Reflection signs a $1 billion compute deal weeks after its SpaceX pact
The open-weight lab is stacking multi-year capacity contracts like an airline books fuel. What agent intelligence costs downstream is being set in deals like this one.
- Reflection AI signed a deal worth more than $1 billion for compute from Nebius running through 2029, including access to Nvidia's GB300-generation chips, per TechCrunch and Reuters.
- It lands weeks after Reflection's compute pact with SpaceX; the startup is valued at $25 billion pre-money with close to $2.6 billion raised from Nvidia, Sequoia, and Lightspeed.
- Nebius has become the capacity dealer of the moment: Meta signed a deal worth up to $27 billion and Microsoft up to $19.4 billion, with Nvidia investing $2 billion in the provider.
Reflection AI, the open-weight frontier lab founded by two former DeepMind researchers, signed a deal worth more than $1 billion to buy compute from Nebius, TechCrunch reported July 14. The capacity runs through 2029 and includes access to Nvidia's GB300-generation chips, per Reuters. It comes just weeks after Reflection signed a separate compute arrangement with SpaceX, and it makes the pattern unmistakable: model labs are booking capacity years ahead, the way airlines hedge fuel.
The capacity market has a market maker
Nebius is quietly becoming the dealer at the center of the table. Meta signed a five-year deal with the provider worth up to $27 billion, Microsoft carries a multi-year commitment worth up to $19.4 billion, and Nvidia has invested $2 billion in the company, per TechCrunch's tally. Reflection, valued at $25 billion pre-money with close to $2.6 billion raised from backers including Nvidia, Sequoia, and Lightspeed, is now on that client list. Reuters' framing of the deal is the beat thesis in one line: AI startups are racing to lock in compute as enterprise demand growth outpaces new data-center supply, per the wire report.
Why a compute contract is a pricing story
Every per-token price on every rate card is downstream of deals like this. When capacity is scarce enough that labs prepay years of it, you get exactly the market we have been covering: a flagship model rationed off subscriptions because its maker cannot serve demand, JPMorgan's CFO forecasting token-cost acceleration on an earnings call, and an infrastructure trade so crowded it can halt a national stock market on a wobble. The agent economy's COGS is being negotiated in private, multi-year contracts, and the public pricing follows them, not the other way around.