The conversation around AI and drug discovery tends to focus on the wrong things. Compute power, frontier models, organizational adoption — these are real levers, but they miss the fundamental nature of what pharmaceutical research actually is. At its core, drug discovery is not an industrial process. It is an idea business. And until the industry internalizes that distinction, AI will remain a productivity tool rather than a transformational force.
The Paul Janssen Problem
Paul Janssen — founder of Janssen Pharmaceutica, later acquired by Johnson & Johnson — personally invented somewhere between half a dozen and a dozen critical medicines, including the first antipsychotic and fentanyl. There are statues of this man. He represents something the industry has largely lost: the singular, creative scientific mind generating breakthrough ideas from first principles.
Modern pharma has replaced that model with committees. Fifteen to twenty departments — toxicology, pre-clinical pharmacokinetics, project management, regulatory — all required to communicate with each other, and all consistently failing to do so cleanly. Project managers, theoretically the connective tissue, often make coordination worse. The result is a system where a CEO kills a promising drug for no reason, a chief scientist pushes a compound because of a personal rivalry, and hundreds of millions of dollars flow toward decisions driven by ego, inertia, and Wall Street optics rather than data.
The most damning estimate: up to half of all drugs currently in clinical trials probably shouldn't be there. That is not a compute problem. That is a decision-making problem.
Where AI Actually Fits
The highest-leverage application of AI in pharma is not replacing chemists or filing paperwork faster. It is augmenting the idea-generation layer — the equivalent of having a digital Paul Janssen who can read every published paper, identify underexplored mechanisms, and propose hypotheses that a single human mind would never reach.
Once you improve the idea pipeline, everything downstream becomes cheaper. Fewer bad drugs enter clinical development. Fewer phase two and phase three trials get run on compounds that a clear-eyed data review would have killed in phase one. The savings are not in headcount — they are in not spending $400 to $500 million proving that a drug doesn't work when the signal was already weak at the start.
The question of whether the bottleneck is raw compute scale, smarter models, or organizational adoption is somewhat beside the point. Pharma's inertia is cultural and structural. A more capable AI that a CEO never deploys meaningfully does nothing. The real unlock is pushing more decision-making authority downward and using AI to hold those decisions accountable to data rather than politics.
The Clinical Trial Cost Problem Won't Be Solved by Efficiency Alone
Drug development economics are brutal and largely misunderstood. Inventing a drug might cost $10 to $50 million. Getting it approved costs four to ten times that — almost entirely because of clinical trials. Recruiting 10,000 patients for a diabetes study requires incentivizing physicians and patients financially, paying data collection organizations, paying for regulatory assembly, and managing a process that is, in large parts, still run on pen and paper.
China has meaningfully lower costs for this work, partly because wages for clinical trial staff are lower, and partly because capital markets dynamics allow Chinese biotechs to deploy the same fundraise — say, $300 million — very differently. A $300 million raise in San Diego buys roughly 40 chemists. The same amount in China buys 800. The structural output per dollar is not comparable.
This creates a legitimate competitive threat. Chinese drug development capability has advanced to a point where the marginal value of Western biotech infrastructure is being questioned. The industry already shows signs of overcapacity — multiple companies developing drugs for patient populations of a few hundred people, fighting over the same thin slices of rare oncology indications.
The Regulatory Gray Zone: Peptides and Unapproved Compounds
A parallel conversation surrounds the boom in self-administered peptides — technically small proteins — that sit in a regulatory no-man's-land. Some of these compounds have real pharmacological effects. Some are manufactured under conditions of unknown quality. The FDA has no framework to approve many of them because the agency doesn't classify their target conditions as diseases.
A representative example: dapoxetine, a compound that demonstrably works for premature ejaculation, was rejected by the FDA not because it lacks efficacy but because the agency declined to classify that condition as an illness warranting drug approval. It is approved in Europe. American patients who want it are left to source it informally.
The deeper structural problem is that many peptides and off-patent compounds will never have formal clinical trials because the economics don't support it. No single company can justify spending tens of millions on trials for a molecule anyone can manufacture — all competitors benefit from the data without bearing the cost. The workaround the industry typically uses — modifying the molecule enough to generate a new patent — exists, but it requires active pharmaceutical development intent.
This creates a genuine policy gap. The libertarian argument says informed adults should be able to access compounds with reasonable evidence profiles. The safety argument says uncontrolled sourcing, no dose standardization, and no adverse event tracking are real risks. Neither side has fully resolved the tension, and the regulatory framework hasn't caught up.
The Structural Ceiling on Pharma Returns
Even the most successful drug companies face a fundamental constraint that no tech company tolerates: a hard expiration date on their best assets. A blockbuster drug generates revenue for roughly ten years before generics erode the market. Eli Lilly's GLP-1 franchise is the current exception, partly because the manufacturing complexity delays generic competition. But the rule holds.
No AI company would be worth what AI companies are worth if their core models entered the public domain a decade after release. That asymmetry — between software compounding and pharmaceutical patent cliffs — is structural, not cyclical. It shapes how capital flows into the industry, how companies make R&D bets, and why even well-run biotechs tend to be one-hit wonders rather than compounding platforms.
AI can improve idea generation, reduce failed trial spending, and make organizational decision-making less irrational. What it cannot do is change the fundamental economics of a regulated industry built around time-limited monopolies. That ceiling remains.
