Nebius Sells $1B in AI Compute to Reflection

The scramble to lock down AI computing capacity just produced another billion-dollar deal, and it's a useful window into how AI infrastructure economics are shifting. Cloud provider Nebius Group agreed to sell more than $1 billion worth of computing power to Reflection AI, a model developer founded by two former Google DeepMind researchers, according to Bloomberg's report. The agreement runs through 2029 and gives Reflection access to Nvidia's latest GB300 chips.

Reflection's CTO and co-founder Ioannis Antonoglou framed the deal around a specific market gap. "The need for open models is clear, and this additional compute capacity will allow Reflection to continue to build and train frontier AI models at scale," he said in a statement reported by Reuters. Reflection develops open-source AI models positioned as an alternative to closed offerings from OpenAI and Anthropic, a category that's drawn growing enterprise interest as rising AI bills push businesses to look for cheaper, more customizable options.

This deal builds directly on Reflection's June agreement with SpaceX, which reportedly provides the startup with around $150 million in monthly computing capacity through 2029. Reflection closed a roughly $2 billion fundraise in late 2025 at an $8 billion valuation, backed by Nvidia, Sequoia, and Lightspeed, and according to Ad-Hoc News's reporting, is now reportedly negotiating a $2.5 billion round at a $25 billion valuation, a striking escalation in less than a year.

For Nebius, the Reflection contract is another brick in an increasingly large wall. The company sealed a $27 billion, five-year deal with Meta in March, beginning in 2027, followed by a $17.4 billion contract with Microsoft, with Nvidia itself taking a $2 billion equity stake in the company. Nebius's total contracted backlog now stands at roughly $50 billion, and the company is investing $20 to $25 billion in 2026 capital expenditure alone, targeting 4 gigawatts of computing capacity by year-end, including a new 1.2-gigawatt site in Pennsylvania.

Here's the part of this story that matters most for understanding where AI infrastructure sentiment actually stands. Despite the blockbuster Reflection contract, Nebius shares moved on investor anxiety rather than celebration. Short interest in the stock swelled from 20.13% of the float in mid-June to 28.55%, driven partly by reports that Meta, one of Nebius's largest customers at $27 billion, is exploring building its own AI computing capacity, a prospect that would turn a major customer into a direct competitor.

Not every analyst sees that threat as real. Analysts at SemiAnalysis argued the market may be overreacting, suggesting Meta's growing compute hunger could actually create more opportunities for hyperscalers like Nebius rather than fewer, a dynamic similar to what we've seen play out in Foxconn's AI server business as demand keeps outpacing even aggressive capacity expansion plans.

I've advised companies on AI infrastructure decisions for four years, and this deal captures something important about the current market. AI compute demand remains genuinely intense, evidenced by billion-dollar contracts getting signed regularly, but investor confidence in the sustainability of that demand is increasingly shaky. Those two things are happening simultaneously, and businesses evaluating cloud AI vendors need to watch both signals rather than just the headline deal size.

For companies relying on providers like Nebius for AI compute, this is worth understanding as a broader indicator of the generative AI market at large. Booked contracts reflect real, committed demand, but stock market skepticism reflects genuine uncertainty about whether that demand curve holds steady over multi-year horizons.

Watch whether Meta actually moves forward with building independent AI computing capacity, and how that would affect its existing $27 billion Nebius commitment if it does. Also watch whether Reflection's reported $25 billion valuation round closes at that level, since a successful raise at that size would validate the aggressive compute-locking strategy startups are increasingly using to compete with better-funded labs.