Lila Sciences Wants to Run Science Like a Data Center

Inside a Cambridge lab, magnetically levitating plates shuttle between instruments the way a PCI bus moves packets between devices. The same lab runs vision-language models that drive legacy Windows-95-era lab equipment, with a single API line separating what a human touched from what a machine did. Lila Sciences, the Flagship Pioneering spinout unveiled in March 2025 and now valued at $1.3 billion with Nvidia's backing, is using this setup to run the most influential idea in modern machine learning against its hardest possible test: physical experiments where nature is the verifier and every data point costs real time and money.

The idea is the "bitter lesson," coined by reinforcement learning pioneer Rich Sutton in a 2019 essay. The short version: scale and general learning methods tend to beat hand-coded rules, again and again, across chess, Go, and language. Lila is applying that logic to wet-lab science. The lab, in CTO Andrew Beam and CSO Rafael Gómez-Bombarelli's framing from a recent Latent Space conversation, is an "infinite token generator," with reinforcement learning as a data-generation mechanism and nature as the verifier. The goal is what they call "scientific superintelligence," wired into biology, chemistry, and materials discovery at once.

The unveiling release named a $200 million committed seed round with Flagship as lead and General Catalyst, March Capital, ARK Venture Fund, Altitude Life Science Ventures, Blue Horizon Advisors, the State of Michigan Retirement System, Modi Ventures, and a wholly owned subsidiary of the Abu Dhabi Investment Authority on the cap table. The $1.3 billion valuation, reported by Reuters in October, came with new Nvidia backing, putting the company inside a small group of well-funded AI-for-science startups with the compute to actually run the experiments. Beam has said a biopharma-rename would put Lila in top-three GPU cluster territory, a rough proxy for how much silicon the lab is actually running.

The substrate, in practice, is a network of instruments treated as nodes on a graph, with orchestration that the team likens to a Slurm queue, the scheduler that splits compute jobs across a cluster. Plates carrying samples levitate on a magnetic track between stations, which removes the slow parts of human handoffs. The team has rebuilt individual measurements to run orders of magnitude faster. Gómez-Bombarelli told Latent Space his materials group rebuilt a gas-sorption measurement, the test that measures how much gas a porous material can hold, to run roughly 2,500 times faster than the standard protocol. The biology side, he said, went from idea to in vivo CAR-T data in non-human primates in about six months, a timeline the team explicitly framed against AbbVie's $2.1 billion Capstan acquisition, built on preclinical in vivo CAR-T data.

Lila's central data claim is what the team calls "over 10 trillion" experimentally validated scientific reasoning tokens, by which they mean units of scientific reasoning, not literal language tokens, all generated by real experiments. The internal benchmark is that models trained on this corpus beat domain-specific models trained on the same number of samples. If the claim holds, the lab's automated experiments are not just producing data. They are producing the substrate for a general scientific reasoner.

Two open bottlenecks sit alongside those numbers. Beam put reinforcement-learning training at "roughly 5 percent mean FLOP utilization," a concrete, attributed efficiency ceiling that explains why the data-center analogy is aspirational rather than descriptive. The other open bottleneck, in Gómez-Bombarelli's telling, is the sim-to-real gap, the persistent mismatch between physics-based simulation and the messy behavior of real physical systems. Both are exactly the problems the bitter lesson has historically failed to solve in domains with expensive, noisy, or low-N labels.

That is the productive tension in Lila's pitch. Sutton's lesson has held across games and language, where the verifier is fast, the data is essentially free, and the system can play itself millions of times before breakfast. Wet-lab science is the opposite. Each data point consumes reagents, instrument time, and human review. Replication is hard. Experimental noise, confounders, and small sample sizes are not footnotes; they are the data. Calling nature the verifier is correct in a thermodynamic sense and inadequate in a statistical one, because "nature's verdict" on a single experiment is often ambiguous.

Lila's response is to insist it is not an automation company. The pitch is not that robots will run the experiments humans already run, but that the loop of model, experiment, and feedback will produce a reasoner that can ask questions humans have not thought to ask. Whether that loop is "the bitter lesson, in a lab coat" or a much more expensive version of what existing high-throughput screening already does, is the open question. The $1.3 billion and the Nvidia logo make it a fair question to ask. The next twelve months of Lila's compute bill, benchmark releases, and replication disclosures will start to answer it.