BMS turns eight Vera Rubin racks into a drug-discovery AI factory

AI compute in drug discovery is moving from a specialist resource to production infrastructure for everyday research. Bristol Myers Squibb is deploying a second NVIDIA DGX SuperPOD built on Vera Rubin systems, a cluster the company internally calls the “SuperDuperPOD.” According to NVIDIA’s official case study, the system uses eight DGX Vera Rubin NVL72 rack-scale systems and will sit alongside BMS’s existing DGX SuperPOD.

The key figure is infrastructure efficiency. Each rack-scale system combines NVIDIA Vera CPUs and Rubin GPUs, and NVIDIA says the new setup delivers up to 10x the performance per megawatt of the infrastructure it replaces. BMS plans to combine the new system and the existing SuperPOD into one global data plane, giving scientists across company sites access to a unified AI platform that includes the NVIDIA BioNeMo Agent Toolkit for biological AI workloads.

The deployment follows three years of internal AI use rather than a speculative pilot. BMS says AI-enabled target identification already saves scientists weeks of manual work. The company has also used AI to expand its library of CELMoD compounds, support lead optimization through a “Predict First” method, run large-scale predictions for large molecules, and build its own foundation models. Those workloads explain why the research group says its current GPU capacity is saturated.

The organizational change may matter as much as the hardware. BMS wants researchers to start complex predictions in plain English instead of routing every job through a narrow group of computational specialists. The company also plans to remove site-specific access restrictions left from earlier acquisitions, so data from one program can feed models and decisions across other therapeutic programs and research locations.

The careful read is that this is not an instant cure machine. It is an attempt to make every experiment, clinical readout, and partnership compound into reusable institutional learning. The metrics to watch are clinical-candidate speed, failed-experiment reduction, model reuse across programs, and whether giving every scientist access to the platform changes R&D throughput rather than just GPU consumption.