Eli Lilly, Roche build supercomputers
- Eli Lilly and Roche are no longer just buying AI software. They are building giant in-house compute systems to run drug discovery themselves. - Roche said in March it expanded to 3,500-plus GPUs with NVIDIA; Lilly said earlier it was building pharma’s most powerful owned supercomputer. - That matters because pharma AI is shifting from cheap pilot projects to heavy capital spending on chips, data pipelines, and wet-lab integration.
Drug discovery is turning into a compute business. Not just in the usual “everyone uses AI now” way, but in the much more concrete sense that big drugmakers are building their own supercomputers. Roche said on March 16 that it had launched an NVIDIA-powered AI factory with more than 3,500 GPUs across on-premise and hybrid-cloud systems. Lilly said in late 2025 that it was building what it called the most powerful AI supercomputer owned and operated by a pharmaceutical company. (roche.com) ### What are they actually building? These are high-performance computing stacks tuned for AI — basically giant clusters of advanced GPUs, data systems, and software tools that can train models, run simulations, and search through huge biological datasets fast enough to matter. Roche’s setup adds 2,176 NVIDIA Blackwell GPUs and brings its announced total to more than 3,500 GPUs. Li(roche.com) infrastructure for research and operations. (roche.com) ### Why does pharma need that much compute? Because modern drug discovery is a mess of giant search problems. You are trying to predict which molecules bind to which targets, which designs are manufacturable, which compounds are toxic, and which candidates might still work in real biology after all the modeling. That means chewing through chemistry data, protein structure data, ima(roche.com)ts this compute embedded across discovery, development, manufacturing, and commercial work — which tells you the ambition is full-stack, not just molecule generation. (roche.com) ### Why not just rent cloud AI? They still will. But the catch is that pharma has awkward workloads — sensitive data, long-running experiments, and expensive pipelines that need to connect directly to internal labs. Owning more of the stack gives companies tighter control over security, latency, and cost at scale. Roche explicitly calls its system a hybrid-cloud AI factory, which i(roche.com)both, stitched together for industrial use. (roche.com) ### Why is NVIDIA in the middle of this? Because NVIDIA is selling more than chips. It is selling a whole operating model for AI-heavy science — GPUs, networking, software, and partnerships. In January, Bloomberg reported that NVIDIA planned to invest $1 billion over five years in a new lab with Lilly in Silicon Valley to speed pharma AI. Then Roche expanded its own NVIDIA tie-up i(roche.com) for AI infrastructure, not just a customer for occasional cloud bursts. (bloomberg.com) ### Does this mean AI is already fixing drug discovery? Not exactly. Lilly’s own AI platform material says the industry is ramping up AI broadly, but drug discovery is “not the primary success story—yet.” That “yet” matters. The promise is real, but biology still punishes overconfidence. Models can rank molecules beauti(bloomberg.com)ween models and labs so companies can kill bad ideas faster and push better ones forward sooner. (tunelab.lilly.com) ### What changed from the last AI wave? Scale and seriousness. A few years ago, pharma AI often meant partnerships, pilot programs, and slide-deck promises. Now the biggest companies are putting hard assets behind it — GPU clusters, dedicated labs, and internal platforms. Lilly also linked AI directly to its drug pipeline in 2024 when it teamed with OpenAI on antimicrobial discovery. This looks les(tunelab.lilly.com)n. (investor.lilly.com) ### So what is the real bet? The bet is that compute becomes a strategic moat in pharma, the way manufacturing scale became a moat in semiconductors. If Roche and Lilly are right, the winners in AI drug discovery will not just have better models. They will have better data, tighter lab feedback loops, and enough owned infrastructure to run the whole machine continuously. (roche.com)
