How they use Nebius AI Cloud

Prima Mente uses Nebius GPUs to train, fine-tune and serve Pleiades biological foundation models for brain-disease research. A representative described an allocated 32-node cluster in a customer interview. Lead AI engineer Luca Giacomoni says Nebius continually improves the orchestration layer, shortening training cycles and letting the team focus on experiment design. Prima Mente said at Inflection 2026 that Nebius helped scale its models in weeks. NVIDIA describes Nebius Managed Kubernetes and distributed file storage supporting multi-host training, shared datasets and checkpoints. Nebius architects helped configure workloads and debug inter-node communication. Ravi Solanki also describes hands-on support for model architectures and training. NVIDIA describes the SANDBOX clinical pilot across 15 UK sites. Nebius says Pleiades is deployed in a live UK clinical study to classify and stratify patients with Alzheimer’s and Parkinson’s.

Prima Mente reports training Pleiades through NVIDIA DGX Cloud Lepton on Nebius Cloud using 256 H200 GPUs and 1.9 trillion methylation tokens. It later reported 12% higher pre-training throughput on a dedicated H200 cluster, reaching 1.19 million tokens per second across 16 nodes, and scaling an ablation model from 1B to 10B parameters in weeks. Nebius reports per-device throughput increased from 8,000 to 9,000 tokens per second and says the team expects throughput to double when that ablation model reaches the full cluster. Nebius’s Q2 2026 shareholder letter separately reports scaling Prima Mente’s epigenetic foundation model from 1B to 68B parameters on a dedicated cluster. Its customer story describes work with more than 60 trillion tokens and plans to scale Pleiades beyond 100B parameters on NVIDIA Blackwell while doubling reserved Nebius capacity. A Nebius representative describes consolidating neurodegenerative disease data from hundreds of sites across Europe, the US and the Middle East, with reserved Nebius GPUs supporting projects with groups in Italy, England and San Francisco.

Prima Mente also uses Nebius in Prima Daedalus, which turns published biological experiments into data and evaluations for its models. Its September 2026 paper describes an optional model-evaluation step that submits GPU inference jobs to a Nebius cluster after compute approval. The paper credits Nebius, Union AI and Seqera for backend infrastructure.