SpaceX Inks $6.3 Billion AI Compute Deal with Open-Source Startup Reflection

SpaceX has signed a landmark compute infrastructure agreement with open-source artificial intelligence startup Reflection AI, a deal that could be worth as much as $6.3 billion over its full term, according to a report by CNBC.

Under the terms of the agreement, Reflection will gain immediate access to Nvidia’s GB300 AI chips, the latest-generation accelerators designed for training and running large-scale advanced models. The startup has committed to paying SpaceX $150 million per month starting July 1, 2026, through the end of 2029. If the contract runs to completion, total payments will reach approximately $6.3 billion.

The deal includes an exit clause that allows either party to terminate the partnership with 90 days’ notice, provided the agreement has been in effect for at least three months.

SpaceX is no stranger to the AI compute business. The company has already inked similar arrangements with Anthropic, Google, and Cursor — the latter of which is currently being acquired by Elon Musk’s xAI. Adding Reflection to the roster broadens SpaceX’s customer base, but more importantly, it brings a strategically distinct kind of partner into the fold: a firm committed to open-source AI development at a moment when governments and enterprises worldwide are reassessing their reliance on closed, proprietary AI systems.

Reflection AI has not yet publicly released a frontier-class open-source model, but its credentials are already drawing attention. The startup is collaborating with the U.S. Department of Energy on the Genesis Mission and is involved in broader AI initiatives at the Pentagon, signaling that its technology carries weight in national-security and scientific computing circles.

The partnership underscores the escalating demand for high-end AI infrastructure and the emergence of a new class of compute brokers — companies like SpaceX that sit between chip manufacturers and the wave of AI startups racing to build the next generation of models.