Reflection AI commits $6.3B through 2029 for SpaceX Colossus GB300 capacity; $150M/month sovereign inference spend

Reflection AI, a pre-revenue AI research lab founded by ex-DeepMind talent and backed by NVIDIA, signed a $6.3 billion deal through 2029 to rent NVIDIA GB300 chips from SpaceX's Colossus Memphis data center, beginning at $150 million monthly. The startup has not shipped a public product—no chatbot, no API, no revenue—yet committed to full-price, long-term compute. The deal includes a 90-day exit option after the first three months, but the willingness to lock in premium-tier silicon capacity signals confidence in open-source frontier model demand outpacing supply.

Reflection joins Anthropic, Google, and Cursor as major Colossus tenants. Anthropic alone leases all of Colossus 1 at $920 million monthly for bridge capacity while its own data centers ramp. Reflection's move demonstrates that demand for bleeding-edge compute is no longer confined to hyperscalers and frontier labs with shipping products; it now extends to well-funded research teams betting on building open-weight models for national security and geopolitical independence. GB300 availability is the limiting factor—NVIDIA cannot build chips faster than customers can commit capital to rent them.

For infrastructure planners, Reflection's $150M/month burn on R&D-stage compute is a north star: inference capacity is now treated as strategic national infrastructure, commanding sovereign capital and long-term leases at hyperscaler-scale pricing. The fact that SpaceX stock fell 10% on news of the deal (implying market uncertainty about the startup's ability to execute) suggests investor skepticism about Reflection's odds—but the deal proves that access to leading-edge silicon, not execution certainty, is the prize in 2026's AI infrastructure arms race.