Getting new helpful medicines to patients is hard, and historically it’s been getting harder. From 1950 to 2010, the number of new approved drugs per billion dollars of R&D spending halved roughly every nine years. In 2012, Jack Scannell and colleagues called this Eroom’s law, or “Moore’s law backwards”.1 One theory that explains this is the “better than the Beatles” problem: we’ve already made substantial progress on various diseases, and statistically speaking, it requires more and more patients (and therefore more and more funding) to demonstrate marginal improvements — just like how the next generation of pop stars needs to compete with the Beatles catalogue.
Despite that progress, the global burden of disease is still enormous. Most positive visions of the future put forward by AI leaders today involve the use of AI to accelerate the eradication of this burden.2 But even if AI produces a miracle drug for every disease, we still need to run the clinical trials to prove that things are getting better.
Saloni Dattani recently made a fun game to help illustrate the challenges in running these sorts of clinical trials.3 My main takeaway is that making trials go faster requires a lot of different innovations to come together; as per Amdahl’s law,4 there’s no quick fix.
This page is about the cost of measuring that progress:
Suppose AI found, for every deadly disease in the world, a therapy that cuts the chance of its worst outcomes in half compared with the best treatment available today. What would it cost to prove that in pivotal trials?
What is a benevolent coordinator? An idealised planner who runs the world’s pivotal trials as a single programme to best address the global disease burden. So no duplicate or competing trials.
This is only the cost of proving that it all works. Discovery, earlier-phase trials, manufacturing, delivery and access all cost more, and are much harder to predict. Conditions that cause illness but rarely kill, such as depression or back pain, are not on the list, so “every disease” would cost more.
List the diseases, price one pivotal trial for each, then add them up:
From the Global Burden of Disease (GBD) study’s list of causes of death from disease, excluding injuries and catch-all groups like “other cardiovascular diseases”, which bundle many conditions no one trial could cover.5 Use one pivotal trial per disease.
Find an RCT that tested what international guidelines recommend today. Among the patients who received it, take their risk of a bad outcome over that trial’s duration, p0.
Work out how many participants a new trial would need to reliably show the risk halving, from p0 to p0/2. Standard hypothesis-testing maths gives the total, N (90% power, 5% significance).
N × $53,904 per participant, summed over diseases. That is Moore et al.’s median cost per participant of a pivotal trial, adjusted to 2026 dollars.6,7 See the maths ↓
Each dot is one way a trial could measure success: a bad event, such as death or a heart attack, that the new treatment would need to make half as common. Rarer events need bigger, more expensive trials. The black tick marks the median cost for each disease, which is what the total adds up. Hover overTap a dot to see the trial its numbers come from.
The headline number surprised me in how small it was. In 2024 the global pharmaceutical industry spent something like on phase 3 trials (in 2026 dollars). I expected the bill for every major disease to be far bigger, even with a benevolent coordinator in charge.
Part of the cost gap is the size of the assumed improvements. Halving the risk of a disease’s worst outcomes is a lot of progress, which needs fewer participants to produce significant results than standard clinical trials. In the pivotal trials establishing today’s best care, the risk of the bad outcomes was cut by a median of . Proving improvements of that size, rather than halvings, would cost about in total.
The other part is redundancy. Many companies compete to treat the same patients: by 2022 there were 5,683 trials testing drugs that block a single immune pathway, PD-1/PD-L1. This got so crowded that combination trials started to outstrip patient demand and the FDA had to call on drugmakers to collaborate to minimise redundancy.8 A newer immune target, TIGIT, has drawn 21 companies developing 30 drugs: about 49,000 patients in 220 trials, at an estimated cost of $3.1–3.6 billion.9 Across the drug R&D pipeline, the number of cancer drugs in development per target rose from 1.8 in 2000 to 9 in 2022.10 The effort is lopsided too in terms of global welfare: cancer accounts for 38% of industry trial starts,11 but 18% of the world’s deaths.5
What would a better-coordinated system look like? During the pandemic, the UK’s RECOVERY trial combined research efforts across NHS hospitals. It showed that dexamethasone reduced deaths among patients hospitalised with COVID-19 without requiring different groups to reproduce trial arms separately.12 I-SPY 2 ran a similar platform trial for breast cancer for more than a decade.13
Other industries have also successfully achieved better coordination to avoid redundancy across firms: in 1987, 14 US chipmakers formed SEMATECH to share pre-competitive manufacturing research. I’m not familiar with the field, but it seems to be widely credited with helping the US recover against Japanese competition.14 Maybe a better system for medicine would look similar: shared trial platforms that companies compete to put their drugs into, rather than dozens of near-identical trials competing for the same patients.
Let p0 be the chance that the bad outcome happens within a fixed follow-up window on today's best care. A therapy that halves it brings the chance down to p1 = p0/2. The trial splits people equally between the new therapy and best care. It must be large enough that a real halving is detected 90% of the time (power), with no more than a 5% chance of a false alarm (two-sided significance). The standard formula for comparing two proportions gives the number of people n in each arm:
Here z1−α/2 = 1.96 and z1−β = 1.28 are the standard normal values for 5% two-sided significance and 90% power, and p̄ = (p0 + p1)/2. The total across both arms is N = 2n. The trial's cost is then N × $53,904: Moore et al.’s median estimated cost per participant across 225 pivotal trials ($41,413), which I assume is in 2019 dollars and adjust by US consumer prices to 2026.6,7 This is a fixed-time approximation, not the event-driven survival design many real trials use.
The gap a trial has to detect, p0 − p1 = p0/2, is squared in the denominator. So the rarer an outcome already is on good care, the more people it takes to see it halve.
| Risk today | Halved | People needed |
|---|
PhRMA’s member companies spent 27–39% of their R&D on phase 3 trials in 2024.15 Applying that share to the $288 billion global total16 and adjusting to 2026 dollars7 gives about .
Written and designed by Conor Igoe, using Kosmos.