Protein design works. Scaling it is the hard part.

Chai Discovery is building AI models that turn biology into an engineering discipline. Their structure prediction and protein design models, now reaching near-atomic accuracy (0.33 angstrom error), enable teams to design antibodies precisely rather than relying on slow, expensive trial-and-error like immunization or yeast display.

The real shift is organizational. Traditional drug discovery follows a waterfall model — target discovery, hit identification, lead optimization — each gate taking months to years. With Chai's models, teams can compress this into an agile loop, testing more hypotheses faster. But the bottleneck is no longer model capability alone.

The tension is between model generality and bespoke product needs. Chai aims for general models but must build custom workflows for each pharma partner — selectivity, cross-reactivity, bispecifics. This creates a consulting-like dynamic where the product must adapt to each partner's scientific and security constraints.

Most teams underestimate the engineering and infrastructure challenges. From compute scarcity (95% of B300s go to hyperscalers) to file format parsing nightmares, making biology work like software requires deep engineering rigor, not just better models.