Rentosertib, a candidate against idiopathic pulmonary fibrosis, reached the preclinical candidate stage in approximately 18 months. The comparison published by Insilico addresses approximately 4.5 to 6 years required to nominate a preclinical candidate through the traditional pathway, without establishing the precise alleged costs of 6 million or 150 million dollars. This gain is real, documented, and it changes something fundamental in industrial biology. But the beneficiaries of this cost compression are not who one might imagine: computing costs, proprietary data, and patents are already reconstituting a barrier that AI had just lowered.
The essentials
- According to Axis Intelligence, 173 clinical programs analyzed and over 200 AI medicines are in development in 2026.
- GPU computing costs constitute a significant entry barrier for academic laboratories, which recreates a structural entry barrier.
- The market is presented as fragmented; no data consulted shows that five to ten actors capture the bulk of gains. Academic laboratories and low-income countries face significant access barriers linked to computing costs, proprietary data, and intellectual property.
- On the 2030–2035 horizon, patent reform and growing public funding of GPU infrastructure could redistribute a significant share of these gains.
Rentosertib as proof of concept
Rentosertib was discovered and developed with support from Insilico Medicine’s Pharma.AI platform and targets idiopathic pulmonary fibrosis. Its development employed machine learning models; Insilico indicates that the program synthesized and tested 78 molecules. Development lasted approximately 18 months until the preclinical candidate stage; no audited cost of 6 million dollars is established by the source.
A growing number of drugs developed with AI tools are in clinical trials. The global market for AI-assisted drug discovery is projected at approximately 13.8 billion dollars in 2033, according to Grand View Research.
A few names appear consistently: Recursion, Insilico Medicine, Exscientia, and the AI divisions of AstraZeneca, Pfizer, or Roche. The promise of open science collides with structural access barriers: computational costs, proprietary data, and concentration of intellectual property.
The entry barrier survives cost compression
The logic seemed simple: if certain development stages cost less, actors capable of conducting them should be more numerous. In practice, structural barriers persist.
GPU computing costs for training pharmaceutical foundation models constitute a significant entry barrier for academic laboratories without dedicated funding. It is trivial for Pfizer or AstraZeneca, which amortize it across pipelines of several hundred candidate molecules. Cost compression benefits first those actors already having access to significant computational and informational resources.
Then comes the question of data. AI models learn on massive biological databases: results from clinical trials, genomic sequences, protein structures. A significant number of high-quality biological databases are proprietary and held by major pharmaceuticals. An academic laboratory works with public data, PubChem, UniProt, ChEMBL, which are vast but less precise for the most recent molecules. The informational advantage precedes the computational advantage.
Patents add a third layer. A competitor using a method covered by patent claims can face infringement litigation—exactly the kind of risk a startup without significant resources cannot absorb.
Five actors, two hundred medicines
Medicines in clinical trials discovered by AI are being developed by several major actors. Recursion is among the leading actors; Insilico Medicine develops drug candidates including rentosertib, targeting idiopathic pulmonary fibrosis. Exscientia, acquired by Recursion in 2024, brings its own oncology candidates.
Major pharmaceuticals combine external partnerships and internal development of capabilities, including AI models. AstraZeneca concluded agreements with Tempus and Pathos, while Recursion concluded a separate agreement with Tempus. Pfizer concluded a research collaboration with Insilico Medicine in 2020, and Merck develops capabilities in-house. These partnerships structure access to the drug discovery pipeline.
This movement resembles other technological concentrations observed in other sectors. On the theme of concentration induced by AI in industries where the promise was decentralization, one can read the analysis published in this journal regarding AI in Africa and the electricity lock-in: the mechanism of barrier reconstruction is similar there, even if the sector differs. The question is posed here with particular urgency because medicines, unlike robots or translation models, directly condition lives.
Low-income countries facing the oligopolistic pipeline
A drug discovered 25 times cheaper does not cost 25 times less to purchase. The economy realized upstream does not automatically transfer to lower prices downstream.
For low- or middle-income health systems, this dynamic creates a concrete problem. The Drugs for Neglected Diseases initiative and similar organizations played this rebalancing role in the 2000–2010 period. They will need to adapt to an environment where intellectual property is even more fragmented.
Unequal access to AI-discovered medicines constitutes a major regulatory issue. No alternative framework has yet been adopted.
This challenge is not unprecedented. Thailand and Brazil used or brandished compulsory licenses for antiretrovirals; India primarily used other flexibilities. Generic competition reduced certain HIV treatment prices, with documented supply at 350 dollars per year in 2001. The question is whether an equivalent mechanism can apply to an era where intellectual property extends to algorithmic methods.
The 2030–2035 horizon according to choices made now
Three trajectories are taking shape. Their respective credibility depends on political and regulatory decisions beginning to be made now.
The first is the consolidation of discovery capabilities. Agreements between contractual partners concentrate access among them and do not allow establishing the existence of a general access redistribution mechanism. Recursion, Insilico, and other companies are important actors in AI applied to drug discovery; Exscientia was further integrated into Recursion. Access for low-income countries would depend on intellectual property regimes and commercial terms to be defined. Academic laboratories without access to GPU infrastructure face significant obstacles to conducting competitive discovery programs. This scenario extends current trends without institutional rupture.
It is the most likely scenario in the short term.
The second trajectory stems from license reform. Since 2024, the WHO has pursued technical cooperation on existing flexibilities in the TRIPS agreement; a project to expand TRIPS agreements is not documented. The NIH claims an essential financing role in the scientific foundations of AlphaFold; Wellcome Trust financing is explicitly declared for work related to the AlphaFold database. Non-exclusive licenses granted before patent expiration can improve access and competition in low- or middle-income countries, but their timeline and effects on innovation incentives depend on the conditions of each agreement. The Medicines Patent Pool enabled a significant number of low- and middle-income countries to access generic HIV and hepatitis C medicines.
Extending it to the AI medicine era would require political will that current WHO negotiations have not yet produced.
The third trajectory is that of shared public infrastructure. The U.S. NIH and the European Commission have engaged still-embryonic reflections on funding GPU capacity accessible to academic laboratories. In Europe, EuroHPC already provides high-performance computing for research; its explicit extension to computational biology is technically feasible. In the United States, several bills have been introduced to create subsidized access to biological foundation models for public universities. None have yet been adopted.
If these initiatives succeed, they could maintain competitive academic research and provide an open base on which middle-income countries would build their own capabilities, similar to what adopting AI without local training produces as dependence in other sectors.
These three scenarios are not mutually exclusive. Consolidation can coexist with compulsory licenses on certain medicines, and partial public infrastructure can reduce without eliminating entry barriers. The signals to monitor are precise: the rate of FDA and European Medicines Agency approval of the first medicines entirely discovered by AI, the market share of the five leading AI pharma actors in two years, and the volume of public funding allocated to GPU infrastructure open to academics.
Intellectual property, the nerve of the next battle
Rentosertib is an example of a drug candidate identified and designed with AI assistance, not conclusive proof that AI discovers efficacious and approved medicines on its own. The institutional issue concerns the ownership of rights attached to these discoveries and the conditions under which they can be reproduced or used.
Current patent regimes reward the inventor and investor, which has economic logic. But they were designed for medicines where the bulk of cost lay in discovery, chemical screening, synthesis, and in vitro testing. Even if certain discovery costs decline, this does not mechanically determine the justification or legal duration of patents. A framework that distinguished ownership of the molecule from ownership of the algorithmic method, and that provided differentiated licenses according to the income level of the purchasing country, would be better adapted to this new economy.
This is nothing utopian. The Medicines Patent Pool, created in 2010 at UNITAID’s initiative, enabled a significant number of low- and middle-income countries to access generic HIV and hepatitis C medicines at prices compatible with their health systems. Participating companies retained their rights in solvent markets. Public research and shared intellectual property models already coexist in other sectors—telecommunications standards, Covid-19 vaccines under COVAX—with variable but non-negligible results.
AI in pharmacology is not different from other sectors where the same tension between productivity gains and concentration is observed. Robotics in Europe illustrates how technological advance and the question of infrastructure access arise simultaneously. What is specific about medicine is the human urgency that makes access delays politically untenable.
The first approvals of AI-discovered medicines will constitute a political as much as scientific test. Its price, its licensing conditions, and its availability in low-income health systems will tell whether the productivity gain produced by the machine is shared or captured.
Sources
- Towards Healthcare, North America AI in Drug Discovery Market : https://www.towardshealthcare.com/outlook/north-america-ai-in-drug-discovery-market
- Axis Intelligence, AI Drug Discovery Statistics 2026
- Grand View Research, AI Drug Discovery Market 2025–2033
- OECD, Intellectual Property and AI report 2026
- Medicines Patent Pool, license history and covered countries: https://medicinespatentpool.org
