Bristol Myers Squibb is building a second NVIDIA DGX SuperPOD using eight Vera Rubin NVL72 systems, expanding the computing capacity available to its global drug-discovery organisation. The pharmaceutical company plans to combine the new installation with its existing SuperPOD as one environment accessible from research sites around the world.
NVIDIA says the new cluster will support predictions, model training and agentic workflows across the drug-discovery pipeline. It will also include BioNeMo Agent Toolkit, NVIDIA’s collection of tools for biological AI work.
A second SuperPOD for wider access
Bristol Myers Squibb has operated its first DGX SuperPOD for about three years. According to the company, that system has helped researchers use AI for target identification and prioritise experimental work, including research on molecules designed to degrade cancer-causing proteins.
The second deployment is intended to remove access constraints. Rather than limiting supercomputing resources to a specialist group, Bristol Myers Squibb says it wants every scientist to be able to use the environment. The company’s existing cluster is already saturated by production predictions and foundation-model work.
Each of the eight rack-scale systems combines NVIDIA Vera CPUs and Rubin GPUs. NVIDIA claims the configuration can deliver up to ten times the performance per megawatt of the infrastructure it replaces. The announcement does not provide a measured workload comparison, total purchase price or deployment completion date.
AI across the research pipeline
Bristol Myers Squibb describes a “Predict First” approach in which computational models help researchers decide which molecules should move into laboratory synthesis. The goal is to focus scarce experiments on candidates that have a higher predicted chance of meeting several required properties.
The company says AI-enabled target identification can save weeks of manual work. With more capacity, teams plan to expand those predictions, train proprietary foundation models and use agents to work across traditionally separate research systems.
The unified environment will use a common data plane across Bristol Myers Squibb sites. NVIDIA Mission Control will manage the infrastructure, while AI-oriented tools are intended to let researchers start complex predictions through plain-language requests instead of requiring deep knowledge of the underlying cluster.
BioNeMo and agentic workflows
BioNeMo Agent Toolkit provides components for building biological and chemistry agents. In this deployment, it is expected to support workflows that combine models, scientific data and computational tools across stages of discovery.
Bristol Myers Squibb argues that agents could help knowledge move across organisational and therapeutic silos. A finding produced by one research programme could be available to models and teams working elsewhere, creating a cumulative learning loop rather than leaving each project with a separate body of evidence.
That vision depends on data quality, permissions and scientific oversight. Agents can make information easier to retrieve or connect, but they cannot determine whether a biological result is reproducible or whether a candidate is safe. The company says human scientists will continue to drive decisions, assess caveats and teach systems how to use institutional knowledge.
Why the deployment matters
The project is a large enterprise commitment to the Vera Rubin platform and a concrete example of pharmaceutical AI moving from isolated experiments into shared infrastructure. It also shows NVIDIA pairing hardware with domain software and operational tooling rather than selling accelerators alone.
However, the claims in the announcement come from NVIDIA and Bristol Myers Squibb. No independent performance results, energy measurements or drug-development outcomes from the new system are yet available. Its value will ultimately depend on whether greater compute produces better validated decisions and shorter research cycles, not simply more model runs.
Operational questions remain
Running a shared global research environment will require careful allocation of expensive capacity. Bristol Myers Squibb will need to decide which tasks deserve the new systems, how interactive agent work coexists with large training jobs and how researchers can reproduce results as models, data and software change.
Life-sciences data also brings strict governance requirements. Access controls, data lineage and validation records must follow information as it moves between sites and tools. Plain-language interfaces can make advanced compute easier to use, but they should not make underlying assumptions or provenance harder to inspect.
The deployment may provide an important case study once it is operating at scale. Useful evidence would include queue times, energy consumption, researcher adoption and the proportion of model-generated hypotheses that survive experimental review. Until those results are available, the announcement demonstrates investment and intended scope rather than proven clinical impact.
The ten-times performance-per-megawatt claim should also be read carefully. It compares the new configuration with replaced infrastructure under conditions that are not described in the announcement. Buyers considering similar systems will need workload-level measurements that include utilisation, cooling, networking and data movement, as well as the accelerator’s peak specifications.