Reflection AI secures $1B Nebius compute deal for frontier models

Locking in more than $1 billion of high‑end Nvidia GB300 capacity puts Reflection squarely in the top tier of labs racing to train frontier‑scale models. Compute access has become the chokepoint of modern AI, and this kind of long‑dated contract effectively reserves a slice of future GPU supply in a market where even hyperscalers struggle to get enough hardware. By pairing this Nebius deal with an earlier multi‑billion‑dollar agreement with SpaceX, Reflection is building a diversified infrastructure stack outside the usual US hyperscalers.

Strategically, the move underscores two big trends. First, open‑model labs are no longer small players scraping for leftover GPUs; they are raising capital and signing infrastructure deals at a scale comparable to closed‑model incumbents. Second, cloud providers like Nebius are emerging as serious alternatives to US giants, especially for customers who worry about export controls or unilateral access restrictions after the recent Anthropic curbs. That diversification of both labs and infrastructure makes the race to AGI more multipolar, with more actors capable of training truly frontier systems.

Competitively, this raises the bar for any would‑be open‑source challenger: without similar multi‑year compute lock‑ins, they risk falling a full hardware generation behind. At the same time, it intensifies pressure on Nvidia and its rivals to keep pushing out ever larger, more power‑hungry accelerator families to satisfy overlapping demands from tightly contracted customers.