GSK Joins Chai Discovery's AI Protein-Design Platform -- the Startup's Sixth Big-Pharma Deal Since January

GSK will use Chai Discovery's protein-folding and design models across its drug-discovery pipeline, the two companies announced this week, after a wet-lab evaluation that GSK itself calls a genuine test: it picked a set of its own research targets, asked Chai's models to generate candidate binders with no target-specific training, and then tested the designs in its own labs rather than taking Chai's word for the results. GSK says the models found a working binder for 100% of the targets it tried. Christopher Austin, GSK's SVP of R&D Technologies, said the designs showed strong binding affinities and that the platform "complements our own internal models that we continuously retrain against our own wet-lab data."

GSK is not a bet-the-company gamble for Chai; it's the sixth name on a list that was already long. Chai signed Eli Lilly in January, Pfizer in June, Novartis and argenx within days of each other in July, and Bristol Myers Squibb in August -- five major pharma platform agreements in roughly seven months before GSK made six in ten. Terms have not been disclosed in any of them: no upfront payments, milestones, royalties, or equity stakes have been made public for the GSK deal, matching the pattern of the five before it.

The valuation climbed with the client list. Chai raised a $30 million seed, then $70 million at a $550 million valuation in August 2025, then $130 million that December, before the July Series C put it at $3.8 billion -- a roughly sevenfold rise in under a year, on the strength of exactly the kind of lab-confirmed design result GSK is now citing. None of those pharma deals has yet produced a drug candidate in human trials; what they've produced, repeatedly, is a company's own lab confirming that Chai's models can find a molecule that binds a target it's pointed at. Chai's own figure for how often its earlier Chai-2 model pulls that off from scratch runs 16% to 20%, by the company's telling.

That gap between "found a binder" and "found a drug" is where the serious reservation sits, and it's one the companies' own disclosures leave open rather than answer. Analysts following the deal have been explicit that the initial results establish laboratory-confirmed binding across the tested targets, but do not establish that the generated molecules demonstrate functional activity, selectivity, suitable pharmacokinetic properties, or safety in biological systems -- the steps that make up most of the time, cost, and failure rate in actual drug development. GSK's own description of the deal is careful on exactly this point: the company frames Chai's platform as a complement to its internal models, not a replacement for the wet-lab work that follows.

What Chai actually sells is a shortcut around one of biologics research's slowest steps. Its current model, Chai-3, is a multimodal, diffusion-based system that designs a complete antibody sequence and three-dimensional structure from scratch, rather than screening a library of existing candidates against a target and hoping for a hit. The traditional version of GSK's test -- hand off a target, get back a short list of candidate binders -- can take a specialized lab months of iterative screening; Chai's pitch is that a model can propose structurally novel candidates on the same target in a fraction of that time, with no prior exposure to it. That speed claim is exactly what the GSK result tests, and exactly what it can't yet prove, since finding *a* binder fast is a narrower achievement than finding the *right* one.

Even the deal count is a little contested, in a way that's worth naming rather than smoothing over: one tracker counts the Bristol Myers Squibb agreement as Chai's fourth big-pharma license, after Lilly, Novartis and argenx, rather than fifth -- apparently a difference in which prior agreements count as full platform licenses versus narrower collaborations. The exact ordinal matters less than what both counts agree on: six major drugmakers, by any reasonable count, have now paid for access to the same unproven pitch inside of ten months. That's a rational bet for each of them individually -- the up-front cost of trying is low, none of the license terms disclosed so far include exclusivity, and the downside of a dry well is just one unused software seat. It is not, on its own, evidence that the pitch is correct.