Part Two · The Stack · Layer 5 of 9

The Hands — Automation, Robotics, and the Orchestration Layer

An AI model that designs a molecule has done nothing until a physical experiment tests it. The bottleneck in modern drug discovery is not ideas; it is the rate at which ideas can be run through wet-lab reality and turned back into data. Every design-make-test-analyze loop has to close in the physical world, and the physical world is slow, error-prone, and staffed by expensive humans doing repetitive work. The "hands" of the industry, the robots that move liquid, the arms that carry plates, and the software that sequences them, are what set the clock speed of discovery. This is where AI's appetite for data meets the hard constraint of atoms.

Understanding this layer means separating three things that are usually lumped together: the instruments that do the work, the physical robotics that connect them, and the software that decides what happens when. The investable insight, as we will see, sits almost entirely in the third.

Liquid handling: the most repeated action in the lab

The single most common physical action in a life-science lab is pipetting, moving a precise, small volume of liquid from one vessel to another. A screening campaign can require millions of these transfers. Doing it by hand is slow, introduces variance (a tired technician at hour six is not the technician at hour one), and caps throughput at human speed. Automating it is the foundational act of lab automation.

A liquid handler is a robot that aspirates and dispenses liquid across a fixed work surface, or deck, laid out with plates, tips, and reagent reservoirs. Two engineering variables dominate. The first is precision: many assays operate at microliter and sub-microliter volumes, where a small percentage error in dispensed volume corrupts the result and, downstream, the training data an AI model will learn from. Dirty data at the pipette propagates into every model built on it. The second is throughput: how many wells the instrument can process per hour, which is a function of channel count (an 8-, 96-, or 384-channel head moves far more liquid per cycle than a single tip) and deck capacity.

The market splits into two philosophies. Deck-based workstations, large, enclosed, high-precision systems, are the workhorses of pharma and contract research. Against them sits a low-cost, open-system approach that trades some precision and speed for affordability, scriptability, and access, opening automation to academic labs and startups that could never justify a six-figure workstation.

The public-market reality is stark. The clean pure-play is Tecan (TECN.SW), the Swiss maker of premium deck-based liquid handlers and detection systems, effectively the only way to own this instrument category directly on a listed exchange. Almost everything else is private or buried. Hamilton (private) is the other premium name serious labs standardize on. Opentrons (private) is the low-cost, open-source disruptor. Eppendorf (private) and SPT Labtech (private, owned by EQT) round out the specialist field. The rest is inside conglomerates: Beckman Coulter's liquid handlers sit within Danaher (DHR), and the Bravo platform sits within Agilent (A). Owning the pipette, in other words, mostly means owning a diversified life-science tools giant and accepting the dilution.

The workcell: turning instruments into a line

A liquid handler alone automates one step. Real discovery workflows chain many steps: dispense reagents, seal a plate, incubate it for hours at controlled temperature, spin it in a centrifuge, read it on a plate reader, then move the results downstream. A workcell is the integration of these separate instruments into one continuous automated line, with a robotic arm, floor-standing or, increasingly, a collaborative arm, as the courier that carries plates between stations.

The engineering challenge here is not any single instrument; each is a mature product. It is the interfaces between them. Every instrument speaks its own control protocol, expects plates presented at a specific position and orientation, and runs on its own timing. Integrating a dozen of them into a system that runs unattended overnight, recovers from a jammed plate without a human, and logs everything for traceability is genuinely hard systems engineering. This integration work is why workcells have historically been bespoke, expensive, and slow to deploy, closer to a construction project than a purchase.

The orchestration layer: the operating system of the lab

" The robotic arm is not the brain. The brain is the orchestration and scheduling software, the layer that sequences every instrument, allocates shared resources, resolves timing conflicts, and reroutes work when something fails. If two experiments both need the plate reader at 2:14 a.m., the scheduler decides who waits and reshuffles the entire downstream plan so neither assay's incubation window is violated. This is a constraint-solving problem running continuously against a live physical system.

This software is the real lock-in, for three reasons. First, switching cost: once a lab's protocols, instrument drivers, and validated workflows are encoded in one scheduler, moving to another means re-integrating and re-validating everything, in a regulated environment where re-validation is expensive and risky. Second, integration equity: the value the customer paid for is not the arm but the months of driver-writing and workflow-tuning that made a specific set of instruments run together, and that lives in the software. Third, margin: hardware is a capital good that commoditizes; orchestration software carries recurring, high-margin economics and deepens with every protocol the customer adds.

The physical arm, by contrast, is commoditizing. General-purpose collaborative robots, cobots, are increasingly good enough to move plates, and they are cheaper and more flexible than the proprietary arms that used to define a workcell. As the arm becomes a generic component, value migrates decisively up to the scheduling layer that tells it what to do. This is the central investment insight of the chapter: do not pay for the muscle; pay for the operating system.

The pure orchestration players are almost entirely private. Automata (private) markets its software explicitly as the "OS for AI-ready labs." HighRes Biosolutions (private) is known for its Cellario scheduler. Peak Analysis & Automation (private) and Hudson Robotics (private) are established integrators. The most investable path runs through the strategics: Biosero, a leading orchestration provider, sits inside Bio-Techne (TECH), and per public reporting, Merck KGaA is in the process of acquiring Bio-Techne, which would fold that orchestration asset into a larger strategic. The pattern to watch is consolidators buying the scheduling software, because that is where the recurring economics and the switching cost concentrate.

General-purpose arms and mobile, autonomous chemistry

The commoditization of the arm has a name and a market leader. Universal Robots, inside Teradyne (TER), makes the default cobot deployed across manufacturing and, increasingly, the lab, a safe, reprogrammable arm that any integrator can drop into a workcell. KUKA, now controlled by China's Midea, is the other major industrial-arm name, and its ownership is a supply-chain consideration for Western pharma buyers wary of Chinese-controlled infrastructure in a sensitive workflow.

Beyond the fixed workcell, the next step is mobility and autonomy. Mobile robots that navigate between stations remove the constraint that everything must sit within one arm's reach, letting a lab scale by adding stations rather than rebuilding a cell. Autonomous chemistry pushes further: closed-loop systems where an AI proposes the next experiment, the hardware runs it, and the result feeds back into the model with no human in the loop. Multiply Labs (private) applies robotics to cell-therapy manufacturing, building its systems on NVIDIA's Isaac robotics and Omniverse simulation stack, a direct bridge between physical AI and biomanufacturing. Kebotix (private) and Chemify (private) are building autonomous, self-driving chemistry platforms that couple generative models to robotic synthesis.

Automation is also reaching the parts of the pipeline usually treated as purely manual craft. VistaPath Bio (private) automates pathology specimen grossing, the hands-on dissection and sampling of tissue that has resisted automation and gates the throughput of the diagnostic lab. It is a reminder that "the hands" extend well past the screening deck into diagnostics and manufacturing.

Proof that the loop can close

The strongest evidence that this stack works end-to-end comes from Ginkgo Bioworks' cloud lab, where an AI agent was set loose to run tens of thousands of experiments autonomously across DNA-focused workflows. That is the full loop operating at scale: a model proposing work, robotics executing it, and data returning to the model without a human sequencing each step. It demonstrates that the orchestration layer, not the arm, is what makes autonomy real, the scheduler is what lets one agent command a physical lab.

How to own the layer

For an investor, the hands are a frustrating category to access, and that frustration is itself the signal. The value sits in the orchestration software, and that software is overwhelmingly private (Automata, HighRes, Biosero pre-acquisition) or buried inside diversified strategics (Beckman in DHR, Bravo in A, Biosero moving into a Merck KGaA-owned Bio-Techne). Tecan (TECN.SW) is the only clean listed pure-play, and it sits in hardware, the commoditizing end of the stack. The general-purpose arm is best owned through Teradyne (TER) via Universal Robots, but that is a bet on cobots broadly, not on drug discovery specifically.

The disciplined read is to track the orchestration layer as the location of durable lock-in and margin, and to gain exposure primarily through the strategics acquiring it, because when a consolidator buys a scheduler, it is buying the operating system of the automated lab, and everything downstream of that decision runs on their software. The muscle is for sale everywhere. The nervous system is what is scarce.

© BEP Holdings · Ben Pouladian. Research and commentary, not investment or medical advice.