Now Live: The LillyPod AI supercomputer Sets a New Bar for Pharma R&D
Eli Lilly has switched on a new AI factory, built with NVIDIA, that it says is the world’s most powerful supercomputing infrastructure wholly owned and operated by a pharmaceutical company. Dubbed LillyPod, the system runs 1,016 NVIDIA Blackwell Ultra GPUs and delivers more than 9,000 petaflops of AI performance—purpose‑built to accelerate AI for drug discovery, development, and delivery across regulated workflows [1][2]. Lilly asserts that owning end‑to‑end infrastructure enables faster iteration and tighter risk control in high‑stakes drug design [1][2].
Inside the LillyPod AI supercomputer: Key specifications and architecture
LillyPod is the first NVIDIA DGX SuperPOD configured with DGX B300 systems and integrates NVIDIA’s full‑stack AI factory architecture—accelerated computing, an optimized AI software stack for healthcare and life sciences, and Spectrum‑X high‑performance Ethernet networking [1][2]. Its unified high‑speed fabric carries GPU‑to‑GPU, storage, and general system communication on one network. Physical buildout required nearly 5,000 connections and more than 1,000 pounds of fiber cabling, and the system uses efficient liquid cooling to limit incremental energy impact [1].
Assembly took roughly four months from start to finish. NVIDIA’s Spectrum‑X networking underpins end‑to‑end performance across training and inference, while the DGX B300 architecture provides the scale to move from pilot workloads to production‑grade research [1]. For readers seeking the vendor perspective, see NVIDIA’s announcement (external) of Lilly’s AI factory, including platform details and partner context [1].
What it enables: Workloads and use cases
The AI factory is designed to support the full lifecycle of pharmaceutical R&D: large‑scale data ingestion, training and fine‑tuning of foundation and domain‑specific models, and high‑throughput inference. Targeted workloads include genome‑scale analyses, modeling of biological systems, and training of foundation models on Lilly’s proprietary experimental and clinical data—aimed at accelerating discovery, development, and delivery to patients [1][2].
For technical teams, the combination of Blackwell Ultra GPUs, DGX B300 systems, and Spectrum‑X networking is designed to shorten iteration loops and scale from exploratory analysis to production inference without moving sensitive data offsite [1][2].
Operational tools and governance
NVIDIA Mission Control manages the DGX SuperPOD, orchestrating workloads, monitoring performance, and automating AI operations. These capabilities are positioned to help operate securely and in a compliant manner, a priority for regulated healthcare and life sciences workflows [1][2]. Lilly notes that owning the infrastructure end‑to‑end improves control over data governance and model risk in drug design [1][2].
Where it fits: Comparisons and context
Among pharma‑owned systems, LillyPod roughly doubles the GPU count of Recursion’s BioHive‑2 (504 H100 GPUs). It would rank among the most powerful commercial systems, though it remains below government exascale supercomputers in overall scale [1][3]. For pharmaceutical teams, the practical difference is the ability to train larger foundation models on proprietary data and run higher‑throughput inference under one roof—reducing bottlenecks that can slow discovery programs [1][2].
Sustainability, build timeline, and operations
LillyPod employs liquid cooling to curb incremental energy demand, and Lilly plans for the AI factory to run on 100% renewable electricity by 2030 [1][2]. The cluster’s physical integration—nearly 5,000 connections and over 1,000 pounds of fiber cabling—was completed in about four months, reflecting a fast path from design to deployment for a system of this scale [1].
For enterprises evaluating an on‑premise build, the LillyPod approach underscores the potential benefits of a unified network fabric and an integrated software stack to reduce operational friction across training, fine‑tuning, and inference in regulated settings [1][2].
What this means for pharma, AI teams, and investors
For executives: the LillyPod AI supercomputer represents a strategic shift toward enterprise‑owned AI factories that consolidate compute, data, and operations for sensitive R&D. The aim is faster iteration, lower latency to insight, and greater control over IP and compliance [1][2].
For engineers and operators: DGX B300 systems with Blackwell Ultra GPUs, Spectrum‑X high‑performance Ethernet, and NVIDIA Mission Control lay out a reference pattern for scaling model training and inference while maintaining centralized observability and governance [1][2].
For analysts and investors: relative to other pharma systems, LillyPod’s scale indicates a competitive push to bring foundation model training onto proprietary data and to industrialize high‑throughput inference—key levers for pipeline velocity and asset quality [1][3].
If you’re mapping your own roadmap, consider how an enterprise‑owned AI factory may complement cloud resources for sensitive workloads and how unified networking and orchestration can simplify regulated operations [1][2]. To go deeper on practical implementation patterns, Explore AI tools and playbooks.
Quick specs and resources
- First NVIDIA DGX SuperPOD with DGX B300 systems owned and operated by a pharmaceutical company [1][2]
- 1,016 NVIDIA Blackwell Ultra GPUs; >9,000 petaflops of AI performance [1][2]
- NVIDIA Spectrum‑X unified fabric for GPU, storage, and system traffic [1]
- Mission Control for orchestration, monitoring, and secure, compliant AI operations [1][2]
- ~4‑month assembly; ~5,000 connections; >1,000 lbs of fiber; liquid cooling [1]
- 100% renewable electricity targeted by 2030 [1][2]
- Positioning: Larger than Recursion BioHive‑2 (504 H100 GPUs); below government exascale systems [1][3]
For official details and announcements, see NVIDIA’s blog post on Lilly’s AI factory (external) [1].
Sources
[1] Now Live: Lilly AI Factory for Pharmaceutical Discovery …
https://blogs.nvidia.com/blog/lilly-ai-factory-live/
[2] Lilly partners with NVIDIA to build the industry’s most powerful AI …
https://www.prnewswire.com/news-releases/lilly-partners-with-nvidia-to-build-the-industrys-most-powerful-ai-supercomputer-supercharging-medicine-discovery-and-delivery-for-patients-302597285.html
[3] Lilly building pharma’s most powerful AI supercomputer with Nvidia
https://www.rdworldonline.com/lilly-says-its-building-pharmas-most-powerful-ai-supercomputer-with-nvidia/
