI is being rebuilt to look like real estate, because real estate is the asset class that already solved the problem of financing enormous, long-lived, location-bound capacity consumed by tenants whose businesses are more volatile than the asset.

What it says about the market

Three observations fall out. First, the frontier is renting from the frontier — the fiction that proprietary compute is itself the moat gets weaker every time a rival trains on your landlord's silicon. What remains proprietary is allocation priority, and it is notable that every Colossus tenancy reported so far is for capacity, not priority; who gets throttled first in a crunch is the clause nobody publishes. Second, the neocloud category — CoreWeave and its cohort — just acquired its strangest competitor: a rocket company with a defense-scale balance sheet and a captive anchor tenant. Third, the 90-day clause cuts both ways, and the side bearing the risk has quietly flipped: if Reflection's next model disappoints, SpaceX re-lets the capacity into a seller's market; if compute prices fall, Reflection walks in a quarter. Both parties have priced the possibility that the other's world changes fast — which is the most honest statement about 2026 AI economics either of them has made publicly.

What to watch: whether the rumored fourth and fifth Colossus tenancies materialize, which would confirm the campus is being run to maximize occupancy like any other REIT-shaped asset; whether Reflection's July 1 start date holds, since GB300 delivery schedules have slipped elsewhere; and whether any tenant ever discloses the allocation-priority terms — the first one to do so will tell us who actually owns the frontier when the capacity gets scarce.