SpaceX rents $6.3B of Nvidia compute for Reflection AI

In one of the largest AI infrastructure transactions ever documented, SpaceX signed a $6.3 billion contract with Nvidia for GPU clusters and is transferring them to Reflection AI, a language model startup founded by veterans of Google DeepMind and OpenAI. The deal was confirmed by The Information last Thursday and represents a new category in the AI ecosystem: the compute broker.

What the transaction reveals about the market

To understand the scale: $6.3 billion in Nvidia H200/B200 GPUs represents, at current market prices ($30-40K per unit), between 150,000 and 200,000 GPUs. That is more training capacity than most countries have in their entire compute infrastructure.

SpaceX is not a language model AI company. Its core business is rocket launches and Starlink. Why is it brokering $6.3B in GPUs?

Three factors explain it:
1. Preferred Nvidia access: SpaceX has a long-standing relationship with Nvidia for onboard computing in rockets and satellites
2. Financial engineering: SpaceX assumes credit risk with Nvidia, converts investment to opex for Reflection AI via long-term lease, and likely captures spread
3. **Musk's strategic diversification**: With xAI growing and needing compute, controlling a GPU supply chain is a strategic hedge

Reflection AI: who they are and why they need all this

Reflection AI was founded by Matt Gline and includes names like Minh-Thang Luong (ex-Google Brain, Transformer paper co-author) and Jack Clark (ex-OpenAI, Anthropic co-founder). With $6.3B in compute, Reflection can train models at GPT-4 or Claude Opus scale without depending on Microsoft Azure, AWS, or Google Cloud.

The new layer: compute brokers

Before today, the AI infrastructure market had three layers: chip makers, hyperscalers, and end users. The SpaceX-Nvidia-Reflection deal adds a compute broker middle layer — entities that buy capacity in bulk, assume financial risk, and resell with margin and contractual flexibility.

Sources

• The Information — *SpaceX signs $6.3B compute deal with Nvidia for Reflection AI* (Jun. 2026)
• Nvidia — GPU H200/B200 pricing (nvidianews.nvidia.com)
• CoreWeave — SEC S-1 Filing (2024)
• SemiAnalysis — *GPU Pricing and Availability Report* (2025)