Reflection AI signed a $1 billion compute deal with Nebius, giving the open-weight model startup access to Nvidia's latest chips through the European infrastructure provider. The agreement lands just weeks after Reflection struck a similar deal for SpaceX computing resources, and slots into a broader scramble by AI developers to lock in training capacity years in advance. Reflection is currently valued at $8 billion and has raised close to $2.6 billion since its 2024 founding.
Nebius, formerly the international arm of Russian tech giant Yandex, has emerged as one of the more aggressive neocloud plays in Europe. The company signed a five-year infrastructure deal with Meta worth up to $27 billion, and last year inked a multi-year deal with Microsoft worth up to $19.4 billion. Nvidia put $2 billion into Nebius directly, tightening the loop between chipmaker, cloud, and model developer.
Reflection was founded in 2024 by two former Google DeepMind researchers and has drawn backing from Nvidia, Sequoia Capital, and Lightspeed Venture Partners. The company is building open-weight models — a category that has picked up momentum as Chinese labs release increasingly capable systems and as customers reassess dependence on closed APIs.
Key facts
- 01Reflection AI signed a $1 billion compute deal with Nebius for access to Nvidia's latest chips.
- 02Reflection is valued at $8 billion and has raised close to $2.6 billion since its 2024 founding.
- 03The deal follows a similar Reflection pact for SpaceX compute signed a few weeks earlier.
- 04Nebius previously locked in a $27 billion five-year deal with Meta and a $19.4 billion multi-year deal with Microsoft.
- 05Nvidia backs both sides — it put $2 billion into Nebius and is an investor in Reflection.
The timing is not incidental. Last month, the Trump administration pressured Anthropic and OpenAI to restrict their most powerful new models, a move that fueled concerns among enterprise buyers that access to closed models could be curtailed on short notice. Open-weight developers pitch a simpler value proposition: download the weights, run them where you want, and the vendor cannot pull the plug.
The Nebius partnership also signals how compute contracts are becoming the real currency of the AI race. A model developer's ability to train and serve depends less on headcount and more on how many GPU-hours it has already reserved. Reflection's back-to-back deals with SpaceX and Nebius, both worth hundreds of millions or more, give it a diversified supply base at a moment when hyperscalers are absorbing most of Nvidia's Blackwell allocation.
For Nebius, the Reflection contract is a smaller but strategically useful counterweight to its Meta and Microsoft anchors. Serving a fast-growing open-model lab gives Nebius a foothold with the segment of the market least tied to any single hyperscaler cloud, and it validates the company's pitch to other independent labs shopping for capacity.
The open questions are execution and price. A $1 billion compute commitment implies a delivery schedule that has to survive Nvidia's supply constraints, and Reflection has not disclosed the timeframe. Nor has either company detailed what generation of Nvidia silicon underpins the deal, though Nebius's recent buildouts have centered on the latest Blackwell-class systems. Reflection has not publicly shared a benchmark result for a flagship model that would justify the scale of its infrastructure spend.
The bigger read is that the frontier is no longer a single race between a handful of closed labs. Reflection is spending like a frontier player, sourcing compute like one, and doing it in service of open weights — a bet that the customers Anthropic and OpenAI cannot fully serve, whether for data-residency, sovereignty, or continuity reasons, add up to a large enough market to fund an $8 billion company. If that bet lands, the more interesting AI infrastructure story of the next year may not be who trains the biggest closed model, but who supplies the compute for everyone building around them.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
