Molecules and Markets

In January 2026, an open-source project called OpenClaw went from a weekend build to the most-starred repository on GitHub in under five months. People were impressed with what it did. One developer let it manage his emails over several days and it negotiated $4,200 off a car purchase. Another had it inspect failed builds, diagnose the bug, fix the deployment, and submit a pull request, independently, on a schedule.

“Claude with hands” became its new name. However, this was misleading, as OpenClaw’s direct reach was purely digital, with the exception of a few cases where real-life jobs were listed by the agent and carried out by actual humans. Still, these were indirect interactions between the agent and the physical world. OpenClaw was limited to clicks, reading inboxes and files, filing GitHub issues, bookings, and sending messages. It was purely agentic, not physical.

The speed of its adoption is the focus of the story. The general public, initially fearful of agentic AI, quickly adopted it into their daily lives. Pretty quickly, people were comfortable handing an AI agent a loose goal and letting it act autonomously. Singapore currently has 51% adoption of agentic AI up from 22% the year before, with the recent introductions into the Armed Forces to improve workflow and efficiency.

Nvidia’s Jensen Huang described 2026 as the “ChatGPT moment for physical AI,” and the phrase has stuck across robots and manufacturing. Physical AI refers to systems that perceive their surroundings and take real physical actions: a robotic arm, manufacturing cell, or lab hardware. Google DeepMind’s Gemini Robotics-ER 1.6, released in April 2026, added the ability to read analog gauges and sight glasses, developed in partnership with Boston Dynamics. Furthermore, newsletters from Pymnts have noted Anthropic’s research on a standard, the Model Hardware Standard (MHS). Built with HHMI Janelia Research Campus, it enables agents to operate instruments such as microscopes, liquid handlers, and robotic arms, via the Model Context Protocol. 

This new venture into physical AI has immense opportunities in industries such as drug discovery, chemical engineering, and much more.

Focusing on chemistry, the integration of agentic AI has its own name: “robotic chemist.” First coined by a 2020 Nature paper, it describes a mobile robot that can run chemistry experiments on its own in a standard lab. What’s changed since then is the AI layer in the robot. A modern self-driving lab pairs a synthesis or liquid handling robot with an optimisation agent. The agent proposes an experiment, the robot runs it, and feeds back to the agent. The agent would then improve on the experiment with no human intervention. A 2026 Nature Synthesis paper describes an open-source version of this, RoboChem Flex, which costs about $5,000 to build from 3D-printed parts.

Escaping the academic demos and simulations, Chemify, a Glasgow-based spinout, runs automated “Chemifarms.” It takes a digital molecule design and physically synthesizes it, and already works with 6 of the top 20 pharmaceutical companies. Emerald Cloud Lab lets researchers anywhere in the world log into a web interface and remotely run experiments across roughly 200 real instruments; Carnegie Mellon built a dedicated $40 million campus around access to it. This is what “AI agents with physical reach” looks like when the reach extends into a chemistry lab instead of a browser tab.

What Could This Mean for Big Pharma and Chemical Engineering

Estimates for the lab automation market in 2026 range from about $6.6 billion to $10 billion, depending on the research firm, growing towards $8.6 billion to $24 billion by the early-to-mid 2030s. However, AI in drug discovery is specifically smaller and even more disputed across analysts, with 2026 figures ranging from $2.9 billion to $24.5 billion. That entire AI layer is still only about 1.5% of the roughly $194 billion the top 45 pharmaceutical companies spend on R&D each year. The money moving into partnerships is a better signal of where this is heading: Eli Lilly’s 2026 deal with Insilico Medicine is worth up to $2.75 billion, Takeda’s with Iambic is worth over $1.7 billion, and total AI/ML pharma R&D partnership value hit $45.9 billion in just the first half of 2026.

If self-driving labs continue to mature, the disruption is more focused on compressing what a lab can do with a given number of people. A robochemist system can run experiments overnight, unattended, in parallel, which changes the unit economics of drug and materials discovery rather than just the speed of a single experiment. This would have two resulting effects: smaller biotech and materials startups get access to lab infrastructure they could never have afforded to build themselves, simply by renting time on a Chemify or Emerald Cloud Lab, potentially lowering the capital barrier to entering pharma and materials R&D. Larger firms, in comparison, get to run more repetitive, high-volume synthesis and screening work without scaling headcount at the same rate. Not to mention, specialised AI chemical engineers and biotechnicians could take on a larger scope, possibly remotely, coming in only when mission-critical issues occur with the robochemist.

The U.S. Bureau of Labor Statistics data projects faster-than-average growth of chemical technician employment, 5% from 2025 to 2035, but for clinical lab technicians the agency flags that “the increasing automation of laboratory workflows may dampen employment demand.” The roles most exposed are the repetitive ones: manual pipetting, routine synthesis, high-throughput sample handling. What seems to be emerging alongside them are the hybrid roles that didn’t really exist a few years ago: Lab Automation Engineer, Digital Twin Lab Engineer, and AI Lab Safety Supervisor, a role specifically focused on overseeing what happens when AI systems are given control of physical chemical and lab processes. Robotics engineers more broadly are already commanding $121,000 to $137,000+ average base salaries in the U.S., with employers reporting they can’t hire enough.

Market-size forecasts for the same category can differ by a factor of eight depending on which analyst firm you ask, and no AI-discovered drug has yet reached the market. What does look durable is the underlying pattern OpenClaw made visible outside the lab: AI agents are being trusted with more autonomy and more consequential actions, faster than most institutions can build the guardrails for them. In a chat app, that means an agent negotiating your car price. In a lab, it means an agent deciding what molecule to synthesize next. The difference in stakes is enormous, and the industries built around chemistry and pharmaceuticals are only just starting to reckon with it.