Reflection AI Signs $1.5B Monthly AI Compute Deal with SpaceX

①Artificial intelligence startup Reflection AI has signed a cooperation agreement with SpaceX, under which it will pay SpaceX $150 million per month starting next month.
②Reflection is an AI laboratory focused on open-source models, offering governments and enterprises greater flexibility in AI deployment.
image description
June 23 – According to the latest reports, open-source AI startup Reflection AI has signed a large-scale computing cooperation agreement with SpaceX, becoming the latest external customer to utilize the Colossus infrastructure.


Based on leaked information, Reflection will immediately gain access to SpaceX’s GB300 for training and running advanced models. Under the agreement, from July 1, 2026, through 2029, Reflection will pay SpaceX $150 million per month.


After the initial three-month period, either party may terminate the contract with 90 days’ prior notice. If the agreement is executed through its full term, Reflection’s total payments to SpaceX will amount to approximately $6.3 billion.


This agreement indicates that, following its record-breaking IPO, SpaceX is further leveraging its extensive data center footprint. The Colossus infrastructure was initially built partly to support the AI chatbot Grok and compete with rivals such as ChatGPT.


Now, SpaceX is using this infrastructure to sell computing resources to external AI companies. Previously, SpaceX had already reached computing-related partnerships with Anthropic, Google, and Cursor, and is currently in the process of acquiring Cursor.


Reflection’s addition as a new client for SpaceX also carries distinct strategic significance—at a time when governments and the business community are reassessing their reliance on closed AI systems, Reflection is an AI lab dedicated to open-source models.


Earlier, Anthropic ceased providing services to Fable and Mythos, raising concerns about the risks of relying on closed-model suppliers for critical operations. This incident further strengthens the advantage of open-source model companies, namely that customers should have greater control, with the ability to inspect, customize, and run models independently.


Reflection, recently valued at $25 billion, is capitalizing on this trend to develop open-source models that can compete with the frontier models of OpenAI, Anthropic, and Google, while offering governments and enterprises greater flexibility than closed systems.
A spokesperson for Reflection stated in a statement that “the series of recent events has highlighted the importance of open source to the AI ecosystem. More and more countries and enterprises are recognizing the risks and costs of fully relying on closed models.”



Reflection noted that this agreement will provide it with additional computing power resources to accelerate the realization of what it calls “American open intelligence.”



Although the company has not yet released a frontier open-source model for the public, it has gradually gained recognition in the government and national security sectors. Currently, Reflection is participating in the U.S. Department of Energy’s “Genesis Mission” project and has joined the broader AI initiatives of the U.S. Department of Defense.



For SpaceX, this deal once again demonstrates that in the AI race, computing power itself has become a strategic resource, and access to advanced Nvidia chips remains one of the biggest bottlenecks in training and deploying frontier models.



By opening Colossus to external customers, SpaceX is positioning itself as a cloud service provider and a member of the AI infrastructure industry. At the same time, this provides new economic support for SpaceX’s expanding AI infrastructure strategy.



Currently, investors have been paying close attention to whether SpaceX can further expand its presence in artificial intelligence, data centers, and computing power services beyond its rocket and Starlink businesses.


Leave a Comment

Your email address will not be published. Required fields are marked *