← Back to feed

Reflection AI has signed a $1 billion compute agreement with Nebius, the European AI infrastructure company spun out of Yandex’s international operations. The deal grants Reflection access to Nvidia’s latest chips through Nebius’s GPU cloud infrastructure.

The announcement comes weeks after Reflection’s $6.3 billion, 48-month agreement with SpaceX for access to Colossus 2 in Memphis at $150 million per month. Together, the two agreements put Reflection’s committed compute at approximately $7.3 billion — a total larger than many sovereign AI infrastructure programs and achieved without building a single data centre.

Reflection’s Compute Stack

Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both formerly of Google DeepMind. The startup raised $130 million in its initial round, then secured $2 billion in 2025. Nvidia, Sequoia Capital, and Lightspeed are investors. Valuation at the last round was $8 billion.

The compute strategy is now visible across three providers:

• SpaceX / Colossus 2: Nvidia GB300 chips, $150M/month from July 1, up to $6.3B total
• Nebius: Nvidia latest chips, $1B total value, terms undisclosed
• Total committed: ~$7.3B

Nebius is the infrastructure vehicle created after Yandex sold its Russian operations and restructured as a European AI cloud company headquartered in Amsterdam. It has been expanding GPU capacity through direct Nvidia agreements and positions itself as a compute provider for AI labs that want European-hosted or US-hosted alternatives to the major hyperscalers.

Why the Compute Accumulation Matters

Reflection’s model is open-weight: it trains at frontier scale and releases weights publicly. That is a different business from Anthropic or OpenAI — no API monopoly, no subscription revenue, no pay-per-token model. The revenue model for open-weight labs at this scale is not yet settled, but the compute investment signals Reflection is aiming to train models that compete with closed-source frontier systems.

The amounts also signal training runs of a scale that require tens of thousands of GPUs over extended periods. A single frontier training run for a model in the GPT-5.5 / Fable 5 class costs in the range of $100M to $500M in compute. At $7.3B committed, Reflection has enough committed infrastructure to complete multiple frontier training runs.

The framing from Reflection’s public statements has explicitly tied its open-source positioning to recent US government intervention in AI model access. In June, the Trump administration blocked Anthropic’s Fable 5 and Mythos 5 from non-US users for 19 days, affecting over 200 institutions that had built workflows on Anthropic’s API. Reflection has pointed to that event as evidence of structural risk in closed-model dependency.

The Nebius Angle

Nebius adds geographic and supply diversification to Reflection’s compute. SpaceX’s Colossus 2 is a US facility. Nebius operates data centres across multiple European and US locations. For an open-weight AI company targeting both US and European enterprise customers — including those with data sovereignty requirements — having compute in multiple jurisdictions matters.

Nebius went public on Nasdaq in 2024 and has been building out GPU density aggressively since. Signing Reflection at $1B is one of its largest disclosed compute agreements.

What Open-Weight Frontier Actually Takes

The $7.3B figure puts Reflection in infrastructure territory that was previously only occupied by hyperscalers (Google, Amazon, Microsoft), integrated AI labs with hyperscaler backing (Anthropic via Google/Amazon, OpenAI via Microsoft), and SpaceX. An independent startup accumulating that commitment level before releasing a publicly-competitive frontier model is unusual.

The comparison point is DeepSeek, which trained V4 — a model competitive with GPT-5.5 and Fable 5 — on a fraction of the compute budget its US rivals used. If Reflection achieves similar efficiency with its frontier weights, the infrastructure position becomes very strong. If it does not, $7.3B in committed compute is a liability that requires revenue to service.

Reflection has not published a public model. Its positioning is at the compute accumulation stage.