Eli Lilly, Roche race to build supercomputers

Drugs are becoming a compute story. That’s the real shift here. Eli Lilly and Roche are no longer treating AI as a side tool for a few research teams — they’re building giant in-house AI factories and supercomputing stacks so model training, simulation, and lab automation sit much closer to the core of drug discovery. Why? Because drug development is brutally inefficient. Most candidates still fail before they become approved medicines, and a lot of that failure comes from picking the wrong targets, the wrong molecules, or the wrong patients too late in the process. The pitch for these systems is simple: use much bigger models on much bigger biological datasets to rule out bad ideas earlier and generate better ones faster. What changed recently is scale. Roche said in March 2026 that it is expanding a hybrid-cloud AI factory with 2,176 NVIDIA Blackwell GPUs added on premises across the U.S. and Europe, bringing its total announced footprint to more than 3,500 GPUs. Roche framed that as infrastructure for the whole value chain — discovery, development, manufacturing, and commercial work — not just one isolated research program. Lilly has been just as aggressive, but with a slightly different angle. In October 2025 it unveiled what NVIDIA called the largest and most powerful AI factory wholly owned and operated by a pharma company, built around 1,016 Blackwell Ultra GPUs. Then in early 2026 Lilly and NVIDIA added a co-innovation lab aimed at drug discovery, robotics, and what they call “physical AI” for medicine discovery and production. In other words, Lilly isn’t just buying compute — it’s trying to wire compute directly into how experiments get designed and run. That last part matters more than the GPU count. A supercomputer does not magically fix biology. If the underlying lab data are messy, sparsely labeled, or trapped in incompatible systems, bigger models just learn faster from worse inputs. That’s why the phrase showing up around this whole push is “lab-in-the-loop” — models generate hypotheses, automated or semi-automated labs test them, and the results feed back into the models. Basically, pharma is trying to close the loop between prediction and experiment. So what are Lilly and Roche really racing to build? Not a single monster machine for bragging rights. They’re building a stack. Compute at the bottom. Foundation models and molecular-design tools in the middle. Then data pipelines, robotics, and manufacturing simulations on top. Roche has been explicit that its AI factory will also support diagnostics and digital twins for manufacturing. Lilly has been explicit that discovery and delivery both sit inside the plan. The catch is that returns may show up unevenly. AI has already been useful in narrowing search spaces, generating candidate molecules, and helping scientists prioritize experiments. But “AI-designed drug” is still not the same thing as a faster approval or a commercial winner. Biology stays noisy, clinical trials stay expensive, and regulators still care about evidence, not GPU totals. That’s why these companies are spreading the bet across discovery, development, and operations instead of promising one miracle model. The practical implication for the rest of the industry is pretty clear. If frontier pharma starts running on AI factories, then suppliers, CROs, and internal lab teams will face pressure to produce cleaner, more reproducible, machine-readable data. The bottleneck shifts. It stops being just “do we have enough compute?” and becomes “can the real-world lab produce data good enough for the compute to matter?” Bottom line — Lilly and Roche are turning drug discovery into an infrastructure race. The winners won’t just have better models. They’ll have tighter loops between algorithms, experiments, and manufacturing.