Bristol Myers Squibb Builds a Drug-Discovery Supercomputer

Bristol Myers Squibb wants to run its own AI factory, and it has bought the hardware to do it. On 20 July the pharmaceutical company said it would deploy an Nvidia DGX SuperPOD built on the company's newest DGX Vera Rubin NVL72 systems, which it describes as the most powerful single-owned Nvidia infrastructure in life sciences.

The appeal of the Vera Rubin generation is as much about the electricity bill as the raw speed. Nvidia says the architecture delivers up to ten times the performance per megawatt of its predecessor, which lets a research operation take on much larger AI workloads without a matching jump in power draw. For a company planning to run these systems continuously, that ratio is the whole argument.

Training on its own data

What Bristol Myers Squibb plans to do with the cluster is the interesting part. Rather than lean on a general-purpose chatbot, the company intends to train AI foundation models on its own proprietary research data and use them to power agentic workflows across its oncology, haematology, cardiovascular, immunology and neuroscience programmes. The work will draw on BioNeMo, Nvidia's platform for biological AI.

This is not a standing start. The company says the deployment builds on nearly three years of work with Nvidia that began with an earlier DGX SuperPOD, and the two systems will be joined into a single environment that researchers can reach from any Bristol Myers Squibb site worldwide.

The move fits a wider pattern in which large drugmakers are bringing serious compute in-house rather than renting it, betting that models tuned on their own molecules and trial data will beat anything trained on the open internet. Whether that translates into medicines any faster is the question the whole industry is spending on, and it will take years of clinical work to answer. The hardware is the easy part.