NVIDIA and Eli Lilly: the AI revolution in drug discovery explained

01The NVIDIA-Lilly partnership and its goals

The partnership between NVIDIA and Eli Lilly, formalized in 2024 and expanded in 2025, represents one of the most significant collaborations between a chip manufacturer and a pharmaceutical company. Nvidia Corporation is an American multinational technology company headquartered in Santa Clara, California. The company develops graphics processing units (GPUs), systems on chips (SoCs), and application programming interfaces (APIs) for data science, high-performance computing, artificial intelligence (AI), and mobile and automotive applications. Founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem, Nvidia has been widely described as a Big Tech company. The deal gives Lilly access to NVIDIA's BioNeMo platform — a cloud-based service for training and deploying AI models in drug discovery — along with dedicated DGX SuperPOD computing infrastructure at Lilly's research facilities.

The stated goal is ambitious: reduce the time from target identification to clinical trial candidate from the industry average of 4-5 years to 18-24 months. Lilly has committed $500 million to the partnership over five years, deploying AI across its entire drug discovery pipeline. The collaboration focuses on three areas: generative AI for novel molecule design, AI-accelerated molecular dynamics simulation, and predictive modeling of drug-target interactions.

The partnership is not exclusive in either direction. NVIDIA has similar agreements with Amgen, Genentech, and several biotech startups. Lilly also partners with AI drug discovery companies including Schrödinger, Insilico Medicine, and Atomwise. The NVIDIA-Lilly deal is distinguished by its scale and the depth of infrastructure integration — Lilly has essentially built an AI-first drug discovery organization on top of NVIDIA's hardware and software stack.

02How GPU computing accelerates molecular simulation

Molecular simulation is the computational backbone of drug discovery. To predict whether a drug molecule will bind to a disease target (typically a protein), researchers simulate the physical interactions between the molecule and the target at the atomic level. These simulations, known as molecular dynamics (MD), calculate the forces between every atom in the system at each time step — a computation that scales cubically with the number of atoms.

Drug design, often referred to as rational drug design or simply rational design, is the inventive process of finding new medications based on the knowledge of a biological target. The drug is most commonly an organic small molecule that activates or inhibits the function of a biomolecule such as a protein, which in turn results in a therapeutic benefit to the patient. In the most basic sense, drug design involves the design of molecules that are complementary in shape and charge to the biomolecular target with which they interact and therefore will bind to it. Drug design frequently but not necessarily relies on computer modeling techniques. This type of modeling is sometimes referred to as computer-aided drug design. Finally, drug design that relies on the knowledge of the three-dimensional structure of the biomolecular target is known as structure-based drug design. In addition to small molecules, biopharmaceuticals including peptides and especially therapeutic antibodies are an increasingly important class of drugs and computational methods for improving the affinity, selectivity, and stability of these protein-based therapeutics have also been developed. Traditional CPU-based MD simulations can handle systems of roughly 100,000 atoms at a timescale of microseconds per day of compute. GPU acceleration, particularly with NVIDIA's CUDA platform, has increased this by orders of magnitude. A single NVIDIA H100 GPU can simulate systems of over 1 million atoms at millisecond timescales per day — a 1000x improvement over CPU-based approaches. A DGX SuperPOD with 576 GPUs multiplies this further.

The practical impact is that researchers can now simulate drug-target interactions that were previously computationally intractable. Membrane proteins, which are targets for approximately 60% of approved drugs but are notoriously difficult to study experimentally, can be simulated in their native lipid bilayer environment. Large protein complexes, including antibody-antigen interactions for vaccine design, are now accessible to computational study. This doesn't replace experiments — but it dramatically narrows the search space before costly wet-lab work begins.

03AI models for protein folding and drug targeting

Protein structure prediction has been revolutionized by AI. DeepMind's AlphaFold2, released in 2020, demonstrated that neural networks could predict protein 3D structure from amino acid sequence with near-experimental accuracy. By 2026, AlphaFold3 and open-source alternatives like ESMFold and RoseTTAFold can predict not just protein structure but protein-ligand, protein-protein, and protein-nucleic acid interactions.

NVIDIA's BioNeMo platform incorporates these models and extends them. It includes generative models that design novel drug molecules optimized for specific targets — not just screening existing compound libraries but creating new molecules predicted to bind with high affinity and selectivity. These generative models can explore chemical space far beyond what exists in nature or pharmaceutical compound libraries, potentially discovering entirely new classes of drugs.

The limitation is that AI predictions require experimental validation. A model can predict that a molecule will bind a target, but only a wet-lab assay can confirm it. The bottleneck has shifted from "which molecules to test" to "how to validate predictions at scale." High-throughput screening facilities, automated by robotics, can test thousands of AI-predicted candidates per week — but even this is too slow for the pace at which AI can generate candidates. The gap between computational prediction and experimental validation remains a key constraint.

04The infrastructure behind AI drug discovery

The computational infrastructure for AI drug discovery is staggering. Lilly's installation includes a dedicated NVIDIA DGX SuperPOD — 576 H100 GPUs interconnected with NVIDIA InfiniBand, delivering approximately 8 exaFLOPS of AI performance. This is comparable to a top-100 supercomputer, installed not at a national lab but at a pharmaceutical company. The system runs 24/7 on molecular simulations, protein folding predictions, and generative chemistry models.

The software stack is equally important. NVIDIA's BioNeMo provides pre-trained models (AlphaFold, ESMFold, generative chemistry models) as cloud services, along with frameworks for fine-tuning on proprietary data. Lilly's computational chemists can access these tools through APIs, integrating AI predictions into their existing workflows. The platform also includes MoleculeNet benchmarks for evaluating model performance and OpenMM for molecular dynamics simulation.

The cost of this infrastructure is substantial. A single H100 GPU costs approximately $30,000, and a full SuperPOD installation runs into hundreds of millions of dollars. The power consumption is equally significant — each H100 draws 700 watts, meaning the SuperPOD requires over 400 kilowatts of power and sophisticated cooling. This is why the partnership model makes sense: few pharma companies can justify this investment alone, and NVIDIA amortizes its R&D across multiple customers.

05From discovery to clinical trials faster

The ultimate measure of AI drug discovery is whether it produces approved drugs faster. As of 2026, the first AI-discovered drugs are entering Phase II clinical trials. Insilico Medicine's INS018_055, an AI-designed drug for idiopathic pulmonary fibrosis, entered Phase II in 2024 — just 30 months from target identification to clinical trial, compared to the industry average of 4-5 years. Exscientia's DSP-1181, an AI-designed obsessive-compulsive disorder drug, reached Phase I in 2023.

Lilly's AI-accelerated pipeline has produced several clinical candidates in oncology and immunology that entered Phase I trials in 2025-2026. The company reports that AI-designed candidates have shown higher hit rates in preclinical testing — approximately 15% of AI-predicted compounds show desired activity, compared to 2-5% for traditional high-throughput screening. This higher hit rate reduces the number of compounds that need to be synthesized and tested, saving both time and cost.

06What this means for pharmaceutical competition

The AI drug discovery race is reshaping pharmaceutical competition. Companies that adopt AI early gain a structural advantage: faster discovery cycles, higher hit rates, and the ability to pursue targets that were previously intractable. This creates a compounding effect — more AI-discovered candidates entering the pipeline means more data feeding back into the models, improving their accuracy and further accelerating discovery.

The competitive landscape is splitting. Large pharma companies like Lilly, Novartis, and Roche are building internal AI capabilities through partnerships with NVIDIA, Google, and specialized AI companies. Biotech startups like Insilico Medicine, Recursion Pharmaceuticals, and Exscientia are built AI-first from the ground up. Traditional pharma companies that are slow to adopt AI risk being outcompeted on both speed and novelty of drug candidates.

The implications extend beyond speed. AI can explore chemical space that human chemists would not consider, potentially discovering drugs for targets that have been considered "undruggable." If AI-discovered drugs reach the market with novel mechanisms of action, the patent landscape could shift dramatically. Companies with strong AI capabilities may file patents on entire classes of computationally-designed molecules, creating intellectual property moats that are difficult for competitors to cross.

07When AI-discovered drugs reach patients

The first AI-discovered drugs will likely reach patients in 2028-2030, assuming current clinical trials succeed. This timeline is set by the clinical trial process, not by AI — Phase II and III trials take 3-7 years regardless of how the drug was discovered. What AI changes is the front end: getting a viable candidate into trials faster, and potentially improving the quality of candidates so that fewer fail in later stages.

The attrition rate in drug development is brutal. Approximately 90% of drugs that enter Phase I fail to reach approval, with the highest failure rates in Phase II (efficacy) and Phase III (safety and efficacy at scale). If AI can improve the quality of candidates entering trials — by better predicting efficacy and safety from molecular properties — the approval rate could increase. Even a modest improvement from 10% to 15% approval rate would represent billions in saved development costs and more drugs reaching patients.

The NVIDIA-Lilly partnership is a bet that computational power, applied systematically across the drug discovery pipeline, can fundamentally change the economics of pharmaceutical innovation. If it succeeds, the model will spread. If the first AI-discovered drugs fail in late-stage trials, enthusiasm will cool. The next 3-4 years of clinical trial data will determine whether AI drug discovery is a genuine paradigm shift or a powerful tool with more modest impact than its proponents claim.