Reflection AI Signs $6.3B Compute Deal With SpaceX as Valuation Hits $25B
An open-source AI company less than two years old just signed a compute contract larger than most countries' annual defense budgets.
Three Key Takeaways
Reflection AI will pay SpaceX $150 million per month for Nvidia GB300 chips at the Colossus 2 data center, starting July 1 through the end of 2029 — roughly $6.3 billion if the contract runs its full term. This is SpaceX's fourth major external compute lease for Colossus. According to CNBC, SpaceX's committed compute revenues from outside clients now exceed $80 billion — a figure it amassed in roughly two months, putting it on par with cloud giants that took years to build comparable books.
Reflection AI's valuation is climbing at rocket speed. It emerged from stealth in March 2025 at a $545 million valuation, raised $2 billion at $8 billion in October 2025 (Nvidia alone invested $800 million), and is now seeking $2.5 billion more at a $25 billion valuation. Nvidia, Sequoia, and Lightspeed are all backing it.
The company's goal is to build American open-source frontier models that can compete head-to-head with OpenAI, Anthropic, and Google. In a landscape where Meta's Llama dominates open-source, Reflection wants to prove that an open-source player with enough capital and top-tier silicon can produce closed-source-caliber models.
WangDou's Take
A 46x valuation jump in 18 months — even SpaceX didn't pull that off in its early days. But look at the math: $150 million a month in compute rent alone means Reflection burns through more on electricity bills than most AI companies raise in an entire year. Its weapon is open source, yet nobody has cracked open-source monetization at scale: Meta funds Llama with ad profits, Mistral is still searching for a revenue model. Reflection is sitting on a $25 billion valuation with no mature product. Its only exit strategy is to outrun GPT and Claude — using someone else's chips in someone else's data center. This isn't a startup. It's a burn-rate championship in the AI arms race.