Reflection AI's $6.3B SpaceX Deal Powers Open-Source Frontier AI

Reflection AI Activates $6.3 Billion Compute Lease on 1 July 2026

Reflection AI, an open-source AI laboratory founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, began paying SpaceX $150 million per month for access to Nvidia GB300 computing infrastructure at SpaceX's Colossus 2 data centre in Memphis, Tennessee, starting 1 July 2026. The arrangement is governed by a compute capacity agreement signed on 22 June 2026, with total payments reaching approximately $6.3 billion if the deal runs to its 2029 end date. The deal positions Reflection AI as the highest-spending independent compute customer in the open-source frontier AI market and places SpaceX at the centre of the AI infrastructure supply chain alongside hyperscalers.

Who Is Reflection AI

Reflection AI was founded in 2024 by Misha Laskin, who led reward modelling on Google DeepMind's Gemini project, and Ioannis Antonoglou, who was DeepMind's sixth researcher and a co-creator of AlphaGo. The company's strategic premise is that the United States needs an open-source frontier AI counterpart to DeepSeek — a publicly accessible, high-capability model from a Western organisation whose weights are available for inspection, fine-tuning, and deployment by governments, enterprises, and researchers without dependence on a single commercial provider.

Reflection raised $2 billion in October 2025 at an $8 billion valuation, with Nvidia, Eric Schmidt, Citigroup, Lightspeed Venture Partners, and Sequoia Capital among the investors. Nvidia subsequently made an additional $800 million strategic investment in the company, reflecting its interest in backing the labs that will consume the most compute in the coming years. Reflection's valuation reached approximately $25 billion by mid-2026 — a three-fold increase from the October 2025 round in eight months.

SpaceX's Colossus 2 and the Commercial Compute Platform

The Colossus data centre was originally built for xAI, Elon Musk's AI company, at SpaceX's Memphis, Tennessee site. SpaceX subsequently began offering compute capacity on Colossus infrastructure to external customers. As of June 2026, SpaceX has landed compute agreements with Anthropic, Google, and Cursor — the AI coding tool that SpaceX acquired earlier in 2026 — alongside the new Reflection AI deal. Committed external revenues from Colossus have exceeded $80 billion through 2029 across its customer portfolio, making Colossus one of the largest third-party AI compute platforms in the world by revenue, rivalling the GPU cloud offerings from Microsoft Azure, Google Cloud, and Amazon Web Services.

The Reflection AI deal provides immediate access to Nvidia GB300 chips, the most capable inference and training accelerators currently available from Nvidia, enabling Reflection to operate training runs at a scale previously accessible only to the largest closed-source AI labs. The $150 million monthly cost implies Reflection is operating compute at a scale comparable to what Anthropic and OpenAI consumed during training of their current flagship models.

The Open Frontier AI Thesis

Reflection AI's stated goal is to release the weights of its frontier models publicly while keeping its datasets and full training pipelines proprietary — a model that mirrors DeepSeek's approach of releasing V3 and V4 model weights publicly while retaining internal infrastructure as proprietary. Misha Laskin has stated that releasing model weights enables governments, enterprises, and research institutions to inspect, audit, and adapt AI systems without depending on a single commercial provider — an argument that has resonated with US government customers concerned about supply chain concentration in closed-source AI.

Reflection has government contracts with the Department of Energy's Genesis Mission and has been part of broader Pentagon AI programmes. Securing government partnerships before a public frontier model release gives Reflection a revenue base and mission credibility that distinguishes it from pure commercial open-source AI labs focused primarily on enterprise API access.

The Scale-First Approach to Open-Source Frontier AI

Reflection has not yet publicly released a frontier model on the scale of its Colossus compute investment. The company previously released Reflection 70B, a fine-tuned model that topped open-source leaderboards on its release. The $6.3 billion compute commitment signals that Reflection's next public release will be trained at a substantially larger scale — likely in the hundreds of billions to trillions of parameters — with the Colossus GB300 cluster as the primary training infrastructure.

The timing of the Reflection-SpaceX deal relative to the broader open-source AI landscape is significant. DeepSeek V4 is completing its mid-July 2026 official production release, and Meta has been releasing Llama model weights publicly across multiple generations. Reflection is positioning itself as the US-based alternative with open weights, frontier-scale training compute, and alignment with US government national security interests in a way that Chinese-origin open-source models are not structured to provide.

What Reflection AI's Trajectory Means for Enterprise and Indian Tech Teams

For enterprise technology teams evaluating open-source AI infrastructure, Reflection AI's compute commitment is a forward signal about where high-quality, permissively licensed frontier model weights will come from in the next two to three years. Teams currently relying on DeepSeek V4 for production workloads — including Indian development teams drawn to its low API cost and open-weight availability — will have Reflection's frontier models as an alternative once they ship.

For Indian government and regulated-sector technology teams, open-weight frontier models from a US-origin lab with Pentagon-verified safety credentials represent a procurement-eligible alternative to both closed-source models and Chinese-origin open-weight models. As RBI and SEBI advance their AI risk frameworks for financial services applications, the provenance and auditability of model weights will matter more in regulated-sector deployments, making Reflection's open-weight thesis directly relevant to Indian fintech, insurtech, and banking technology builders considering local or sovereign model deployment.

The Bottom Line

Reflection AI began paying SpaceX $150 million per month on 1 July 2026 for Nvidia GB300 compute access at the Colossus 2 data centre in Memphis — a $6.3 billion commitment through 2029. The company, founded by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, is valued at $25 billion and backed by Nvidia, Sequoia Capital, Lightspeed, and Eric Schmidt. SpaceX's Colossus platform now holds over $80 billion in committed compute revenues from Anthropic, Google, Cursor, and Reflection. Reflection's thesis — US-origin, open-weight frontier AI for government and enterprise — positions it as the Western counterpart to DeepSeek, with a public frontier model expected at scale from this compute investment. For Indian enterprise teams evaluating open-source frontier AI, Reflection's trajectory is the key development to track in the second half of 2026.