Novo Nordisk announced a multi-year partnership with OpenAI on April 14, embedding the lab’s models across drug discovery, clinical trials, manufacturing, supply chain, and commercial operations. Pilots have started. Full integration is targeted for the end of 2026. The framing across most of the press was that AI is about to compress the next obesity blockbuster into clinics years sooner. That framing is partially right and partially off, and the interesting question is which parts.
A drug timeline is several pipelines stitched together. The part that AI demonstrably accelerates today is the front end. The part that determines when a drug actually ships is mostly somewhere else.
What AI changes at the front of the pipeline
The front end of drug discovery is a sequence of decisions. A target has to be picked, a molecule has to be generated, the molecule has to be filtered for absorption, distribution, metabolism, excretion, and toxicity, and a small set of candidates has to be selected for animal and then human testing. Until recently, each of those decisions was bounded by laboratory throughput and chemist time. They are now substantially bounded by compute and model quality, which is a different bottleneck with a different scaling curve.
Target identification is the easiest case. The set of plausible biological targets for a given disease is large, and the search across it is naturally combinatorial. Models trained on protein structure, expression data, and clinical-trial outcomes can produce ranked target lists that would have taken a team of biologists months to assemble. The output is not a decision; it is a starting set. The team still has to pick.
Molecular generation is the noisier case. Generative models can produce candidate compounds that satisfy a target specification, and the volume of candidates per unit time is several orders of magnitude higher than what wet-lab medicinal chemistry produces. The harder problem is filtering. A novel compound has to clear in-silico predictions for binding affinity, off-target activity, metabolism, and toxicity before it becomes worth synthesizing. The predictions are imperfect but useful, and the imperfection compounds across the filter stack. The teams that consistently get good leads out of this pipeline have spent five or more years tuning their filters.
The Insilico Medicine partnership Eli Lilly signed in March sits squarely in this category. The deal is for AI-discovered molecules, and the company has been refining its discovery-to-IND pipeline since 2020. Its public timeline numbers show roughly thirty months from target to IND for the cases that have hit the milestone, against an industry baseline of four to six years.
Numbers like that are why the press framed the Novo announcement the way it did. The framing skips a step.
What sits between a candidate and a shipped drug
The thirty-month number ends at IND, which is the regulatory permission to start human trials. The drug is then somewhere in Phase 1, which is short and small, and the rest of the timeline is dominated by what comes after.
Phase 2 trials for chronic-disease drugs run twelve to twenty-four months and are gated by patient enrollment, not by anything a model can predict. Phase 3 trials for an obesity indication run twenty-four to thirty-six months, gated by enrollment scale and by the duration of follow-up the FDA requires. Manufacturing scale-up for a biologic like a GLP-1 receptor agonist takes eighteen to twenty-four months, runs in parallel with late-stage trials, and is its own engineering problem with its own tail of unknowns. Regulatory review after a submission takes six to ten months on a standard track. Sum the tail honestly and a molecule that clears IND in 2026 has at least five years of work between it and a shipped drug.
None of that is shaped like discovery. Enrollment timelines are bounded by how many qualifying patients exist, how many sites can be run, and how fast those sites can recruit. Manufacturing scale-up is an engineering problem inside a regulated facility. FDA review is a regulatory queue. Models help at the margins in each of these, but the marginal months are months, not years.
The Novo deal does cover trial-cohort selection and manufacturing optimization. The honest read of what those integrations will compress is on the order of months across a multi-year tail. That is real value. It is not the same as cutting years off a blockbuster.
The two shapes of an AI deal in pharma
Looking at Novo with OpenAI alongside Lilly with Insilico, the structural difference between them is more interesting than the head-to-head framing the press settled on.
The Lilly deal is vertical. It is an integration in a single function, drug discovery, with a partner whose entire product is that function. The expected return is concentrated in one part of the pipeline, the front end, and shows up in a small number of high-value bets.
The Novo deal is horizontal. It is an integration across every function in a large operating company, with a partner whose models are general-purpose. The expected return is diffuse, modest in any single integration, and aggregates across a company that does roughly thirty billion dollars of revenue annually. The pattern is closer to what enterprise AI integration looks like in any large incumbent than to what AI-first biotechs do.
These patterns have different operational tails. Vertical integrations get judged on the success rate of the molecules they generate. Horizontal integrations get judged on cost reductions across operations, fewer stockouts in supply chains, faster trial-cohort matching, and small productivity improvements in functions that already employed thousands of people. The first is the story journalists prefer. The second is the story that shows up in margin numbers.
What this probably changes first
The most plausible near-term effects of the Novo deal are in manufacturing forecasting and supply chain. Ozempic and Wegovy have been in chronic shortage since 2022, which is a demand-prediction and capacity-allocation problem with a lot of headroom for better models. Fewer empty pharmacy shelves is a less photogenic outcome than a faster blockbuster, but it is the one that will show up first and the one that will matter to people on prescriptions today.
Trial-cohort matching is the next-most plausible. The work of finding qualifying patients across a global trial network is information-retrieval work with a slow tail, and faster matching can pull months out of a Phase 2 enrollment window. Multiply across the dozen-plus active trials Novo runs at a time and the months add up.
Discovery acceleration, if it works at all, shows up in clinics in the early 2030s. The horizon is too far to read confidently from where we sit. The deal will have to renew at least once for that compression to land in patients, and the integration may have changed shape several times by then.
The next obesity blockbuster is not going to arrive years sooner because of this partnership. It may arrive close to on schedule, with a more predictable supply chain, at margins that fund the next set of bets. The horizontal AI deal inside an incumbent looks unglamorous from outside and produces results that show up in operations long before they show up in clinics. The vertical AI deal at a discovery-first biotech is louder and concentrates its bets. We will probably see more of both for the rest of the decade, and most of the running coverage will be about the loud one.