Market assessment · Advise

AI for scientific discovery: why the bottleneck is below the model

Something real happened this year. The most decorated research team of the cycle left the largest AI lab to automate the experimental loop, the US government committed more than five billion dollars to AI for science under a Manhattan Project framing, and a generative-AI-designed molecule showed a clinical signal in a peer-reviewed trial. None of it is hype. All of it is upstream of the question that decides the outcome: whether these systems can reach trustworthy inputs, and whether anyone can check what comes out.

21 PAGES EVIDENCE-LABELED FREE TO READ
Commission an assessment

Four markets are being reported as one

Structure prediction, literature and data agents, autonomous labs, and the automation of research engineering get counted as a single "AI for science" market. They have different proof standards, different capital intensity, and time-to-falsifiable-proof ranging from months to a decade. Averaging them into one number is how the category gets oversold. Published figures here need heavier discounting than usual, because vendors are the primary source for almost every capability claim.

What the record actually shows

Six of the assessment's findings. The full report has more.

01

Four markets, one label

Structure prediction, literature agents, autonomous labs, and research-engineering automation have different proof standards and capital intensity, with time-to-falsifiable-proof running from months to a decade. Reporting them as one market is the first error.

02

The talent movement targets the loop, not the model

Jeff Dean left Google after 27 years, with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, to found Discovery Loop as a public benefit corporation on 5 August 2026, with Google a founding investor. The stated target is automating the experimental loop itself.

03

A real clinical gain at the easy end, none yet at the hard end

Roughly 173 AI-originated programmes were in clinical development in early 2026, with Phase I success reported at 80 to 90 percent, well above the industry norm. The gains concentrate where the biology is already understood, not at the frontier.

04

The most cited materials result did not survive review

A 2023 model proposed 2.2 million candidate structures and 380,000 predicted stable, and an autonomous lab reported synthesising 41 novel materials in 17 days. The field's own correction of that result is the more useful signal.

05

The binding constraint is the inputs

Of 1,792 manuscripts whose authors stated data were available on request, 6.7 percent produced usable data. The numbers on input quality are worse than the numbers on model capability, and inputs are what these systems consume.

06

The corpus is being priced, per buyer

Wiley reported 49 million dollars in AI licensing revenue for the year ended April 2026, and more than 110 million lifetime, on flat underlying revenue. Access to the scientific corpus is becoming a structural cost, not a passing controversy.

Read the full 21-page assessment for the government implementation plan that concedes the point, the scenarios, and the tripwires worth watching instead of the forecasts.

Every claim carries its evidence

This isn't a vendor summary. Every sentence is labeled by what stands behind it: verified fact, vendor claim, third-party estimate, my assessment, hypothesis, or scenario. Sources are numbered and clickable, and vendor capability figures are discounted rather than repeated. Forward-looking sections use scenarios with observable tripwires, not forecasts. It's the same method behind every market assessment I write.

The Bottleneck Is Below the Model

Twenty-one pages, built from public sources with no client brief and no interviews. Read it in the browser or take the PDF.

This is real, published work. Commission one for your decision.

Each report here answers a real question, directed and researched against public sources and evaluated against a stated assumption, then delivered as Word and PDF. If you're weighing a platform, sizing a category, or defending a number to a board, tell me the decision behind it and I'll tell you honestly whether a report is the right tool.

Commission an assessment

See more reports →