SpaceX Is Quietly Becoming an 80 Billion Dollar AI Compute Landlord (and Nvidia Funds Both Sides)
SpaceX signed a 6.3 billion dollar compute deal with Reflection AI (150 million per month from July 1), pushing committed AI-infrastructure revenue past 80 billion. The tell: Nvidia sits on both sides.
TL;DR
- SpaceX signed a compute deal with open-source AI startup Reflection AI worth up to $6.3 billion , $150 million per month starting July 1, 2026, running through 2029 (CNBC, citing Bloomberg).
- Reflection gets Nvidia GB300 chips at SpaceX's Colossus 2 data center in Memphis. Either side can exit on 90 days' notice after the first three months.
- This is one of many: SpaceX has now landed compute deals with Google (~$30B), Anthropic (~$45B), and Cursor , committed outside-client revenue exceeds $80 billion.
- The tell: Nvidia is on both sides , an investor in Reflection (last valued at $25B) and the supplier of the chips it's renting.
A rocket company just became one of the largest compute landlords in AI. Days after its record IPO, SpaceX signed a $6.3 billion deal to rent Nvidia GB300 capacity to Reflection AI , and it's the latest of several. Strip away the branding and a quiet truth emerges: under the rocket company, SpaceX is building an AI-infrastructure business with more than $80 billion in committed customer revenue, financed in a loop that runs straight through Nvidia.
What the deal actually is
Reflection AI will pay SpaceX $150 million per month from July 1, 2026, totaling roughly $6.3 billion if it runs through 2029, for immediate access to Nvidia GB300 chips housed at Colossus 2 in Memphis (CNBC, via Bloomberg). The structure is notably loose , either party can walk with 90 days' notice after an initial three months , which says something about how fast both sides expect this market to move.
Reflection itself is a two-year-old startup founded by ex-Google DeepMind researchers, Nvidia-backed, last valued around $25 billion, with no public frontier model yet but deep ties to government clients (the DoE's Genesis Mission, Pentagon AI programs). It is, in other words, a company renting enormous compute on the strength of its team and backers, not a shipping product , a recurring shape in this cycle.
The real story: SpaceX is now a neutral compute landlord
Colossus was originally built to train Grok, the chatbot from Musk's xAI. SpaceX has since turned it into one of the largest third-party compute platforms in the world, renting to whoever pays , Google (~$30B), Anthropic (~$45B), Cursor, and now Reflection. Committed revenue from outside clients now tops $80 billion. That makes SpaceX a "neocloud": a specialized landlord selling raw AI compute, the same structural bet we saw in the economics squeezing model makers. The company most associated with rockets is quietly becoming an infrastructure utility for the AI industry , including for direct rivals of its own xAI.
The circular-financing tell
Here's the part worth circling: Nvidia is on both sides of this trade. It's an investor in Reflection AND the supplier of the GB300 chips Reflection is renting. So Nvidia's capital helps fund a startup whose spending flows to SpaceX to buy , Nvidia chips. This is the circular financing pattern now threaded through the whole AI buildout: the chip vendor seeds demand for its own chips. It's not necessarily wrong, but it inflates "committed revenue" numbers that are partly the same dollars cycling, and it concentrates the entire structure's risk on one vendor's continued willingness to fund the loop.
Who wins, who's exposed
- Wins , SpaceX. It converts a data center built for its own model into an $80B+ revenue platform, diversifying far beyond launch.
- Wins , Nvidia. Every one of these deals is GB300 demand, partly of Nvidia's own making.
- Exposed , the loop. Committed revenue that depends on vendor-financed startups is only as solid as the financing. The 90-day exit clauses hint nobody treats these as permanent.
- Watch , capacity reality. Musk claimed Colossus 2 offered 1GW at launch; satellite imagery in January reportedly showed cooling for ~350MW. Committed dollars are racing ahead of confirmed physical capacity.
What this means for you
- If you track AI capex: read "committed revenue" claims with the circular-financing lens. When the chip vendor funds the buyer, headline backlog numbers overstate independent demand.
- If you buy compute: a neutral landlord like SpaceX renting to everyone is good for you , more capacity, more pricing competition against the hyperscalers.
- If you invest or plan: the binding constraints are power and chips, not press-release dollars. Watch megawatts actually energized (the 1GW-vs-350MW gap) over committed-revenue totals.
Frequently asked questions
How big is the SpaceX-Reflection deal?
Up to $6.3 billion , $150 million per month from July 1, 2026 through 2029 , for Nvidia GB300 access at SpaceX's Colossus 2 data center in Memphis, per CNBC citing Bloomberg.
Why is SpaceX selling compute?
It turned Colossus, originally built to train xAI's Grok, into a commercial compute platform. It has now signed outside-client deals (Google, Anthropic, Cursor, Reflection) exceeding $80 billion in committed revenue.
What's the Nvidia conflict of interest?
Nvidia is both an investor in Reflection AI and the supplier of the GB300 chips Reflection is renting , an example of the circular financing running through the AI buildout.
Is the capacity real?
Partly unconfirmed. Musk claimed Colossus 2 offered 1GW at launch, but January satellite imagery reportedly showed cooling capacity for ~350MW. Committed dollars are ahead of verified physical capacity.
Sources
- CNBC , SpaceX signs computing power deal with Reflection worth up to $6.3 billion
- Data Center Dynamics , SpaceX secures $6.3bn compute deal from Reflection
- TFN , Nvidia is on both sides of the trade