The Self-Driving Lab Is Becoming Infrastructure

In January 2026, NVIDIA and Eli Lilly said they would put up to $1 billion into a single laboratory. Not a building full of chemists. A closed loop where robotic wet labs run experiments around the clock and feed the results straight back into AI models that design the next test. That is the self-driving lab (SDL), and 2026 is the year it stopped being a conference demo.

What actually runs inside the loop

A self-driving lab is not an automated pipettor. It is a closed control system. The AI planner proposes an experiment, robotic liquid handlers and synthesis stations execute it, analytical instruments read the result, and a model updates its beliefs within minutes rather than weeks. The loop then queues the next test on its own. NVIDIA describes the Lilly build as a "continuous learning system" that links agentic wet labs with computational dry labs in a 24/7 cycle, built on its BioNeMo platform, a collection of pre-trained biological and chemical foundation models.

The phrase "agentic" matters. Earlier lab robots followed fixed protocols. The new generation proposes the protocol, writes the robot code, and re-plans when a reaction misses its target. Andrew White, a co-founder of the chemistry agent ChemCrow, put the shift plainly: large language models became flexible enough to do the analysis, design the protocol, and emit the machine instructions. That is what turns a bench into something closer to an autopilot.

Why capital is showing up now

The pattern is the same one that already reshaped compute. A few years ago, training frontier models looked like a research problem. It is now a capital-intensive infrastructure business with fabs, power purchase agreements, and multi-year buildouts. The self-driving lab is following the same curve, except the scarce asset is physical experimentation and the data it produces.

Three things changed at once. Robotics got cheap and reliable enough to run repetitive synthesis and characterization. Foundation models got good enough to plan experiments and write instrument code. And the economics of discovery got brutal enough that large players needed throughput they cannot get from people. A research group running a tight loop produces more high-quality data in a week than a manual lab does in a quarter, and that data compounds into better priors for the next campaign.

The field is already crowding in. Atinary opened a self-driving chemistry lab in Boston in February 2026 and drew an early visit from an AstraZeneca team. Radical AI runs a similar facility in New York after raising $65 million. Dunia Innovations raised $11.5 million for a materials-discovery lab. Novo Nordisk signed a partnership with OpenAI in April 2026 to push intelligence across its R&D, manufacturing, and commercial operations. None of these is a rounding error on a pilot budget.

A self-driving lab is like a Waymo. You can get in and tell it where you want to go. You get in, close your eyes, and you end up at that destination. That's true self-driving.— Joseph F. Krause, co-founder and CEO, Radical AI

The hype and the reality

The honest version is that no lab yet runs without a human in the loop. Krause describes his own AI as "still getting its PhD." Integration is the hard part: most legacy instruments were never built to talk to an orchestration layer, and data sits fragmented across lab information systems, spreadsheets, and notebooks. The labs that work today are narrow, and they win in domains with a clean, repeatable physical experiment, not in open-ended biology.

The investment thesis underneath

For a private investor, the interesting question is not whether SDLs work. They demonstrably accelerate narrow discovery. The question is who captures the value. There are three layers. The compute and model layer is dominated by the large AI platforms. The physical automation layer is fragmented across robotics and instrument makers. The data layer, the experimental results a loop generates, is where the defensible moat sits, and it is why a pharma company rather than a tool vendor is co-funding a $1 billion lab.

Cloud labs add a fourth layer. Emerald Cloud Lab and Strateos let a team run rigorous chemistry or biology by submitting code, paying per experiment instead of building a facility. That lowers the activation energy for a startup and creates a recurring-revenue rail that looks more like infrastructure than a one-off sale. When the hardware, the models, and the execution all sit behind an API, the lab becomes a line item, not a capital project.

Who owns the lab in 2030?

Development scenarios

Sources