SpaceX has signed a compute contract with Reflection AI worth up to $6.3 billion, adding the open-source AI startup to a customer list that already includes Anthropic and Google. The deal, reported by CNBC on June 22 citing materials it reviewed directly, confirms a pattern that has been building for months: SpaceX is operating Project Colossus as a merchant compute facility, not just an internal resource for Grok.
Under the terms, Reflection gains access to Nvidia GB300 accelerators starting July 1, 2026, at $150 million per month through 2029. Either party can exit with 90 days’ notice after the first three months, which caps SpaceX’s revenue lock-in but gives Reflection a practical off-ramp if something better emerges.
AI Insiders previously reported SpaceX compute deals with Anthropic and with Google. The Reflection agreement is the third publicly named customer, and the first whose primary business is open-source model development. That distinction matters for the compute market. Open-source labs have historically been squeezed on GB300 access because hyperscalers prioritize their own closed offerings and large enterprise customers. SpaceX, with no competing model product in the open-source space, is a structurally neutral supplier for these labs.
Reflection, last valued at $25 billion, has so far shipped no open-weight frontier system to the public. Its existing traction comes from government and national security customers, including the Department of Energy’s Genesis Mission and Pentagon AI programs. Its pitch, sharpened after Anthropic discontinued access to its Fable and Mythos models, is that open models let governments and enterprises inspect, customize, and run AI without dependence on a single closed provider. A Reflection spokesperson quoted by CNBC said the deal accelerates what the company calls “American open intelligence.”
For Colossus, the Reflection deal is evidence that SpaceX’s monetization thesis is holding across customer types. Anthropic and Google are closed-model shops with massive internal compute needs. Reflection is an open-source lab with government contracts and a frontier training run still ahead of it. The fact that Colossus is attractive to both categories suggests SpaceX’s GB300 supply and pricing are competitive with the hyperscaler alternative, not just a convenience play for companies already inside Musk’s orbit.
The structural constraint here is what happens when Colossus capacity fills. SpaceX built the Memphis facility fast, but GB300 allocation is finite. Three named customers paying at Reflection’s rate represent enormous committed capacity. Any open-source lab currently searching for GB300 time outside AWS, Azure, or Google Cloud should treat Colossus as a real alternative to evaluate now, before the remaining slots are committed.
Open-source labs with active frontier training runs and government-friendly positioning should move quickly. Colossus is no longer a rumor or a single-customer story. It is a named, multi-customer compute platform with a price point and contract structure that has closed at least three deals above $1 billion.
Reporting by Deirdre Bosa for CNBC, published June 22, 2026.
