How they use Nebius

Converge Bio trained Converge-SC on Nebius infrastructure using more than 36 million single cells and approximately 2TB of transcriptomic data. Nebius says training required 7,000 NVIDIA H100 GPU hours accessed through distributed clusters, NVIDIA CUDA and Nebius AI Studio. Its healthcare solutions page specifies H100 80GB GPUs.

The model reads gene-expression patterns for biological research and drug discovery. It processes more than 20,000 genes per cell using a 30K context length, preserves raw numerical expression values and supports patient-level analysis. Nebius says Converge-SC was built for interpretability and usability, with a goal of delivering explainable, patient-specific insights for drug discovery and precision medicine. Nebius reports that Converge-SC outperformed baseline models on disease classification and drug perturbation tasks. Nebius also says the model was released via Hugging Face. Oded Kalev also described building the single-cell model on Nebius’s AI platform. In a BioPharma APAC interview, Nebius’s Ilya Burkov said Converge Bio used its AI platform to analyze 36 million single cells in days rather than months.

Converge Bio’s co-founder identified Nebius on-demand GPUs as part of the company’s ML-first research stack built during 2024–2025. He described the GPUs as providing elasticity while the team developed production-grade biological model pipelines.