Reflection AI, an American startup developing open models, has signed a one-billion-dollar agreement with Nebius for access to computing capacity. Nebius, the former international arm of Yandex, will provide Reflection with Nvidia’s latest chips.

The deal comes weeks after a similar arrangement with SpaceX for computing resources, and fits into the race among AI companies to secure infrastructure for training and running models. Reflection is one of the open-weight model startups that has drawn significant attention in recent weeks, at a time when debate is growing over the value of proprietary closed models—especially after the Trump administration restricted access to Anthropic and OpenAI’s most powerful models, and following the release of more capable open models from China.

Why it matters to you. When a young startup locks in a billion dollars in infrastructure, it signals that capital is betting on open models and independent training. If you’re working with models in production, this reduces the risk of losing access to frontier models overnight. If you’re building around open models, you see the ecosystem receiving serious resources.

In detail

The context. Reflection AI was founded in 2024 by two former Google DeepMind researchers and has already raised nearly $2.6 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners. The startup is valued at eight billion. The billion-dollar deal with Nebius comes after Nvidia invested two billion in the startup.

Nebius, for its part, is building an enterprise customer portfolio: shortly after the Nvidia investment, it signed a five-year agreement with Meta worth up to $27 billion, and last year closed a deal with Microsoft worth up to $19.4 billion over multiple years.

What “open models” means. Reflection develops open-weight models: the weights are publicly available, so anyone can download, modify, and run the model on their own infrastructure. This differs from the closed models of OpenAI and Anthropic, where you access only via API and don’t see the weights. Interest in open models has grown in recent weeks after the Trump administration pressured Anthropic and OpenAI to limit their most powerful models, raising concerns that access could be revoked overnight.

The timing. The deal comes weeks after a similar arrangement with SpaceX for computing resources. The pattern is clear: open-model startups are locking in guaranteed computing capacity for years, not months. This suggests that training large, competitive models requires enterprise-scale commitments, not experiments on shared cloud.

The figures in perspective. A billion in compute isn’t symbolic: for comparison, training GPT-4 cost (by external estimates) between $50 and $100 million. A billion-dollar agreement therefore covers many training cycles on large models, or continuous training at scale.

What remains to be seen. TechCrunch contacted Reflection and Nebius for details on the contract (duration, resource allocation, milestones), but as of publication there are no public confirmations beyond the deal announcement.