For two years, the strongest argument against AI in regulated biotech was that the models were not good enough. Over the last two weeks, that argument quietly collapsed.

Pfizer licensed Chai Discovery's frontier antibody-design platform. The FDA approved Tempus's xT CDx for tumor-only use, a first for comprehensive genomic profiling. Biohub open-sourced a protein model trained on 6.8 billion sequences, with lab-validated designs working on the first try. Tempus reported a billion-parameter foundation model beating classical survival statistics with no fine-tuning.

These are not the same story, but they rhyme. Capability is no longer the constraint in this field, and capability was never the question a regulator was going to ask. That question is narrower. Not "can the model do it," but "can you prove what it did, and why." Last cycle the FDA enforced on exactly that gap. This cycle the labs widened it, shipping capability faster than the evidence layer underneath could be built.

The capability argument is over.

On June 4, Chai Discovery licensed its full antibody-design platform to Pfizer, including a previously undisclosed model called Chai-3. Chai says Chai-3 doubles the experimental hit rate of Chai-2, which was itself the first zero-shot antibody design platform to reach double-digit hit rates, roughly a hundredfold gain over earlier computational methods. Pfizer also receives a variant trained on its own proprietary data, and Forbes reported Chai is raising $400 million at a $3.4 billion valuation.

The word that matters is "zero-shot." The model designs binders it has never seen, and they work in the lab at rates that used to take months of wet-lab iteration. A top-five pharma is now running that capability against its own pipeline.

Here is what the press release does not address. When the model proposes a binder and a program advances on the strength of it, a chain of inferences sits behind the recommendation. Three years later, when an FDA reviewer or an internal auditor asks why this candidate and not another, "the model suggested it" is not an answer. The capability is real. The traceable reasoning behind each output is a separate problem, and it is the one that survives into the regulatory file.

Open models just raised the floor.

On May 27, Biohub, the nonprofit institute backed by Priscilla Chan and Mark Zuckerberg, released an open suite it calls a world model of protein biology: the ESMC language model trained on roughly 2.8 billion sequences, the ESMFold2 structure predictor, and an atlas of 6.8 billion proteins with 1.1 billion predicted structures. Lab-validated designs reached 36 to 88 percent success for compact mini-binders and 15 to 29 percent for antibody-derived formats across five disease targets.

This deserves a caveat, and the caveat is instructive. Last cycle, two preprints pressure-tested the "more parameters, more biology" thesis and found it wanting. A 2018-era variational autoencoder still beat every modern transformer on a single-cell benchmark, and AlphaFold-family samplers collapsed to a single conformational state where physics-based methods recovered the full ensemble.

Both things are true at once. Some foundation models are overhyped, and some now design working proteins in the open. The discipline does not change. Demand the benchmark on your actual task, and demand the record that lets you reproduce the result. Open versus proprietary changes the cost. It does not change whether you can defend the output.

A regulated AI product cleared the FDA, and the reason is the whole story.

On May 29, the FDA approved a tumor-only indication for Tempus's 648-gene xT CDx assay, making Tempus the first laboratory with companion-diagnostic clearance for both tumor-only and tumor-normal comprehensive genomic profiling. The tumor-only label removes the prior requirement for a matched normal sample, and the assay identifies colorectal cancer patients eligible for cetuximab and panitumumab. The stock rose about nine percent.

Strip away the genomics and this is a clean demonstration of a principle usually argued in the abstract. An AI-driven diagnostic expanded its regulated indication. That did not happen because the model got smarter. It happened because the validation and change-control evidence supported the expansion.

This is what regulated AI looks like when the documentation holds. The label moves when the evidence chain is complete enough to move it. Every team deploying clinical AI should be reverse-engineering this approval into a predetermined change-control plan of its own. The FDA has now shown it will expand indications for sponsors who can document the change. It will not for sponsors who cannot.

The models are starting to beat the statisticians.

The same week, at ASCO, Tempus disclosed initial results from a multimodal foundation model trained on 2.5 million longitudinal patient records, drawn from a database of more than 45 million de-identified patient journeys. Used with no fine-tuning, it predicted overall and progression-free survival on a validation set of more than 1.2 million records, with a proof of concept in EGFR-mutant lung cancer. Tempus reports it outperformed standard Cox proportional-hazards modeling, the workhorse of clinical trial analysis for the better part of fifty years.

If that holds, it changes how external control arms, feasibility analyses, and biomarker discovery get done. That is a large "if," and the announcement has earned a specific piece of skepticism. It claims outperformance without publishing a single headline accuracy number. Read the abstracts before repeating the claim.

But assume it replicates. When a sponsor uses a foundation-model survival estimate to justify a synthetic control arm, the regulator will ask how that estimate was produced, on what data, with which version of the model. A point estimate is not an answer. The reasoning path is.

AI moved into the part of the pipeline that actually fails.

On May 28, Quotient Sciences began what it describes as the first clinical study of an AI-designed drug formulation, a Phase I oral solid-dose trial in healthy volunteers cleared by the UK's MHRA. The claim is a vendor self-designation and should be treated as one until independently confirmed. The significance, if it holds, is the location.

Almost all the attention on AI in drug development sits on molecule design. Far less sits on formulation and chemistry, manufacturing, and controls, which is where a meaningful share of late-stage programs actually fail. Moving AI into that layer is a larger development than the modest framing suggests, and it carries the same evidentiary burden one step further down the pipeline. The MHRA cleared this one. The next regulator will want to see how the formulation was designed, not just that it was.

Europe is writing the rules. The FDA is between commissioners.

In June, the European Medicines Agency and the Heads of Medicines Agencies published the 2025 AI Observatory Report. Its framing is direct: 2025 marked the shift from AI exploration to real-world implementation. The report confirms that January's joint FDA and EMA principles on good machine-learning practice will be followed by a roadmap for further AI guidance, feeding the European regulatory strategy through 2028.

Set that against the American picture. Marty Makary, the public face of the FDA's AI-modernization push, resigned last cycle, and the agency now has three top positions filled by acting heads. The structural reforms are likely to survive, but the pace of new AI guidance will probably slow for a few months, exactly as the workload accelerates.

The takeaway is not which regulator is ahead. It is that the expectations are converging. Audit trails, version control, documented human oversight, reproducible reasoning. A sponsor that builds its evidence infrastructure to the stricter of the two standards is covered on both sides of the Atlantic. A sponsor waiting for a regulator to hand over a checklist will wait through the window in which this advantage is still cheap to build.

The throughline.

Speed without traceability is a liability in regulated industries. This cycle, the speed arrived and the traceability did not keep pace.

The bottleneck now is evidence, and specifically the distance between two things that sound alike and are not. Recording what an AI system did is traceability, and it is the easy win. Most platforms can produce a log. Reconstructing why the system reached a given conclusion, in a form a different reviewer can follow and verify months later, is reproducibility, and it is where most programs are thin.

Chai-3 inside Pfizer, a foundation model beating Cox regression, an AI-designed formulation in a human trial. Every one produces a decision someone will eventually have to explain. The capability that generated the decision is no longer the hard part. The defensible account of how it was reached is.

The companies that win the next eighteen months will not be the ones with the most agents or the largest model. They will be the ones whose AI decisions, every one of them, can be reconstructed when an inspector or a board member asks. That is what evidence infrastructure is for.

Prior-cycle updates.

The Isomorphic Labs round I flagged as a rumor last cycle closed at $2.1 billion, led by Thrive Capital with Alphabet participating. The FDA's early-phase AI trials RFI had its comment window extended, so sponsor engagement on that pilot is still live. And the asymmetry I predicted after the Makary resignation is beginning to show. Europe published guidance this cycle while three FDA seats remain filled by acting heads.