SpaceX Signs $150M/Month Compute Deal with Reflection AI
Reflection AI, an open-source AI lab, has signed a $150 million per month compute contract with SpaceX starting July 1, 2026, securing access to Nvidia GB300 chips at SpaceX's Colossus 2 data center through 2029.
Original sourceReflection AI has committed to one of the largest known compute contracts in AI history, agreeing to pay SpaceX $150 million per month — roughly $1.8 billion annually — for access to Nvidia's GB300 AI chips and supporting infrastructure at SpaceX's Colossus 2 data center. The deal runs from July 1, 2026 through 2029, locking in roughly $5.4 billion in total compute spend over the contract's life.
Colossus 2 is SpaceX's second large-scale AI data center, following the original Colossus facility built in Memphis, Tennessee. The GB300 chips represent Nvidia's most recent generation of AI accelerators, offering substantially improved memory bandwidth and inference performance over the H100s that dominated the previous training cycle. Securing guaranteed access to this hardware at scale is increasingly the bottleneck separating labs that can train frontier models from those that cannot.
Reflection AI positions itself as an open-source lab, which makes the scale of this commitment notable — open-source labs have historically operated with smaller compute budgets than closed frontier labs. Whether the label holds at this compute level, and what open-source means for models trained on $1.8 billion per year of hardware, will be worth watching as the lab's first major model releases approach.
The deal also underscores SpaceX's growing role as a commercial AI infrastructure provider, diversifying its data center business beyond internal use for Grok and xAI workloads. With Colossus 2 now signed as a commercial compute facility, SpaceX is competing directly with hyperscalers and GPU cloud providers for the largest AI training contracts in the market.
Panel Takes
The Futurist
Big Picture
“The thesis here is that compute access is the primary moat in AI, and that locking in GB300 capacity at scale before demand peaks is worth paying a premium that would look irrational on a per-token basis. If that bet is right, Reflection AI just bought three years of frontier-class training capacity while competitors scramble for spot instances. The second-order effect nobody is talking about: SpaceX is quietly becoming a hyperscaler, and the margin structure of a rocket company running GPU clouds is structurally different from AWS or Azure — that's a competitive dynamic that doesn't resolve cleanly for the incumbents.”
The Founder
Business & Market
“$150 million a month is a number that ends companies, not scales them — which means Reflection AI either has a revenue model that hasn't been announced yet, or they have funding commitments that make this math work on paper. The open-source positioning is strategically interesting as a distribution play, but open-source doesn't pay $1.8 billion a year in compute bills; something downstream does, and until we know what that is, the business model is the story. The moat question is real: if your defensibility is 'we trained on more GB300s than the other open-source lab,' that's not a moat, that's a head start with an expiration date.”
The Skeptic
Reality Check
“The phrase 'open-source AI lab' doing $5.4 billion in compute spend over three years deserves more scrutiny than it's getting — open-source is a go-to-market strategy at this scale, not a mission statement, and the lab's actual release commitments should be examined carefully before that label is accepted at face value. The specific scenario where this breaks: compute costs drop 40% by 2028 due to hardware improvements and Reflection AI is locked into 2026 pricing through 2029, which is a real risk in a market where GPU economics have moved faster than most multi-year contracts anticipated. My prediction is that either the open-source framing quietly narrows to weights-only with commercial restrictions, or a better-capitalized closed lab ships first and reframes what 'frontier' means before this compute spend produces a competitive model.”
The Builder
Developer Perspective
“What I actually care about here is whether 'open-source' means weights, training code, and evaluation harnesses — or whether it means a model card and a Hugging Face upload with a custom license buried in the README. $1.8 billion a year of compute means nothing to a developer trying to fine-tune or extend these models if the artifacts that come out of it are open-weights-but-not-open-source in any meaningful sense. The GB300 access story is interesting infrastructure news, but the primitive I'm waiting to evaluate is the actual API surface and weight release format — everything else is a press release.”