Reflection AI Locks In $1 Billion Compute Deal With Nebius Through 2029

Reflection AI has signed a multi-year compute agreement worth more than $1 billion with Amsterdam-based cloud provider Nebius Group NV. The deal, announced July 14, 2026, grants the U.S. open-source AI startup access to Nvidia's latest GB300 chips through 2029.

The agreement marks Reflection's second major infrastructure deal in recent weeks. Last month, the company secured a separate compute agreement with SpaceX reportedly worth $150 million per month through 2029, totaling approximately $6.3 billion.

Racing for Hardware

The back-to-back agreements underscore the intense competition for AI computing capacity as startups race to secure the hardware needed to train frontier models. Reflection, founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, has positioned itself as an open-source alternative to closed systems like OpenAI and Anthropic.

The company is currently valued at approximately $8 billion and has raised close to $2.6 billion from investors including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, according to reporting from The Next Web and TechCrunch.

Nebius Emerges as Infrastructure Power

Nebius, which split from Russian tech giant Yandex in 2024, has rapidly emerged as a critical infrastructure hub for major AI players. The Amsterdam-based "neocloud" provider has secured massive agreements with Meta worth up to $27 billion and Microsoft valued at up to $19.4 billion.

The rise of neoclouds like Nebius reflects the growing capital requirements for training frontier AI models. These specialized providers offer an alternative to traditional hyperscalers, though they remain dependent on Nvidia's proprietary chip architecture.

Open Models, Closed Hardware

Reflection's strategy highlights a central paradox in the open-source AI movement: while the company advocates for open-weight models as an alternative to proprietary systems, its infrastructure relies entirely on expensive, rented proprietary hardware.

This tension has gained strategic urgency following mid-2026 reports that the Trump administration pressured Anthropic and OpenAI to restrict access to certain powerful models. The government intervention has accelerated interest in open-source alternatives, though dependence on Nvidia's supply chain creates its own vulnerabilities.

Antonoglou stated in a company announcement that the additional compute capacity will enable Reflection to continue building and training frontier AI models at scale.

The Compute Bottleneck

Industry analysts view these massive capacity deals as essential for any company attempting to compete at the frontier of AI development. Training runs for state-of-the-art models now require tens of thousands of advanced GPUs running for months, making access to reliable hardware a make-or-break factor.

Reflection's aggressive infrastructure buildup suggests the company is preparing for multiple large-scale training runs through the end of the decade. Whether open-source models can match closed systems while relying on the same proprietary hardware remains an open question.