SpaceX inks compute deal with Reflection AI, an open-source AI lab

The Colossus Compute Pact: What SpaceX’s $150M Monthly Deal with Reflection AI Signals

SpaceX has secured a massive compute services agreement with Reflection AI, an open-source AI lab, worth $150 million per month from July 2026 through 2029. Under the deal, Reflection AI will gain priority access to Nvidia’s forthcoming GB300 AI chips and supporting infrastructure housed within SpaceX’s Colossus 2 data center near Memphis, Tennessee. This is not an acquisition or a joint venture—it is a pure compute-for-cash arrangement at an industrial scale.

What Actually Happened

Reflection AI is effectively pre-buying a dedicated slice of SpaceX’s data center capacity. The $150 million monthly figure—totaling roughly $5.4 billion over the contract’s life—covers immediate access to Nvidia’s next-generation GB300 GPUs, which have not yet been formally released. The deal begins July 1, 2026, suggesting SpaceX expects Colossus 2 to be operational and stocked with GB300 hardware by that date. For context, the original Colossus cluster in Memphis was built in record time (roughly 122 days) and currently houses over 100,000 Nvidia H100 and H200 GPUs. Colossus 2 appears to be a follow-on expansion targeting the GB300 generation.

Why This Matters

This deal reveals several structural shifts in the AI infrastructure market. First, it confirms that frontier AI training is becoming a capital-intensive, pre-committed utility—not a spot market. Reflection AI is locking in pricing and capacity years in advance, a pattern we have seen with Microsoft’s CoreWeave deals and Oracle’s multi-year GPU reservations. Second, it underscores SpaceX’s pivot from a launch provider into a serious data center operator. The company’s ability to build massive clusters quickly (the original Colossus was built in months, not years) gives it a speed advantage over traditional cloud providers. Third, the choice of an open-source AI lab as the customer is notable. Reflection AI’s commitment to open models means this compute will likely be used to train publicly released weights, potentially accelerating the availability of frontier-class open-source models.

Implications for AI Practitioners

For developers and researchers, this deal has two direct consequences. The first is capacity tightening: if a lab is willing to pay $150M/month for GB300 access, the supply of next-gen chips for smaller players will be squeezed further. Expect spot pricing for GB300 compute on cloud markets to be high through 2027. The second is that open-source model quality may leap forward. Reflection AI’s access to this hardware suggests they intend to train models competitive with closed-source frontier labs. Practitioners should watch for Reflection AI’s model releases in late 2026 or early 2027 as potential alternatives to GPT-5 or Gemini Ultra-class systems.

Key Takeaways

• Reflection AI is paying $150M/month from July 2026 for dedicated access to Nvidia GB300 GPUs in SpaceX’s Colossus 2 data center, totaling ~$5.4B over the contract.
• SpaceX is emerging as a major AI infrastructure provider, leveraging rapid construction timelines to compete with traditional cloud hyperscalers.
• The deal signals that frontier AI compute is shifting to long-term, pre-paid contracts, reducing spot-market availability for smaller firms.
• Practitioners should expect a significant open-source model release from Reflection AI in the 2026-2027 timeframe, potentially narrowing the gap with closed-source frontier labs.