The $1 Billion Bet: How Eli Lilly is rebuilding drug discovery around “AI-native” science

For nearly 150 years, pharma has run on a familiar operating system: talented scientists, careful trial-and-error, and a long wait to see which molecules survive the clinical gauntlet. That model still works — but it’s slow, expensive, and painfully exposed to “patent cliff” cycles.

Lilly’s current move is different from the usual “we added AI to our workflow” press release. They’re trying to rebuild the workflow itself around compute, models, and automation — an “AI-native” approach where the lab and the algorithm form a tight loop.

LillyPod: turning compute into a scientific instrument

On February 26, 2026, NVIDIA described LillyPod as “now live”: the world’s first NVIDIA DGX SuperPOD with DGX B300 systems deployed for pharma, powered by 1,016 NVIDIA Blackwell Ultra GPUs and rated at more than 9,000 petaflops of AI performance. Lilly inaugurated the system at a ribbon-cutting in Indianapolis. The framing from both sides was consistent: this isn’t “IT.” It’s a new lab instrument.

NVIDIA also highlighted a historical comparison from Lilly leadership: computational power that once required millions of Cray-class supercomputers is now concentrated inside a single modern GPU — and LillyPod contains more than 1,000 of them. The practical point is simple: you can ask bigger biological questions, faster, and with fewer compromises.

R&D World added a useful benchmark: the idea that a productive wet-lab team might test on the order of a few thousand molecules per target per year — while large-scale compute enables exploring vastly larger hypothesis spaces in parallel. That doesn’t remove the need for wet-lab validation. It changes what gets sent to the wet lab in the first place.

What models are they actually talking about?

NVIDIA’s own write-up is refreshingly concrete about the kinds of workloads LillyPod is meant to support: protein diffusion models, small-molecule graph neural networks, and genomics foundation models. That shortlist matters because it signals where the industry is heading: not “chatbots for science,” but models that can learn the structure of biology and chemistry at scale.

Platform strategy: TuneLab and the “give/get” model

Lilly isn’t keeping everything behind glass. In September 2025, the company announced TuneLab — a platform offering biotech companies access to AI-enabled drug discovery models built on Lilly’s internal R&D work. The interesting part is the mechanism: federated learning, which allows models to improve using distributed data contributions without forcing participants to hand over raw proprietary datasets. It’s a classic platform play: provide tools, attract users, learn from usage, improve the tools.

This matters because “AI advantage” in biopharma doesn’t come only from buying GPUs. It comes from data flywheels and tight feedback loops — and TuneLab is designed to create exactly that.

Background: The $1B co-innovation lab with NVIDIA

On January 12, 2026, Lilly and NVIDIA announced an AI co-innovation lab in the San Francisco Bay Area, with up to $1B in combined investment over five years (talent, infrastructure, compute). The stated goal is to tackle hard pharma problems by co-locating Lilly domain experts with NVIDIA AI engineers — basically, fewer handoffs and more “build it together.”

Reuters described the same effort as aimed at accelerating drug discovery and development through advanced AI, and noted the lab would be in the Bay Area with more location detail expected later. Whether you love or hate corporate partnerships, this is a serious signal: Lilly doesn’t see AI as a side project. They see it as a core capability worth institutionalizing.

Illumina’s Billion Cell Atlas: why single-cell scale is the new baseline

On January 13, 2026, Illumina announced the Billion Cell Atlas — a massive CRISPR perturbation dataset intended to help train advanced AI models on biologically grounded data. Eli Lilly is listed as a founding participant alongside AstraZeneca and Merck.

The “one billion cells” headline isn’t just a flex. It’s an admission that biology is too complex for small datasets to settle big questions. If you want models that generalize across cell types, disease contexts, and genetic backgrounds, you need scale that looks ridiculous by 2016 standards. This is what that looks like.

So where does vitiligo fit in?

Let’s be careful and honest here: Lilly has not publicly announced a vitiligo-specific drug program tied to LillyPod. So what follows is not a “Lilly is doing vitiligo” claim. It’s a “this infrastructure makes certain vitiligo questions much more attackable” argument.

Many researchers now frame vitiligo relapse as an “immune memory” problem in the skin: tissue-resident memory T cells (TRM) can persist locally and help drive recurrence after treatment stops. One of the pathways repeatedly discussed in this context is IL-15 signaling (including CD122), which supports TRM survival and function. In mouse models, blocking IL-15/CD122 signaling has produced durable repigmentation after treatment, suggesting a potential route toward longer-lasting control — not just temporary clearance.

If you zoom out, this is exactly the kind of problem that benefits from hyperscale biology: which patients carry which immune programs, which pathways dominate in which skin microenvironments, and why some people respond to one class of therapy while others don’t. Single-cell atlases, large-scale genomics, and better biological foundation models won’t “solve” vitiligo on their own. But they can compress the time between a plausible idea and a tested, patient-relevant hypothesis.

The bottom line

Lilly is trying to escape the classic pharma cycle by making discovery more like engineering: build a powerful internal compute backbone, connect it to high-quality data, run tight experiment loops, and turn learning speed into a competitive advantage.

The optimistic take is not “AI will replace scientists.” It’s “we can finally test more of our best ideas before the decade is gone.” For diseases like vitiligo — where the biology is real, the burden is real, and the historical underinvestment is also very real — faster learning is not a luxury. It’s the difference between progress and another generation of ‘almost.’

by Yan Valle

Prof. h.c., CEO VR Foundation 

Suggested reading:

• AI and Vitiligo: The Digital Frontier Gets Real
• 📌 Vitiligo Drug Pipeline Analysis and Market Insights
• AI in Drug Discovery – Big Promises, Slow Progress

Listen to Deep Dive In Vitiligo Podcast:

• The State of Vitiligo 2025: A Fast-Moving Field With Slow-Moving Funding (Ep. 57)
• AI for Vitiligo Patients — Beyond the Hype (Ep. 51)
• AI Now Speaks the Language of Cells (Ep. 40)

Definitions

• AI factory: an integrated stack (compute + networking + software + workflows) designed to run large-scale AI workloads reliably, like an industrial plant for model training and inference.
• DGX SuperPOD: NVIDIA’s reference architecture for building large-scale AI supercomputing clusters.
• Federated learning: an approach where models learn from distributed data without centralizing raw datasets in one place.
• TRM cells: tissue-resident memory T cells; immune cells that can remain in tissue (like skin) and contribute to relapse/recurrence in some autoimmune diseases.

References

1. NVIDIA Blog (Feb 26, 2026). “Now Live: The World’s Most Powerful AI Factory for Pharmaceutical Discovery and Development.” Link
2. Eli Lilly Investor News Release (Sep 9, 2025). “Lilly launches TuneLab platform…” Link
3. NVIDIA Press Release (Jan 12, 2026). “NVIDIA and Lilly Announce Co-Innovation AI Lab…” Link
4. Illumina Investor Press Release (Jan 13, 2026). “Illumina introduces Billion Cell Atlas…” Link
5. R&D World (Feb 27, 2026). “Eli Lilly’s LillyPod supercomputer goes live…” Link