SpaceX's Colossus Becomes the World's Most Valuable AI Compute Landlord as Reflection AI's $6.3 Billion Lease Goes Live
Reflection AI's $6.3 billion compute lease with SpaceX officially commenced on July 1, 2026, making the open-source artificial intelligence startup the latest — and most strategically significant — outside tenant at Elon Musk's Colossus data centre complex in Memphis, Tennessee. As reported by CNBC and confirmed by TechCrunch, Reflection will pay $150 million per month through the end of 2029 for immediate access to Nvidia's latest GB300 Blackwell Ultra chips at SpaceX's Colossus 2 facility. The contract includes a 90-day termination clause exercisable by either party after the first three months, giving both sides structural flexibility in a market where chip availability and model timelines can shift rapidly.
The deal cements SpaceX's extraordinary pivot from rocket company to AI compute landlord. According to Tech Funding News, SpaceX has now secured more than $80 billion in committed compute revenue from outside clients through 2029. Anthropic signed on in May 2026 to lease all capacity at the original Colossus 1 site for approximately $45 billion, paying roughly $1.25 billion per month. Google followed in early June with a commitment worth approximately $30 billion over the same period at $920 million per month. The Colossus complex — originally built by xAI to train Grok, Elon Musk's AI assistant — now houses approximately 555,000 Nvidia GPUs, with a planned total power capacity of two gigawatts, according to Tech Funding News. SpaceX priced its own IPO at $135 per share in June 2026 at a valuation of $1.77 trillion, the largest in history.
Reflection AI was founded by two former Google DeepMind researchers — Misha Laskin, who led reward modelling for the Gemini project, and Ioannis Antonoglou, DeepMind's sixth-ever researcher and a co-creator of AlphaGo — and carries a valuation of $25 billion. The company is backed by Nvidia, Sequoia, and Lightspeed, according to reporting aggregated by Build Fast with AI. Nvidia's involvement creates an unusual circular structure: the chipmaker invested approximately $800 million in Reflection and simultaneously supplies the GB300 hardware that SpaceX sold to house it — making Nvidia an investor in, and an indirect supplier to, the same customer. As The Next Web observed, this kind of circularity is becoming structurally common across the AI infrastructure buildout.
Reflection has positioned itself explicitly as an American open-source alternative to China's DeepSeek, arguing that recent government interventions against closed models — most acutely the nineteen-day suspension of Anthropic's Fable 5 — have exposed the fragility of depending exclusively on proprietary providers. A Reflection spokesperson stated that the compute deal gives the company the runway to build the world's best open models at scale. The company has also established formal ties with the US government, working with the Department of Energy's Genesis Mission and various Pentagon AI programmes, according to Open Source For You. However, as of early July 2026, Reflection had not yet publicly released a frontier open-weight model, and its code research agent Asimov remained on a waitlist, a gap between pedigree and product that investors will be watching closely.
The broader significance of the Colossus deal is what it reveals about the new logic of the AI arms race. Access to top-tier Nvidia chips has become the central constraint for any company attempting to train or serve frontier models, and leasing capacity from an established compute landlord — rather than building data centres from scratch — allows startups to deploy capital toward research rather than concrete and cooling systems. As Tech Funding News noted, SpaceX's transformation from internal training cluster to multi-tenant platform mirrors moves by cloud hyperscalers, but at unusual scale and speed. Whether Reflection ultimately ships models capable of competing with Anthropic, OpenAI, and Google is the question the next twelve months will answer — but with $6.3 billion of compute time now contracted, it will have the runway to try.