10 trillion experimentally validated scientific reasoning tokens now exist in Lila Sciences' automated laboratory, representing what the company calls the last untapped source of training data for AI systems.

The Boston-based startup has built what resembles a data center filled with robotic lab equipment. Floating plates zip around on tracks while vision-language models control legacy Windows 95 systems, all generating scientific data around the clock.

Andy Beam, Lila's CTO, and Rafa Gómez-Bombarelli, Chief Science Officer for physical sciences, described their approach as treating the lab like an infinite token generator. The company operates across biology, chemistry, drug discovery, and materials science simultaneously in the same facility.

Why Science Beats Internet Data

Lila's thesis centers on the "bitter lesson" — that scaling data and compute consistently outperforms domain-specific approaches. The company argues that scientific reasoning tokens, verified through physical experiments, represent a fundamentally different data source than internet text.

"The coding model got better because it also read Shakespeare and carnitas recipes," Beam explained, defending their broad approach against specialized competitors.

The startup has rebuilt traditional laboratory processes for speed. Gómez-Bombarelli's team accelerated gas sorption measurements by roughly 2,500x compared to standard methods.

From Boring to Breakthrough

The company's AI has generated suggestions for platinum-group-free electrocatalysts that initially seemed "stupid" to a 40-paper expert but became their best-performing materials. The system has also produced CAR-T therapy data in non-human primates within six months.

Lila operates what it calls a "zero-FTE virtual startup" model for commercializing discoveries. For context, AbbVie paid $2.1 billion for Capstan Therapeutics based on similar preclinical CAR-T data.

The company faces unique challenges when AI reasoning meets physical reality. Models sometimes skip experiments entirely while still reaching correct conclusions, raising questions about trusting reasoning versus experimental verification.

Lila emerged from Flagship Pioneering, unusual for the biotech incubator known for single-asset companies. Beam noted that if Lila called itself a biopharma company, it would rank among the top three for GPU cluster size.