Lila Sciences: Science as Data for AI — an Autonomous Lab Backed by $550M

On July 16, 2026, Lila Sciences explained on the Latent Space podcast how its autonomous lab works: AI controls robots and equipment, designs and runs experiments on its own, and has already generated more than 10 trillion experimentally verified tokens for training scientific models.

Why science instead of the internet?

Lila Sciences is betting that the scientific method is the last untapped internet-scale dataset. The company's logic is simple: text scraped from the web for training large models is nearly exhausted, while verified experiments are a practically endless source of new data.

Training here is built as reinforcement learning in which nature itself acts as the verifier: the model proposes a hypothesis, a robot runs the experiment, and the result from the physical world confirms or refutes the prediction. As a result, Lila has accumulated more than 10 trillion tokens of scientific reasoning tied to real measurements rather than to internet text.

How the robot lab works

Lila's lab is AI-driven robotics layered on top of ordinary scientific equipment, not an assembly line built for speed. The company stresses that it optimizes for flexibility and generalizability rather than throughput: a single setup should be able to run a wide variety of experiments.

• Samples travel on pucks that float over magnetically levitating tracks
• Vision-language models operate legacy instruments by "looking" at them the way a human would
• Experiment orchestration works like a SLURM queue, modeled on a compute cluster
• A gas sorption measurement was reworked to run roughly 2500 times faster
• A model proposed electrocatalysts free of platinum-group metals that outperformed earlier results
• In six months, the team reached in vivo data on CAR-T therapy in primates

"The lab of the future should feel like a data center," is how

Andy Beam, CTO, and Rafa Gómez-Bombarelli, Lila Sciences' head of physical sciences, describe their approach in the Latent Space episode.

How much has the company raised?

Lila Sciences has raised about $550M in total and reached a valuation of roughly $1.3B. The company grew out of the venture incubator Flagship Pioneering, the same one behind Moderna.

The first round, $200M, closed in March 2025, led by Flagship Pioneering. In October 2025, as Reuters reported, Lila added another $115M as part of a Series A extension that included Nvidia's venture arm, bringing the Series A to $350M and the valuation to ~$1.3B. The company is led by CEO Geoffrey von Maltzahn; the chief scientist is geneticist George Church.

What this means

If Lila's bet pays off, the AI race will shift from scraping the internet to producing proprietary scientific data: the winner will be whoever has not more servers full of text but more robots running real experiments. So far this is supported only by individual results, such as faster measurements and new catalysts, and not by a universal "scientific superintelligence."

Frequently asked questions

What is Lila Sciences?

Lila Sciences is a startup from the Flagship Pioneering incubator that is building autonomous AI labs for "scientific superintelligence": AI designs experiments and runs them through robots. The company was founded by Geoffrey von Maltzahn together with a team that includes Andy Beam and Rafa Gómez-Bombarelli.

How much has Lila Sciences raised?

About $550M in total: a $200M seed round (March 2025) and a $115M Series A extension with Nvidia's participation (October 2025). The company's valuation is roughly $1.3B.

Why does Lila need robots and a lab?

The robots exist to generate verified experimental data for training models: the company considers experiments, not internet text, the main inexhaustible source of data for AI. More than 10 trillion verified scientific tokens have already been collected this way.