Bristol Myers Squibb Deploys Second AI Supercomputer for Drug Discovery

Bristol Myers Squibb is expanding its computational infrastructure with a second NVIDIA DGX SuperPOD, this one built on eight DGX Vera Rubin NVL72 systems that the company says will be the most powerful and energy-efficient AI cluster in life sciences.

The deployment marks a strategic shift from rationed supercomputing access to what Erin Davis, vice president of research business insights and technology at BMS, calls "Limitless Compute." Rather than restricting the resource to a small group, the new system will be available to every scientist across the company's global research organization.

Why it matters

Pharmaceutical companies face a fundamental bottleneck: the time and cost required to identify promising drug candidates, optimize their properties, and move them through development. BMS's approach—combining massive compute capacity with unified data access and AI-native tooling—represents a bet that removing infrastructure constraints will accelerate the entire discovery pipeline. If successful, the model could reshape how large research organizations deploy AI at scale.

Performance and capacity gains

The eight rack-scale systems, each comprising NVIDIA Vera CPUs and Rubin GPUs, deliver up to 10 times the performance per megawatt compared to the infrastructure they replace. Researchers will access a unified AI platform that includes NVIDIA BioNeMo Agent Toolkit for biological AI, supporting predictions, model training, and agentic workflows across drug discovery stages.

BMS has operated its first DGX SuperPOD for approximately three years, and Davis reports the system is now saturated. The company is running large-scale predictions for large molecules, building proprietary foundational models, and supporting production workloads that require substantial GPU resources.

Proven results from existing infrastructure

The current system has already produced measurable outcomes. AI-enabled target identification saves scientists weeks of manual work. The company has used AI to expand its library of CELMoD compounds—molecules engineered to selectively degrade cancer-causing proteins—opening pathways to new targets across a wider range of diseases.

Payal Sheth, senior vice president of therapeutic discovery sciences at BMS, describes a "Predict First" methodology in lead optimization that uses AI predictions to prioritize which molecules to synthesize. This approach ensures laboratory experiments focus on compounds with the highest probability of meeting required property profiles.

Unified data and agentic workflows

Davis's team is combining both SuperPODs into a single environment with a unified data plane accessible from every BMS site globally. The system will be managed through NVIDIA Mission Control, enabling researchers to initiate complex predictions using natural language rather than requiring deep computational expertise.

Sheth emphasizes the importance of institutional learning: datasets from one program location can feed models that teams elsewhere draw upon. Agentic workflows will enable learning across traditional organizational silos and research programs.

"When you as a scientist can go to an army of well-vetted, fully trained virtual scientists that have BMS knowledge baked in, now you're a whole team in and of yourself," Davis said.

The new system already has detailed allocation plans mapped across modalities, from small and large molecule design to clinical applications and digital twins. Davis notes the expansion wasn't about scale for its own sake—the compute capacity is tied to specific research applications at every stage of the pipeline.

These details were first reported by NVIDIA.