Reflection AI signs $6.3 billion compute deal with SpaceX for NVIDIA GB300 chips at Colossus 2

Open-source AI startup Reflection AI has signed a compute agreement with SpaceX worth up to $6.3 billion, gaining exclusive access to NVIDIA's latest GB300 chips at the Colossus 2 data center near Memphis, Tennessee. The deal is one of the largest compute contracts ever signed by an AI startup and underscores the intensifying race among frontier labs to secure dedicated GPU capacity.

What's new

Reflection AI will pay $150 million per month starting July 1, 2026, through 2029, for immediate access to NVIDIA GB300 AI chips and their supporting high-speed network fabrics at SpaceX's Colossus 2 facility. If the agreement runs its full course, total payments will reach approximately $6.3 billion.

Key deal terms:

• GPU hardware: NVIDIA GB300 (Blackwell Ultra) chips — NVIDIA's most current large-scale training and inference silicon
• Facility: SpaceX's Colossus 2 hyperscale data center, Memphis, Tennessee
• Duration: July 2026 through 2029, with a 90-day exit option after the first three months
• Reflection AI valuation: approximately $25 billion, backed by NVIDIA

For context, SpaceX has also signed compute agreements with Anthropic ($1.25B/month) and Google ($920M/month) at its Colossus facilities, though those are larger deals anchored to older Colossus 1 capacity.

Context

Reflection AI was founded in 2024 by two former Google DeepMind researchers. The company has staked out a position as an open-source alternative to closed frontier labs — building open-weight models rather than proprietary systems. NVIDIA's backing of Reflection AI aligns with NVIDIA's broader strategy of funding multiple competing frontier labs, reducing dependency on any single customer.

Colossus 2 is SpaceX's second hyperscale data center, built to host GB300-class clusters. SpaceX has increasingly positioned its Colossus data centers as a compute-rental business separate from xAI — leasing capacity to outside labs while Musk's xAI company separately operates Colossus 1 for Grok model training.

Why it matters

The scale of this deal signals that even well-funded open-source AI labs now require compute commitments rivaling those of the largest closed-model companies. Reflection AI's bet on GB300 hardware at this price point suggests the company is preparing to train frontier-class models in the coming years, not just fine-tuning or inference.

The deal has a clear strategic logic: open-source AI requires the same compute as closed AI, but the business model lacks the recurring API revenue to fund it conventionally. Locking in GB300 access before broader cloud availability gives Reflection AI the raw capacity to train and release open-weight models at a scale that could challenge GPT-class systems.

For SpaceX, Colossus 2 is quietly becoming a significant revenue line independent of the satellite and launch businesses — a position in the AI infrastructure market that is only growing as training and inference compute demands rise.