The Dry Lab

Last week's biomanufacturing story followed the work of making useful products with living cells. This week moves earlier in drug development, to choosing which molecules to make and test. Eli Lilly is building substantial computing power around that choice.

On February 25, 2026, the company inaugurated LillyPod in Indianapolis, a supercomputer built with NVIDIA hardware. Its work includes training models to investigate proteins, chemical compounds and genetic data.

In the wet lab, scientists work with physical materials, making compounds and measuring what they do. The dry lab investigates ideas through computation. Predictions can help researchers choose an experiment, and its results can become evidence for later predictions. Lilly and NVIDIA want to connect those activities into a continuous learning system, with scientists involved throughout.

Lilly sells tirzepatide under the names Mounjaro and Zepbound, two of the medicines associated with the rise of GLP-1 drugs. Together, they generated $36.5 billion in worldwide sales in 2025, more than half of Lilly's revenue. A useful molecule can support an enormous business.

Tirzepatide works by activating two kinds of receptors, proteins that receive chemical signals. One is the GLP-1 receptor, which spans a cell's outer membrane. The drug binds to its outward-facing side, initiating a signal inside the cell. The picture below shows that interaction using an experimentally determined structure.

A protein like this receptor can be a target for drug discovery. Researchers use computers to explore which molecules might act on it, narrowing the possibilities before making and testing compounds in the lab.

Lilly's research computing grew quickly. Thomas Fuchs, its chief AI officer, recalled that Lilly's research GPU count grew from about eight when he joined in 2024 to 100 before LillyPod arrived. LillyPod contains 1,016 NVIDIA Blackwell Ultra GPUs. These processors perform the calculations used to train and run AI models.

Lilly owns and operates the system. NVIDIA supplies the computing hardware, networking and software, while Lilly brings its research questions and decades of experimental data. Scientists use those records to train models, then use the models to predict how new molecules might behave. The computing infrastructure supports both steps.

For a sense of that computing power, consider three systems used in research. Amgen's Freyja studies human genetics, Recursion's BioHive-2 models biology and chemistry, and Denmark's Gefion serves a wider research community.

Their chips come from different generations, so counting GPUs misses part of the picture. On the same theoretical measure of AI computing capacity, LillyPod nearly matches those three published configurations combined. That puts the size of Lilly's machine in context. How much useful research it produces depends on the questions, models and experiments built around it.

A computer can consider a molecular structure without anyone first having to make the compound. One approach, called docking, estimates how well a molecule might fit and interact with a target protein. Researchers can use those estimates to help decide which physical tests are worth doing.

A 2019 study of the dopamine D4 receptor makes the difference in scale visible. Researchers computationally screened more than 138 million virtual structures. They chose compounds the computer thought likely to bind to the receptor, along with less promising candidates, so lab tests could check the predictions. They successfully made and tested 549.

This example predates LillyPod and used physics-based docking. It shows how computation can help direct physical work, without establishing how productive Lilly's new system will be. Binding to a protein is an early finding; researchers still have to investigate whether a candidate can become a safe and effective medicine.

The return trip through the loop matters as much as the search. A prediction can be compared with what happened in the experiment. Records of molecules that disappointed researchers can also supply training examples, helping later models account for behavior the earlier predictions missed.

TuneLab, launched in September 2025, gives other biotechnology companies access to prediction models trained on Lilly's research. For example, a company developing an antibody, a type of protein, can use a model to estimate whether its candidate will clump. In exchange for access, participating companies let the models learn from their own experimental results.

The approach is called federated learning. A third party keeps each company's records in a separate, protected space. A copy of the shared model learns from the experiments in each space. The system combines the changes learned by those copies; the underlying experimental records remain separate.

The companies gain tools built from Lilly's research, while Lilly and its partners can train shared models on a wider range of experiments than any one participant supplies. TuneLab checks performance before and after training and offers partners the best-performing shared model. Better predictions are the aim, not a guarantee. TuneLab is a software platform, separate from the LillyPod supercomputer.

One specific project is taking shape at Lilly's Biotechnology Center in San Diego. In an October 6 recruiting notice, Lilly describes a program with dedicated capacity on LillyPod to design new antibodies and other protein-based medicines.

The team aims to improve how those molecules work and how reliably they can be manufactured, then use experimental results to guide the next designs. Lilly has described the program, without naming a resulting drug or reporting a success rate.

For this team, choosing what to try next includes designing new proteins for scientists to make and test. The dry lab proposes a design. What happens when scientists make it in the wet lab helps determine the next one.

Anthony

The Luniter Observatory

Disclosure: I own shares of Eli Lilly (LLY) and NVIDIA (NVDA).

Research and educational analysis only — not financial advice. Every investment carries risk.