A billion dollars buys many things in the AI business. For Reflection AI, it buys time on someone else’s chips. The startup, founded in 2024 by two former Google DeepMind researchers, has signed a computing deal worth more than $1 billion with infrastructure provider Nebius, securing access to Nvidia’s newest GB300 processors on a contract that runs through 2029.
The agreement, announced on July 14, is Reflection’s second major compute arrangement in a matter of weeks. In June the company reached a separate deal to tap computing resources from SpaceX. Taken together, the contracts show a young company spending at a scale that would have seemed absurd three years ago, and doing it before its most ambitious models have even shipped.
The open-weight bet
Reflection is building open-weight models, systems whose underlying parameters are published so that companies can download them, customize them, and run them on their own hardware. That approach puts the startup in direct opposition to OpenAI and Anthropic, which sell access to their models through subscriptions and APIs while keeping the weights themselves locked away.
The pitch is straightforward. Plenty of businesses are uneasy about depending on a handful of closed-model providers, especially as API prices shift and governments impose new restrictions on who can use what. A model you can inspect, host, and modify yourself removes some of that risk. Investors have bought the argument. Reflection is reportedly valued at around $8 billion, a striking number for a company that did not exist before 2024.
Compute is the new moat
Why does an open-model startup need a billion dollars of processing power? Because training frontier-scale systems requires vast clusters of specialized chips running for months, and demand for that hardware still comfortably outstrips supply. Research talent used to be the scarce resource in AI. Now it is electricity, data center space, and above all access to Nvidia’s latest silicon.
Nebius, which grew out of the international operations of Russian tech group Yandex, has built its entire business on that shortage. It buys and operates GPU capacity at scale, then rents it to AI companies that cannot wait years to construct their own data centers. The GB300 chips covered by this agreement sit at the top of Nvidia’s current lineup and are designed for exactly this kind of work: training and serving the largest models in existence.
The structure of the deal says something about where the industry is heading. AI startups can no longer compete on clever research alone. They need long-term contracts covering chips, power, networking, and floor space, the kind of commitments that used to be the preserve of hyperscale cloud providers. Reflection’s SpaceX arrangement and now the Nebius contract lock in that capacity years ahead of need, insulating the company from a market where GPU queues are measured in quarters, not weeks.
A crowded race with real stakes
Reflection is not alone in betting on openness. Meta has released successive generations of open-weight models, and labs in China have shipped capable open systems that businesses around the world now run in production. The competitive question is whether a well-funded American startup can produce open models good enough to pull serious enterprise workloads away from the closed incumbents.
There are reasons for skepticism. Training frontier models is punishingly expensive, and giving the results away narrows the obvious revenue paths. Open-model companies typically earn through enterprise support, hosted services, and custom work, businesses that are harder to scale than a metered API. Reflection will need to prove its economics work at the same time as it proves its research does.
Still, the timing may favor it. Enterprises have grown more sophisticated about AI procurement, and many now split workloads between closed APIs and self-hosted open models depending on cost and sensitivity. Every price increase from a closed lab makes the open alternative look a little better, and every new restriction on model access strengthens the case for weights you control.
The next test is delivery. Reflection now has the chips, the capital, and the valuation of a major lab. What it does not yet have is a flagship model that shifts the leaderboard, and contracts of this size have a way of turning into deadlines. For more coverage of machine learning and the companies building it, visit Mylistingo.
