SpaceX signs $150M/month compute deal with Reflection AI for Nvidia GB300 access at Colossus 2

What Happened

Reflection AI has agreed to pay SpaceX $150 million per month for dedicated compute access starting July 1, 2026, through 2029. The deal provides the open-source AI lab with exclusive access to Nvidia's upcoming GB300 AI chips and supporting hardware, hosted at SpaceX's Colossus 2 data center near Memphis, Tennessee. Over the three-and-a-half-year term, this equates to a staggering $6.3 billion infrastructure commitment.

Technical Details

The explicit focus on Nvidia's GB300 indicates a highly forward-looking architectural strategy. The GB300 represents the next evolution in Nvidia's Blackwell roadmap, expected to offer massive leaps in memory bandwidth and NVLink interconnect speeds necessary for training trillion-parameter dense models. By securing rack space directly in Colossus 2, Reflection AI is bypassing traditional cloud service providers (CSPs) to get closer to the bare metal. This deployment will undoubtedly rely on advanced direct-to-chip liquid cooling and high-radix networking topologies required to cluster GB300s efficiently at frontier-model scale.

Why It Matters

From an engineering perspective, this deal is a bellwether for two major industry shifts. First, it proves that open-source AI labs are securing the astronomical capital required to compete head-to-head with closed-source giants like OpenAI and Google on raw compute. Second, it highlights SpaceX's aggressive expansion into the hyperscaler space. By leveraging their immense power agreements and rapid infrastructure deployment capabilities in Memphis, SpaceX is effectively monetizing its data center expertise and capacity, transforming from a pure consumer of AI compute into a massive merchant provider.

What to Watch Next

Monitor the power provisioning and cooling infrastructure build-out at the Memphis Colossus 2 site; deploying GB300s at a $150M/month scale will require hundreds of megawatts of highly reliable power. Additionally, track Reflection AI's model release cadence leading up to 2026. They will need to demonstrate exceptional algorithmic and distributed training efficiency to justify this massive CapEx burden to their financial backers.