Bristol Myers Squibb is expanding its computational research capabilities by deploying eight NVIDIA DGX Vera Rubin NVL72 systems. This new infrastructure, which the company refers to as its second DGX SuperPOD, is designed to support large-scale AI models and agentic workflows across the organization's global drug discovery pipeline. By integrating these systems, the company aims to provide researchers with broader access to high-performance computing resources, moving away from previous limitations that restricted access to smaller groups of specialists.
The new deployment utilizes NVIDIA Vera CPUs and Rubin GPUs, which the company reports will deliver up to 10 times the performance per megawatt compared to its existing infrastructure. This efficiency gain is intended to support the company's ongoing efforts in target identification and lead optimization. By utilizing predictive modeling, researchers can prioritize the synthesis of molecules with the highest probability of success, effectively filtering out candidates that do not meet necessary property requirements before they reach the laboratory stage.
The integration of these systems into a unified data environment is a central component of the company's strategy to institutionalize research findings. By creating a single data plane accessible from all global sites, the company intends to ensure that datasets and model learnings are shared across teams in locations such as Lawrenceville, New Jersey, and San Diego, California. This approach is designed to replace site-specific restrictions and technical barriers with AI-native tooling managed through NVIDIA Mission Control, allowing scientists to initiate complex predictions using natural language.
The company has already utilized AI to expand its library of CELMoD compounds, which are engineered to selectively degrade proteins associated with cancer. These computational methods have allowed researchers to explore larger chemical spaces and identify new potential medicines for a variety of diseases, including those in the field of brain health. The expanded compute capacity is expected to further support these efforts by enabling the training of proprietary foundational models and the execution of large-scale molecular predictions.
Beyond raw processing power, the deployment focuses on enabling agentic workflows that can operate across different research silos. By providing scientists with access to virtual agents trained on internal knowledge, the company aims to create a cumulative learning loop where every experiment and clinical readout contributes to a broader intelligence framework. This architecture is intended to help researchers focus on scientific decision-making rather than the logistics of resource allocation, ultimately aiming to compress discovery timelines for new therapeutic treatments.