Bristol Myers Squibb is deploying its second NVIDIA DGX SuperPOD, built on the upcoming NVIDIA Vera Rubin platform, to create what it calls the life science industry's most advanced AI factory. The company already runs one of the largest AI clusters in life sciences and is doubling down to accelerate drug discovery and development.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Bristol Myers Squibb announces deployment of its second DGX SuperPOD based on NVIDIA Vera Rubin, aiming to open AI compute to every scientist and achieve prediction and agentic workflows across the entire drug discovery process.
Bristol Myers Squibb (BMS) VP Erin Davis jokingly calls the new system 'SuperDuperPOD.' BMS has been running one DGX SuperPOD for about three years with tangible results, but Davis says current compute is saturated: 'We're running large-scale macromolecule predictions in production and building our own foundation models, which require a lot of GPUs.'
The newly deployed second DGX SuperPOD is based on 8 NVIDIA DGX Vera Rubin NVL72 systems, each containing Vera CPUs and Rubin GPUs, delivering 10x performance-per-watt improvement over existing infrastructure. BMS plans to integrate old and new clusters into a unified environment managed by NVIDIA Mission Control, allowing researchers to initiate complex predictions using natural language.
BMS SVP Payal Sheth notes that AI has moved from abstract concept to measurable impact. AI-assisted target identification saves scientists weeks of manual effort; the team used AI to expand the CELMoD compound library—molecules that selectively degrade cancer-causing proteins, applied in blood cancer treatments—and opened new possibilities for new targets and broader diseases.
In lead optimization, BMS adopts a 'predict-first' approach: predicting which molecules to synthesize first via multi-parameter optimization, filtering out candidates that don't meet property requirements, ensuring precious lab experiments focus on molecules with the highest probability of success. Sheth emphasizes: 'This ensures that precious lab experiments are aligned with advancing molecules with the highest probability of success.'
Sheth says that in the past, each project was treated independently, and learnings couldn't accumulate into an intelligent framework. Now, BMS uses AI to turn every experiment, clinical data, and partnership into higher-confidence scientific decisions. Davis adds that agents 'don't care about boundaries; they cross all domains,' enabling teams to learn from cross-project and cross-department decisions.
Davis describes: 'When scientists have a well-validated, well-trained, BMS-embedded virtual scientist army at their fingertips, you yourself become a complete team.' Sheth emphasizes that human intuition won't be replaced but enhanced by more quantitative insights and predictions.
The new system already has a detailed resource allocation plan covering modalities like small molecules, large molecule design, clinical applications, and digital twins. Davis says: 'We didn't just buy the biggest compute and call it done; SuperDuperPOD will run through every node of R&D.' When asked by Chief Digital and Technology Officer Greg Meyers whether they can saturate the new system, Davis replied: 'Give us time.'
Davis's father passed away from Alzheimer's five years ago, and she understands the difficulty in brain diseases. BMS has significant investments in brain health, and Davis hopes AI can accelerate the discovery of symptom-relief drugs. She believes the bottleneck is not technology itself but how to get scientists to actually use it and learn from it.
This article's information primarily comes from an NVIDIA official blog interview with BMS executives, a primary source. Specific AI achievements (e.g., saving weeks of time, expanding CELMoD library) are statements from BMS executives without independent verification. The 10x performance-per-watt improvement is an NVIDIA product claim.
By deploying its second DGX SuperPOD, Bristol Myers Squibb is transforming AI from a tool for a few experts into infrastructure for all scientists, aiming to make the entire drug discovery process predictive and intelligent. The core challenge lies in effectively utilizing 'infinite compute' and translating it into measurable drug development outcomes.
Primary report
Primary source