Moats in Motion: Federated AI and the Next Pharma Gap
I recently had the pleasure of speaking at Licensing Executives Society (U.S.A. and Canada), Inc. in Boston. I tried to make the points that the next edge in pharma isn’t a single breakthrough – it’s going to be systems that convert data + compute + partners into repeatable velocity. Eli Lilly and Company’s recent moves are exactly that. They’re building an on-prem NVIDIA Blackwell-class DGX SuperPOD in Indianapolis with 1,000+ GPUs, powered by renewables and targeted online by January. Crucially, this isn’t a lab showpiece. They intend to point it at discovery, clinical-trial design, manufacturing quality, and enterprise analytics. Owning the stack lowers latency where it matters, keeps decades of proprietary data under their control, and reduces exposure to supply and geopolitical shocks. This is how I believe that you can wire an organization for compounding learning, not just cost take-out.
The second piece is TuneLab – Lilly-trained models distributed to selected biotechs to fine-tune locally in a federated setup. Partners like Circle Pharma and insitro keep their data in place and contribute only model updates, while tapping into what Lilly says reflects >$1B of internal data investment under the Catalyze360 umbrella. Functionally, this widens the learning loop without centralizing sensitive datasets, and it turns Lilly’s model family into a living system that improves as the network grows. Whoever at Lilly is stitching this together has caught my eye – it is really blending the art of owning differentiated capabilities and scaling them through partners without bleeding IP. Kudos to whoever is helping drive strategy there.
If they execute, success looks practical and measurable. Here is how I am thinking about it: First, throughput: more design-make-test cycles per quarter and the ability to prosecute tougher targets and modalities. Second, quality at the source: predictive ADMET/tox pushes attrition left, so fewer costly failures surface late; that matters in a world where overall clinical success rates hover in the single digits. Third, clinical execution: smarter inclusion/exclusion, lower screen-failures, and tighter protocols compress timelines – even as trials remain multi-year. Patient-visible wins will ramp later this decade and into the 2030s, but the intermediate gains (earlier go/no-go calls, cleaner trials) should show up much sooner.
Now what most do not think about – the ripple effects. Second order (12-24 months): the federated approach creates partner gravity – each local fine-tune improves the foundations without moving raw data – raising switching costs for competitors and making the platform a magnet for the best biotechs. Portfolio discipline improves as earlier kill signals and better priors redirect capital toward higher-quality assets. This also pushes CROs/CDMOs to evolve offerings around AI-native trial design and model ops. And because the R&D stack increasingly matches the FDA’s 2025 NAMs roadmap to reduce animal testing, preclinical packages should become more aligned and review ready.
My timeline to remain competitive: in my view, companies have 18-36 months to implement many of these shifts. After that, I’m not convinced catching up is possible. If you are committed to not doing anything or just cost cutting – please sell the company now, for the sake of shareholders. (I am not trying to be negative; it’s how compounding advantages work in practice.) Remember I do not have a horse in this race other than getting older and wanting better therapies :).
Third order (3-5 years): if Lilly compounds learnings quarter after quarter, we could see a winner-take-most tilt: data + governed compute + a federated flywheel form a moat that’s hard to close. Capital flows follow the stack – M&A to acquire federated tooling and AI-ready data assets; valuation premiums accrue to teams that can prove model performance, provenance, and governance. Talent mixes shift toward model engineers, ML-ops, and data stewards who can run audit-ready AI workflows. Beyond R&D, the same infrastructure lifts manufacturing yield and QC and powers more precise commercial execution. In my mind and through the work I have been doing the throughline is simple: this is not a “cute database.” It’s an operating-model change that can supercharge the company. Firms that can execute now will compound; those that don’t risk getting lapped to the point where catching up is no longer possible. My two cents – hope it was helpful and gives you a different lens to look through.
