Biomedical AI · us
When Computing Power Enters the Drug Laboratory: Bristol Myers Squibb Bets on a Next-Generation AI Supercomputer
Eight Vera Rubin systems will bring molecular design, target assessment, and model training into a single computing environment; but a larger search space for drug candidates still does not mean a higher rate of clinical success.
The bottleneck in drug development is often not a lack of molecules to test, but how to select the small number of candidates most worth sending to the laboratory from an almost boundless range of options. Bristol Myers Squibb (BMS) is now betting part of the answer on larger-scale computing: the company will deploy a second NVIDIA DGX SuperPOD in an effort to turn AI from an individual research tool into shared infrastructure connecting the drug discovery process.
The new system consists of eight DGX Vera Rubin NVL72 systems and will be integrated with BMS’s existing SuperPOD to create a single environment accessible to research sites worldwide. According to the two companies, the architecture can deliver up to 10 times the performance per megawatt of the equipment it replaces, enabling it to handle larger biomedical foundation models, predictive workloads, and AI agent workflows without increasing energy consumption at the same rate. This is an architectural performance claim made by the supplier and purchaser, not yet a comparative result for drug development output.
Its specific applications lie at the front end of research and development decision-making. BMS plans to train proprietary models on decades of accumulated internal data and use NVIDIA BioNeMo tools as needed, enabling the system to help organize target evidence, predict molecular properties, design candidate structures, and test research hypotheses. The so-called “Predict First” approach uses models to screen out designs that fail to meet multiple criteria before deciding which molecules are worth synthesizing and advancing to wet-lab experiments; AI agents can handle cross-database searches and repetitive analyses, leaving researchers more time to interpret results and make experimental trade-offs.
BMS says AI is already used in all small-molecule programs and most large-molecule programs; automating target identification and validation can save weeks of manual work. Reuters cited Robert Plenge, the company’s head of research and development, as saying that AI allows teams to evaluate dozens of drug candidates at an early stage, rather than about 10 in the past, and has already shortened the development time of some drugs entering clinical trials by about 20% to 30%. These figures come from the company’s internal experience; the report did not provide calculation baselines, control groups, the number of programs covered, or peer-reviewed validation data.
Greater computing power also cannot replace biological and clinical evidence. Models can rank molecules, but efficacy, toxicity, and suitable patient populations must still be confirmed through cell, animal, and human trials; if training data contain biases, omissions, or incomparable experimental conditions, models may also amplify existing errors. When AI-generated analyses feed into drug candidate selection and clinical development, companies must also establish mechanisms for data provenance tracking, model version management, result reproducibility, and human review to address quality audits and regulatory inspections.
This investment extends nearly three years of collaboration between BMS and NVIDIA and also reflects how major pharmaceutical companies are treating proprietary AI computing capacity as part of research and development competition; Roche and Eli Lilly have also recently announced NVIDIA infrastructure plans. However, BMS has not disclosed the purchase amount, deployment timeline, or quantified research and development success metrics. The real test is not how many predictions the system can generate, but whether, under auditable conditions, it can consistently eliminate incorrect hypotheses and increase the probability that final drug candidates pass clinical validation.