Bristol Myers Squibb Deploys Second AI Supercomputer for Drug Discovery

Bristol Myers Squibb announced today the deployment of its second NVIDIA DGX SuperPOD, built on eight DGX Vera Rubin NVL72 systems. The announcement represents a major escalation in the company's AI infrastructure commitment after running its first DGX SuperPOD for approximately three years.

The new system delivers up to 10x the performance per megawatt of the infrastructure it replaces and comprises NVIDIA Vera CPUs and Rubin GPUs. BMS will integrate this system with its existing SuperPOD into a unified environment—a single data plane—accessible from every BMS site globally. This unified architecture replaces earlier barriers that made the first system difficult to access, including site-specific restrictions inherited from past acquisitions and the requirement for deep computational expertise. Researchers will now be able to initiate complex predictions in plain English through NVIDIA Mission Control.

Erin Davis, vice president of research business insights and technology at BMS, framed the shift in terms of access: "Instead of equipping a small group of researchers with access to the supercomputer, we're opening it up to literally every scientist. No one has to wait, and no one is told they have a limit." She described the current state of compute demand at BMS as saturated, noting that the company is running large-scale predictions around large molecules and building its own foundational models, both of which require significant GPU capacity.

The company has already demonstrated measurable returns from its first SuperPOD. AI-enabled target identification saves scientists weeks of manual work. BMS has used AI to expand its library of CELMoD compounds—molecules engineered to selectively degrade cancer-causing proteins for blood cancer treatment and beyond—opening doors to new disease targets and potential medicines. In lead optimization stages, BMS applies a methodology called "Predict First," where predictions inform experimental gating based on design predictions. According to Payal Sheth, senior vice president of therapeutic discovery sciences at BMS, this approach uses "predictions as a way to prioritize synthesis of molecules with multi parameter optimization" to weed out molecules unlikely to meet desired properties, ensuring "precious laboratory experiments are aligned with progressing molecules that have the highest probability of success."

Sheth, who transitioned from conducting drug discovery research into an expanded leadership role in January, described the mandate as moving from "sort of this abstract position of what AI can do to actually translating that to measurable impact." The unified compute infrastructure is designed to enable cumulative learning across programs: datasets from a program run in Lawrenceville, New Jersey, feed models that teams in San Diego, California, can draw on. This contrasts with traditional drug discovery, where each project was treated discretely and learnings did not compound into an intelligence framework. Davis emphasized the value of agentic workflows in breaking down silos: "Agents don't care. They go all across. And that is a huge game-changer because now we can learn from decisions across the silos and across programs." The detailed allocation plan for the new system spans small and large molecule design, clinical applications, and digital twins across every node.

Bristol Myers Squibb's investment in a second advanced AI supercomputer reflects a maturation in how large pharmaceutical companies view computational infrastructure. After three years of running its first DGX SuperPOD, BMS has moved past the pilot phase—the system has already demonstrated concrete value through accelerated target identification and expansion of its CELMoD compound library for cancer treatment. The constraint driving the new deployment is not skepticism about AI's utility but scarcity: researchers are saturated with demand and unable to access enough compute for large-scale predictions and foundational model development.

The strategic shift from a small-group research tool to a unified global platform accessible to all scientists represents a recognition that AI's return on investment increases when the barrier to entry drops. By removing site-specific restrictions leftover from past acquisitions and replacing deep computational expertise requirements with AI-native tooling, BMS is betting that democratized access will unlock compound learning loops—where datasets from one program feed models that other teams can draw on. This cumulative learning architecture did not exist in traditional drug discovery, where each project was treated discretely. The integration of the existing and new systems into one data plane is designed to institutionalize those learnings across all BMS programs globally.

No comments yet. Be the first to share your thoughts!