Eli Lilly, Roche build supercomputers

Drug discovery is starting to look a lot more like cloud infrastructure. That’s the real story here. Eli Lilly and Roche aren’t just buying AI software or signing splashy model deals — they’re building giant supercomputing systems because they think the hard part of AI in pharma is now scale, speed, and integration. The old pitch was “AI will find drugs.” The new one is more grounded: “AI might help, but only if you can feed it huge amounts of proprietary biology, run models constantly, and close the loop with real experiments.” (blogs.nvidia.com) ### What are they actually building? Lilly has built what Nvidia describes as the largest AI factory wholly owned and operated by a pharmaceutical company — a DGX SuperPOD system with 1,016 Blackwell Ultra GPUs. Roche took a slightly different route. In March, it said it was adding 2,176 Blackwell GPUs and that its combined on-premise(blogs.nvidia.com)hough — massive compute dedicated to biomedical work. (blogs.nvidia.com) ### Why does a drug company need that much compute? Because modern drug discovery has become a data-and-search problem. You’re trying to model proteins, molecules, patient subgroups, toxicology signals, manufacturing constraints, and trial design — all at once. A single promising molecule is not enough. You need systems that can gener(blogs.nvidia.com)ast. That only works if model training and inference stop being scarce internal resources. (blogs.nvidia.com) ### Why now? The industry seems to have hit the limits of “just add AI.” Pharma companies spent the first wave testing copilots, discovery models, and partnerships. But a lot of that work ran into the same wall — fragmented data, slow wet-lab validation, and not enough compute to run serious foundation models on proprietary datasets. (blogs.nvidia.com)he factory before promising the miracle. (blogs.nvidia.com) ### What does Lilly want from its system? Lilly says the machine will train large biomedical foundation and frontier models for discovery and development. It also plans to plug some of those models into TuneLab, its federated AI platform, which lets outside biotech partners use Lilly’s models without pooling raw data together. That ma(blogs.nvidia.com)ter is useful — but only if it can sit inside that reality. (blogs.nvidia.com) ### What’s Roche doing differently? Roche is spreading AI infrastructure across more of the business. It’s using the Nvidia stack not just for therapeutics, but also diagnostics, pathology, manufacturing, and digital health. In research, it ties BioNeMo into its “Lab-in-the-Loop” workflow — meaning model outputs feed into real biology(blogs.nvidia.com)loop is the whole game. Without it, pharma AI is just very expensive autocomplete. (roche.com) ### Is this really about supercomputers — or about failure rates? Mostly the second one. Drug development still fails far more often than it succeeds, and late-stage failure is brutally expensive. These compute builds are an attempt to kill bad ideas earlier, surface better ones faster, and maybe design smarter trials. They won’t repeal biology. But they could im(roche.com)n a single breakthrough claim. (wsj.com) ### What’s the catch? More GPUs do not automatically mean more drugs. Models are only as good as the data, the assays, and the organizational plumbing around them. If lab systems, clinical datasets, and chemistry workflows don’t connect cleanly, the supercomputer just makes the bottleneck happen faster. And because these are proprietar(wsj.com)ental data — not the companies with the flashiest demo. (roche.com) ### So what changes now? The center of gravity in pharma AI is shifting. Less talk about one magical model. More spending on compute, data pipelines, and closed-loop experimentation. Lilly and Roche are betting that the next edge in drug discovery won’t come from having AI in the abstract. It’ll come from owning the machinery that lets AI run at industrial scale. (blogs.nvidia.com)