DGXSuperPODa
Eli Lilly said today it is building what it calls the pharmaceutical industry’s most powerful AI supercomputer in partnership with Nvidia. The system, billed as the first Nvidia DGX SuperPOD configured with DGX B300 systems, will power an “AI factory” for drug discovery and development. Lilly and Nvidia say the installation will use 1,016 Nvidia Blackwell Ultra GPUs and is rated at more than 9,000 petaflops of AI performance, enabling workloads from genome-scale analyses to training foundation models on Lilly’s experimental data.
Beyond early discovery, Lilly says it will apply the system to digital twins of manufacturing lines for predictive maintenance, AI-assisted clinical-trial document workflows, and medical-imaging research for biomarker discovery. Select proprietary models will be exposed to partners via Lilly TuneLab, a federated platform that uses Nvidia’s FLARE framework so biotechs can fine-tune models without sharing their underlying data. Rau explained that Lilly has data on “millions of well-described molecules. Almost all of those molecules don’t work. But to train an AI model, you need the data on what doesn’t work. Public datasets mostly contain molecules that do work.” Lilly says the supercomputer will run on 100% renewable electricity in existing facilities and use chilled-water liquid cooling.
In plain terms, the supercomputer will have roughly twice the GPU count of Recursion’s BioHive-2 (1,016 versus 504). Recursion previously boasted that its BioHive-2, built with 63 DGX H100 systems totaling 504 H100 GPUs, was the largest supercomputer owned and operated by a pharmaceutical company. BioHive-2 ranks 69th on the June 2025 Top500 list; Lilly’s 1,016-GPU Blackwell Ultra DGX B300 SuperPOD, which the company says delivers more than 9,000 AI petaflops, would likely rank among the world’s more powerful commercial systems, though it would trail the exascale leaders built primarily for government research labs. Lilly has not disclosed Linpack benchmark results that would determine its precise Top500 ranking.
Lilly is shifting from using AI as a tool to embracing it as a scientific collaborator. —Thomas Fuchs, senior vice president and chief AI officer
“By embedding intelligence into every layer of our workflows, we’re opening the door to a new kind of enterprise: one that learns, adapts and improves with every data point,” said Thomas Fuchs, senior vice president and chief AI officer, in a statement. Fuchs said the company’s foundation models are “spawning new possibilities for our chemists, helping them uncover new motifs and configurations of atoms that were out of reach with traditional methods.” For readers, the near-term signal is less about AI-discovered drugs hitting the market, clinical timelines still apply, and more about Lilly consolidating in-house compute to train and serve large biomedical models, then selectively extending those models to external biotechs through TuneLab.
Rau tempered expectations about near-term clinical impact, noting that “for the medicines we’re working on now, we’re going to see those benefits in the 2030s at the earliest.” He told an interviewer at the GTC conference it takes 10 years on average and about $2.6 billion to bring a single medicine to market. “Biology just takes time… So it’s going to be very hard to speed up biology, even if you speed up everything else around it.” Still, Rau argued that “I don’t believe any other company in our industry is doing what we do at this scale. As a 150-year-old medicine company, one of our most powerful assets is decades of data.” The supercomputer is part of Lilly’s $50 billion commitment to expanding its U.S. manufacturing and R&D footprint, which includes a proposed $4.5 billion lab in Indiana called the Lilly Medicine Foundry. CB Insights ranks Lilly as the most AI-ready pharma company, citing the highest AI innovation score and most partnerships with AI drug development platforms.
Lilly frames the buildout as a hedge against generic, public models that can be overly agreeable and error-prone in scientific settings. In a recent fireside chat with Anthropic CEO Dario Amodei, Lilly CIO Diogo Rau argued that enterprise AI must be optimized for truthfulness and domain depth, not engagement, and urged companies to rework end-to-end workflows rather than swap AI into isolated steps. That stance helps explain Lilly’s push to train specialized, “graduate-level” biomedical models on proprietary data and serve them internally at scale, while exposing select models to partners through TuneLab’s federated setup. Safety researchers, and Anthropic’s Amodei in the video below, have documented “sycophancy,” where models extol a user’s ideas over facts. Such behavior is unacceptable in many business contexts, but especially in high stakes areas like drug design. Lilly’s move bets that controlled training, evaluation, and serving on its own infrastructure can reduce these failure modes while speeding iteration.
