For three years, the story in AI drug discovery had one main character: the model. Whoever trained the best foundation model for proteins, or chemistry, or biology would win. Capability was the bottleneck, and capability lived in the model.

In the last two weeks, that story quietly broke.

On June 16, Merck signed a multi-target collaboration with Protillion Biosciences worth an undisclosed upfront payment plus up to $510 million in milestones. Merck did not pay for a model. It paid for a data engine: Protillion's Prot-MaP platform, which prints protein variants onto a sequencing flow cell and measures up to one million of them in a single run, returning results in as little as 48 hours. Two weeks earlier, on June 3, Alnylam committed up to $2 billion to a three-year deal with Inceptive Nucleics, the startup run by Jakob Uszkoreit, a co-inventor of the Transformer architecture that underlies ChatGPT. The headline was the AI pedigree. The substance was the data: Inceptive's wet-lab data generation paired with Alnylam's two decades of proprietary siRNA measurements. Merck had already signed the same logic in March, paying up to $2.2 billion for Quotient Therapeutics' somatic-genomics data.

The pattern is hard to miss. The money is no longer chasing the model. It is chasing the data.

Capability stopped being the scarce thing.

This shift is rational. Frontier biology models have commoditized faster than almost anyone predicted. Open ecosystems like ESM and structure predictors of AlphaFold's lineage put state-of-the-art capability within reach of any competent team. Merck itself already has models, plus a $1 billion agentic backbone it announced with Google Cloud in April. When a top-five pharma writes a nine-figure check, it is not buying something it could download. It is buying something it cannot replicate: proprietary, measured biological data, generated on demand at a scale no public dataset can match.

So the new moats are data moats. Whoever can manufacture high-quality experimental data fastest has the durable advantage. On that much, the market and I agree.

Here is where I part company with the celebration.

Generating data faster is not the same as proving it is right.

A data engine that produces a million measurements in 48 hours is an extraordinary thing. But notice what it actually optimizes. It makes the loop faster and more repeatable. It produces a clean record of what was measured. In the vocabulary I keep returning to, it improves traceability, the record of what happened, and reproducibility, the property that the same process run again yields the same result.

It does not, by itself, touch the third property, the one that matters most in a regulated setting: verifiability. Verifiability is independent confirmation that a result is correct, established without relying on the system's own account of itself. Traceability tells you what the model did. Reproducibility tells you it will do the same thing tomorrow. Only verifiability tells you the answer is right.

Those are not the same question, and the difference is not academic. A system can be reproducible and wrong. It can generate, learn from, and act on data in a perfectly repeatable loop, and arrive reliably at the incorrect conclusion. Reliability is not safety. A model that is confidently and consistently wrong is more dangerous than one that is obviously erratic, because the consistency reads as trustworthiness.

A system can be reliably wrong, and the evidence is piling up.

This is not a thought experiment. A cross-modality review published this month catalogued how medical AI fails across CT, MRI, PET, ultrasound, and pathology: fabricated anatomy, missed findings, wrong laterality, invented measurements that look entirely plausible. Earlier benchmarks put medical hallucination rates as high as 64% without mitigation prompts, falling only into the low 40s with them. These are not glitches at the margins. They are the normal behavior of fluent systems operating beyond their competence.

And note what does not fix this. Explainability, the feature every newly cleared diagnostic now advertises, shows you where a model looked and why it flagged what it flagged. That is genuinely useful. But it is still traceability. Showing where the model attended is not the same as confirming that what it concluded is correct. The bounding box tells you the system's story about itself. It does not independently verify that the story is true.

This is the core error running through current clinical AI governance: it treats traceability as if it were verification. It asks systems to show their work, then mistakes a legible record for a correct answer.

The regulator is about to make this concrete.

The abstract argument is becoming a deadline. The FDA's draft guidance on using AI to support regulatory decisions, first issued in January 2025, built its entire approach around a risk-based credibility framework: before an AI output can support a decision about safety, effectiveness, or quality, the sponsor must demonstrate that the model is credible for that specific context of use. The agency signaled that the final version would arrive in the second quarter of 2026. That quarter ends on June 30. The joint principles the FDA and EMA published together in January 2026 point the same way, foregrounding data governance, documentation, and traceability across the entire product lifecycle.

Read those documents closely and the throughline is unmistakable. Regulators are not asking how fast your AI is. They are asking whether you can prove its output is fit to stand behind. The FDA does not care that Prot-MaP can characterize a million variants in 48 hours. It cares whether the path from a raw measurement, to a model's decision, to a claim in your submission can be reconstructed and independently confirmed.

The chain of evidence is the infrastructure nobody bought.

Here is the gap the data-engine race leaves open. A data engine generates data. It does not, by default, generate evidence. Those are different artifacts. Data is a measurement. Evidence is a measurement plus an unbroken, verifiable account of where it came from, how it was used, and why the conclusion drawn from it can be trusted by someone who was not in the room.

Think of the difference the way aviation does. A flight recorder gives you traceability: after the fact, you can reconstruct exactly what happened. An airworthiness certificate gives you verifiability: before the aircraft flies, an independent authority confirms it is safe to fly. The industry is currently racing to build better flight recorders. It is barely building airworthiness certification at all. Both matter, but only one gets you cleared for takeoff.

The deeper bind is that the two needs collide. You need evidence before you deploy an AI system in a regulated workflow, but you often need to deploy it to generate the evidence in the first place. The only way out of that loop is to stop treating evidence as something assembled at the end, and start generating it continuously, as a built-in property of the pipeline. Every measurement becomes a link. Every model decision carries its provenance forward. The result is something closer to double-entry bookkeeping for AI, a ledger in which each entry is cross-checked against the record that produced it, so the whole chain can be audited by an outsider years later.

That is the layer the new data moats are missing. The companies pouring billions into faster data generation have solved for throughput and reproducibility. They have not yet solved for the chain of evidence that turns a fast, repeatable data factory into a system whose outputs a regulator will accept.

The moat moved to data. The gap moved to proof. The next advantage will not belong to whoever generates the most data, but to whoever can prove that every datapoint, and every decision built on it, is correct and accountable from raw signal to final claim.

That is what evidence infrastructure is for.