We now design more drugs than we can test

What does an answer cost?

The last three pieces I wrote were about drug discovery, mostly about what AI has and has not changed in it. Discovery got much better at producing candidates and the part that finds out which of them are worth anything did not. Filling the pipe with more of them wastes the billions already spent on generating them.

Better tools and worse odds

Between 2014 and 2024 the likelihood that a drug entering Phase 1 eventually gets approved fell from 10.4% to 6.7% on Citeline’s Biomedtracker numbers, with oncology at 4.7%.

Over a decade in which the industry got dramatically better at generating candidates, no corresponding improvement showed up in the rate at which candidates become drugs. None of that can be pinned on AI. Most of the technology is too recent to have moved a ten-year number, the easier targets went first, and crowded mechanisms leave only differentiated assets commercially viable. Whatever was gained upstream did not survive contact with the clinic.

The odds got worse and the clock got longer. Tufts CSDD compared drugs developed between 2008 and 2013 with those from 2014 to 2018 and found the mean clinical phase had grown by 6.7 months while the mean approval phase had shrunk by 1.9. Regulators got faster and testing got slower, by more than enough to cancel the gain. Tufts puts the direct cost of a Phase III day at $55,716, from an analysis of 447 protocols. Six extra months on one Phase III trial is about $10 million.

The half nobody funded

Most of the effort and investment went into molecular design, protein structure and genomics. That work is computational, the feedback loop is fast, the training data is largely public, and the results are publishable.

The clinical half has none of those properties. A trial takes years to read out. What went wrong inside it is rarely written down anywhere a stranger can reach it. The record that would answer the question sits in sponsor files and CRO systems, and the people who need it most are one-molecule companies with no budget line for it.

Nearly every new tool of the last decade went to discovery. Almost none went to the step that decides whether a candidate was worth chasing.

Trials that cannot answer their own question

Trials are expensive and long, and a striking number of them reach the readout without the power to settle anything.

Villacampa and colleagues, writing in ESMO Open in June 2024, reviewed 282 adjuvant phase III oncology trials published between 2013 and 2023. 22% finished recruiting without reaching their target sample size. Nearly 40% were underpowered at readout, carrying less statistical power than their own protocol specified. Median accrual period was 4.3 years. Trials that lost power produced statistically significant results 21.9% of the time against 37.9% for those that kept it.

Underpowered is not the same as under-recruited. A trial loses power when it misses enrolment, and also when the event rate comes in below what the protocol assumed, or the effect size was optimistic, or the endpoint behaved differently than expected. Most of those are assumptions made before recruitment started.

And patients are the hard input. Thanks to AI you can generate more candidate molecules than ever before. But you will still have a fixed number of people with a given cancer. Tran and colleagues, in JCO Clinical Cancer Informatics, took 4,598 actively recruiting interventional oncology trials with US sites and estimated 12.6 newly diagnosed patients available per trial slot. For all of those trials to accrue successfully, 7.9% of recently diagnosed cancer patients would have had to enrol. Unger and colleagues put actual enrolment in cancer treatment trials at 7.1%. Demand already runs slightly ahead of supply, and Villacampa shows you where the shortfall lands. Every trial that opens competes for patients an existing trial already needs.

The least experienced buyers, and a vendor with no reason to help

In 2024, emerging biopharma sponsored 63% of all clinical trial starts, 5,318 of them, up from 56% in 2019 according to IQVIA. BIO’s June 2025 report has emerging biotechs behind 36 of the 54 novel drugs the FDA approved the previous year. So two thirds of new medicines come from outside large pharma, which now buys or licenses the asset more often than it discovers one. The organisation with the deepest capability to generate a clinical answer would rather buy one, and pays up for assets that already survived a readout.

A large and growing share of clinical development choices are thus made by organisations making them for the first or second time, with no institutional memory of how the last twenty trials in that indication actually went. A large pharma company has run enough studies to have opinions about which sites deliver and which protocol assumptions tend to be optimistic. A company with one molecule and a Series B has none of that, and there is nowhere to buy it.

So it buys execution instead, from a contract research organisation, which has no reason to make the trial cheaper or shorter. A CRO is paid to run the study in the contract, to protocol, in compliance, at good utilisation. Nothing pays it to finish early, and nothing pays it to tell you the study is the wrong one. The sponsor wants the most reliable answer per dollar and per month spent, which sometimes means running a different trial from the one it signed for, or not running it at all.

A founder who has run several trials recently told me you cannot tell a good CRO from a bad one before you sign, because almost nothing about their real performance is visible from outside. And once the trial is running and the monitor assigned to you turns out not to be up to it, replacing that person is painful enough that most sponsors do not. You choose blind, then you live with the choice.

Much of what determines a Phase 2 outcome is fixed before the first patient is dosed, in the choice of indication, population, endpoint, dose, comparator and recruitment assumptions. Those are the choices being made for the first time.

Three questions a sponsor cannot get answered

A sponsor with a molecule and $40 million should be able to ask three questions and get real answers.

Of my candidate indications, which one yields a readout I can raise on inside the capital and the time I have? Answering that needs historical event rates, achievable effect sizes and realistic accrual, and none of it sits anywhere I can reach.

Which sites have actually recruited this population in the last five years, and at what rate? My CRO will propose the sites it already holds contracts with.

Which choices in my draft protocol have historically produced underpowered or slow studies? Flag the assumption that is about to cost me two years.

None of that requires new science. It requires knowing what happened in trials that already ran.

Everyone ends up owning the drug

Intelligence tells the sponsor what to do. Infrastructure helps them plan and manage development. Execution runs the trial. Execution holds the large revenue, and it is a services business, which is a different bet with different economics.

Can anyone build the first two without ending up in the third? Recent attempts say something stranger. Both passed through the third and out the other side.

Formation Bio was TrialSpark, which spent its early years building tools for patient recruitment and site management. It rebranded in December 2023 and now acquires clinical-stage drugs from other companies to develop itself, with four active programmes as of December 2025, all in Phase 1 or 2. It claims its trials are 30 to 50% faster and recruitment 2.5 times faster than benchmark.

Lindus Health had raised over $80 million by its January 2025 Series B, called itself the “anti-CRO” CRO. In August 2026 it sold those CRO assets to Curavit, renamed itself Lindus Therapeutics, and started hunting drug assets to in-license.

Two companies with different starting points ended up at the same destination. Both got better at running trials and they are now using it on molecules they own. That should worry anyone thinking of building here. If you get good enough at running trials to have a real edge, you can point it at your own assets and keep all of it, or sell it to sponsors at a services margin. The first pays better.

But both were building with the last generation of tools. Formation Bio built its trial software before the 2023 rebrand, when none of the current models existed, and both sold a service where people did the work. Margins are thin in that business and headcount grows with revenue, and neither company found a way out.

That has changed. The AI-first services companies now coming through our deal flow take work that used to need a person and do it without one, and the economics that pushed Lindus and Formation Bio out of services may not apply to them. Lindus was still building through 2026 and left anyway, so nothing here is settled. But someone may finally break the CRO model instead of joining it.

What is worth building

I argued in an earlier piece that in drug discovery the model was never the moat. The same holds here. Anyone can buy the models. Nobody can buy a record of how trials actually ran, gathered across many sponsors and kept over years: trial design linked to sites, recruitment velocity, timelines, costs, protocol choices and outcomes.

Parts of this exist already. Large pharma has a deep view of its own trials, CROs have a wide view of the work they were hired to run, and registries hold the design and the outcome. No comparable cross-sponsor version exists to serve as an independent benchmark, because everyone holding a part has every reason not to pool it.

It is a slow asset and it takes years of trials before it means anything. That is why nobody has built it, and why the new tools matter. A company that can run trials without a services headcount can afford to keep serving sponsors while the record accumulates, instead of taking the exit Lindus and Formation took.

Citeline counted close to 6,000 drugs in development in 2001 and around 23,000 now. The industry holds close to four times as many candidates as it did twenty years ago and converts them at a worse rate than it did a decade ago. Large pharma has already priced this. It buys assets that cleared a readout and pays up for them, which is what you do when the scarce thing is a credible answer rather than a molecule.

We will keep investing in discovery and in therapeutics. That is where most of the good companies are and that is not about to change. But we now ask what one answer costs, in money and in months, before anyone knows whether a candidate is worth another round.

If you are building something that brings that number down, I would like to talk.