Elias Tharakan posted two numbers side by side this week, and the gap between them is the most important story in biopharma right now. Since January 2026, pharmaceutical companies have committed more than $7 billion to AI drug discovery partnerships with Insilico Medicine alone, including deals with Servier in oncology, Eli Lilly for AI-designed oral therapeutics, SK Biopharmaceuticals in neuroimmune disease, and Takeda across multiple therapeutic areas. Approved AI-discovered drugs today: zero. That arithmetic is not a scandal. But the industry’s failure to interrogate what it reveals very nearly is.
The numbers look contradictory. They are not. They mark exactly where AI’s proven value ends and where the next, harder problem begins. Insilico’s rentosertib, the first fully AI-discovered and AI-designed molecule to publish positive Phase IIa efficacy results in a peer-reviewed study, is a genuine milestone. It is also a molecule that still must navigate Phase IIb, Phase III, manufacturing validation, and a full regulatory review. Every dollar committed to its discovery has bought the industry a candidate, not a drug. And the cost of converting candidates into drugs, per Tharakan’s post, remains approximately $2.6 billion per approved product, with only about 12% of molecules entering clinical trials ever reaching patients.
The capital allocation has not caught up to that arithmetic.
Optimizing Keys, Ignoring Locks
Milad Alucozai framed the structural flaw precisely this week, and his framing deserves to be read slowly. He invoked Lipinski’s Rule of 5, the classic molecular filter that has guided medicinal chemistry for decades. Lipitor, Gleevec, and Cyclosporine all violate it. Three of the most consequential drugs in modern pharmacology would have been screened out by the same logic now being automated at billion-dollar scale. Alucozai’s conclusion: “Filters eliminate garbage. They don’t create gold.” AI drug discovery, trained on historical molecular data, is a more powerful filter. The ceiling on what a better filter can do has not changed.
His sharper point concerns biology. Breakthrough drugs work, he argues, because they do something unexpected in a living human system. Train a model on prior success and you get a system that biases toward consensus, one that is constitutionally incapable of producing biological black swans. The real uncertainty in drug development was never whether chemists could make an interesting molecule. It was always whether that molecule would survive contact with human biology, a regulatory agency, and a clinical trial infrastructure running at 80% enrollment delay rates and costing sponsors $55,000 per day in late-stage delays, per Alucozai’s post.
Those operational figures are where the argument pivots from interesting to urgent. The industry has spent the better part of five years building what Alucozai calls “an infinite library of beautifully bound books” while the system that determines whether anyone ever reads them remains largely unreformed.
The Infrastructure Gap Nobody Is Funding
Dr. Guy Stephens identified the specific technical reasons why clinical development has remained AI-resistant even as discovery has not. His post, published in response to Anthropic’s Claude Science release, cuts to the two structural blockers that capital alone cannot solve. First, pharmaceutical companies hold decades of trial data that is neither standardized nor model-ready. The population-scale sources that could give AI genuine predictive power in development, biobanks, real-world evidence repositories, integrated EHR networks, represent a second data layer that current foundation model platforms do not yet reach. Sponsors who have run fifty Phase II trials do not automatically have fifty trials’ worth of usable training data. They have fifty silos.
The second blocker is trust sequencing. In clinical development, a failed pivotal trial can wipe hundreds of millions in pipeline value in a single readout. New tools earn trust in low-stakes pilot settings first. This is not conservatism for its own sake. It is a rational institutional response to a failure mode with asymmetric consequences. The enterprise-deployment model that governs most software adoption does not translate to an environment where the cost of a wrong decision is not a bad quarter but a dead program.
Stephens’ framing of Claude Science is worth taking seriously here. The auditable output architecture that Anthropic has built into the platform represents a meaningful step toward the regulatory requirements that would need to be satisfied before AI contributes to clinical decision-making in any formal sense. But a step toward regulatory readiness is not regulatory readiness. The FDA has not issued guidance that would permit an AI-generated clinical recommendation to substitute for investigator judgment in a GCP-regulated trial. The gap between “useful workflow tool” and “accountable clinical decision support” remains wide, and the roadmap to close it runs through data standardization and sequenced adoption pilots, not through discovery partnerships.
Where the Next Race Will Be Run
The counterintuitive read on the $7 billion committed to Insilico and its peers is that it validates clinical development AI more than it validates discovery AI. Pharma companies have now publicly confirmed, through their own capital allocation, that they believe AI can find better molecules faster. If that belief is correct, the pipeline pressure flowing toward clinical development will intensify, not ease. More candidates entering Phase I means more candidates competing for site capacity, for enrollment bandwidth, for regulatory review time. A faster discovery engine attached to an unreformed development engine does not accelerate drug approval. It accelerates congestion.
Alucozai’s operational math is the right frame: enrollment forecasting, site optimization, and workflow automation are not moonshots. They are infrastructure plays that compound across every program in a portfolio. A sponsor running fifteen concurrent Phase II trials does not need fifteen separate AI discoveries. It needs one enrollment prediction model that works reliably across all fifteen. The ROI calculation on clinical operations AI is portfolio-level, not asset-level, which means the business case is arguably stronger than the business case for any single discovery partnership, even a $6.6 billion one.
Tharakan closed his post with the question the industry has not yet answered at scale: “If the industry is willing to invest $7 billion to discover better molecules faster, when will it invest at that scale to help those molecules reach patients faster?” Rentosertib will provide a partial answer. It is now inside the development machine. The machine has not changed. And $2.6 billion, a 12% success rate, and $55,000 daily burn rates are still waiting on the other side.
References
- Elias Tharakan, LinkedIn post on AI drug discovery partnerships and Insilico Medicine, 2026. https://www.linkedin.com/posts/eliastharakan_artificialintelligence-drugdiscovery-clinicaldevelopment-activity-7479126910111932416-bWs7
- Milad Alucozai, LinkedIn post on AI drug discovery valuation and clinical trial bottlenecks, 2026. https://www.linkedin.com/posts/miladalucozai_ai-biotech-drugdiscovery-activity-7478106106741936128-p5q_
- Dr. Guy Stephens, LinkedIn post on Claude Science and AI in clinical development, 2026. https://www.linkedin.com/posts/guy-stephens_clinicaldevelopment-artificialintelligence-activity-7478392929410236417-7Rn1
Moe Alsumidaie, MBA, MSF, is founder and Chief Editor of Vanguard Publications, which publishes Clinical Trial Vanguard, Pharma Vanguard and BullScope, and Head of Research at CliniBiz. He has two decades in clinical trial operations and data science, with earlier roles at Genentech, Abbott Vascular and Stanford University Medical Center, and is a guest lecturer in clinical trial sciences at Rutgers University.
