Bristol Myers Squibb buys Nvidia Vera Rubin SuperPOD

Bristol Myers Squibb has purchased an Nvidia DGX SuperPOD built on the Vera Rubin architecture. The cluster includes eight DGX Vera Rubin NVL72 systems and will be joined with the company’s existing, older SuperPOD in a shared computing environment accessible to research teams worldwide. Financial terms were not disclosed.

Each NVL72 rack-scale system combines Nvidia Vera central processing units and Rubin graphics processing units. BMS intends to use the combined infrastructure to train proprietary models and run predictions across programmes that involve compounds, proteins and other scientific data. The SuperPOD software stack will schedule training, prediction and development workloads across the unified environment to reduce wait times for researchers.

BMS has operated an earlier DGX SuperPOD for about three years and characterised it as two or three generations behind Vera Rubin. The two SuperPODs will run through a common data environment so datasets and model outputs produced at one site can be used by teams at other sites. An example offered by the company describes programme data from Lawrenceville, New Jersey, feeding models used by researchers in San Diego. Nvidia Mission Control will handle cluster provisioning, infrastructure monitoring and workload management.

Company executives report rising demand for compute as larger AI models are deployed across research teams. Erin Davis, vice president of research business insights and technology, noted that the existing infrastructure is operating at capacity because of large-scale predictions involving large molecules and the development of internal foundation models. She added the new system will be available across the research organisation rather than limited to a small group of computational specialists.

Greg Meyers, chief digital and technology officer, highlighted energy efficiency, saying the Vera Rubin cluster will deliver substantially more compute capacity per unit of electricity. BMS and Nvidia estimate the eight-system cluster can provide up to ten times the performance per megawatt compared with the infrastructure it replaces.

BMS applies AI across every small-molecule programme and most large-molecule programmes for target identification, lead optimisation, large-molecule prediction and internal model development. The company uses a Predict First approach that applies model-generated predictions to exclude molecules that do not meet required properties before synthesis, reducing the number of compounds sent for laboratory testing.

Payal Sheth, senior vice president of therapeutic discovery sciences, explained researchers use predictions to prioritise synthesis of molecules that meet multiple property requirements. Robert Plenge, chief research officer, said the added compute will allow scientists to evaluate more potential drug candidates during early stages, giving an example of scaling from about 10 candidates to dozens. Plenge reported that AI tools have shortened the time to identify and produce candidates for clinical testing by 20 to 30 percent so far, with potential to reach about 50 percent in coming years. He cited an experimental sickle cell treatment now in early clinical development that the company believes likely would not have been discovered without AI-supported work.

The Vera Rubin SuperPOD will provide access to Nvidia’s BioNeMo Agent Toolkit, which includes tools for protein-structure prediction, molecular generation, molecular docking, sequence analysis and genomics and can link multiple computational tools within the same workflow. BMS is introducing user-friendly interfaces so some prediction requests can be started with natural-language instructions, reducing the specialist knowledge required to launch complex computing tasks. Company officials emphasise that human researchers will continue to review model outputs and decide which compounds or programmes should advance.

BMS said it will allocate the expanded computing capacity across small- and large-molecule design, clinical research and digital-twin applications but did not specify how capacity will be split or provide a deployment date or hosting location for the new cluster. The combined SuperPOD environment aims to give more scientists direct access to high-performance computing resources and to retain institutional knowledge from experiments, clinical readouts and research partnerships across the organisation.