Part Two · The Stack · Layer 8 of 9

The Cloud Lab — Running the Whole Loop as a Service

Everything so far has described the pieces of a modern lab: the models that design, the machines that build, the instruments that measure, the software that remembers. This chapter is about the companies that assemble all of it and rent the finished loop to someone else. If the rest of the primer is a tour of components, this is the tour of the general contractors, the organizations you can hire to run the Design-Make-Test-Learn cycle without owning a single instrument yourself. It is also the layer where the gap between the marketing and the reality is widest, so it demands a skeptical read.

What a Cloud Lab Actually Is

A cloud lab is a physical laboratory that a scientist operates entirely through software, from anywhere, without ever entering the building. You do not ship yourself to the bench; you ship instructions to the robots. A researcher writes an experiment in code, specifying the reagents, volumes, temperatures, instruments, and readouts, submits it, and a remote automated facility physically executes it, then returns the data. The bench becomes an API.

The analogy that gets used, correctly, is cloud computing. Before Amazon Web Services, running software at scale meant buying and racking your own servers. AWS turned that capital expense and operational headache into a metered service you call over the internet. A cloud lab proposes the same trade for wet-lab science: instead of buying instruments, hiring technicians, and maintaining a facility, you rent physical experimentation by the run. The payoff is not only convenience. It is reproducibility, because a protocol written in code executes identically every time, eliminating the silent variation between technicians and days that quietly corrupts so much biological data, and it is throughput, because the facility runs continuously and in parallel in ways a human team cannot.

The Self-Driving Lab, and an Honest Distinction

A cloud lab executes a human's instructions remotely. A self-driving or autonomous lab goes one crucial step further: the AI, not the human, decides what experiment to run next. The machine forms a hypothesis, designs the experiment to test it, runs it, reads the result, updates its model, and chooses the following experiment, the full Design-Make-Test-Learn loop with the human removed from the inner cycle. This is the endpoint the whole field is pointed at, the lab that improves itself.

The overwhelming majority of what is marketed today as an "autonomous" or "self-driving" lab is really the cloud-lab capability above, remote execution plus scheduling, with a human still choosing the experiments. Genuine closed-loop autonomy, where an algorithm meaningfully drives discovery, exists and is real, but it lives mostly in materials chemistry (where the experiments are more standardized and the readouts cleaner) and inside a handful of well-funded private labs. In biology, the messiness of the systems, the difficulty of the measurements, and the regulatory stakes have kept true autonomy closer to the frontier than the product catalog. When you evaluate a company in this layer, the first question is always the honest one: is the AI actually choosing the experiments, or is it scheduling a human's?

The Old Way to Rent the Loop: The CRO

Long before "cloud lab" existed, there was a mature, enormous industry for renting drug-discovery work: the Contract Research Organization (CRO). A CRO is a company that runs research, testing, or clinical trials on behalf of a drug developer. Pharma has outsourced huge swaths of its pipeline to CROs for decades, everything from early safety testing to running the Phase 3 trials that decide a drug's fate. The CRO industry is the incumbent, at-scale way to rent the loop, and it is where most of the public-market money in this layer actually sits.

Automation and AI press on the CRO model from both directions. On one hand, they threaten to compress the labor-intensive, high-margin services CROs sell, if a cloud lab or an AI can do in software what a CRO charged for in people, some CRO revenue is at risk. On the other, the largest CROs sit on something increasingly precious: vast, structured datasets from thousands of past studies and trials, exactly the kind of proprietary data that trains models. The strategic question for each CRO is whether AI erodes its services faster than its data becomes a moat.

Who Rents You the Loop

The cloud labs are mostly private, which tells you the model is still maturing toward public scale:

• Emerald Cloud Lab (private) is the fullest realization of the concept, a facility offering well over a hundred distinct experiment types, all driven by code, effectively an "AWS for the bench." It is the reference implementation of software-defined wet-lab science.
• Strateos and Arctoris (both private) offer robotic, remotely-operated discovery platforms. Culture Biosciences (private) specializes the model to bioprocessing, renting cloud-connected bioreactors so teams can run and monitor fermentation experiments remotely.
• Ginkgo Bioworks (DNA) is the one public pure-play worth naming, and its cloud-lab effort is the layer's headline proof point: an AI agent set loose to run tens of thousands of experiments autonomously across its automated foundry. Ginkgo's own troubled history as a public company is a caution about platform economics, but the demonstration itself, one agent commanding a physical lab at scale, is the clearest evidence the full loop can close.

The autonomous labs are largely private or academic, and cluster, tellingly, in materials science:

• Lila Sciences (private, roughly $550 million raised, backed by Flagship Pioneering) is the best-funded commercial bet on "AI science factories" that run the scientific method autonomously across chemistry, materials, and biology.
• The A-Lab at Lawrence Berkeley National Laboratory is the canonical academic proof point, an autonomous facility that planned, synthesized, and characterized new inorganic materials with minimal human intervention. The Acceleration Consortium at the University of Toronto is the flagship academic program organizing the field. Radical AI (private) is commercializing self-driving materials discovery. Note the pattern: the genuine autonomy is in materials, not yet in drugs.

The tech-enabled CROs are where public investors actually get exposure to "renting the loop":

• Charles River Laboratories (CRL) is the preclinical and safety-testing backbone of the industry, the physical wet-lab work of turning a candidate into something safe enough to put in a human. It is the most direct public exposure to the early-discovery half of the loop.
• ICON (ICLR) and IQVIA (IQV) dominate the clinical half, running trials, with IQVIA layering an enormous health-data-and-analytics business on top that is itself an AI asset. Medpace (MEDP) is the high-margin, small-biotech-focused clinical CRO. Certara (CERT), met earlier in the Design layer, belongs here too as the biosimulation CRO whose models can substitute for physical work in a regulatory filing.

China is the unavoidable complication. WuXi AppTec (2359.HK) and WuXi Biologics are among the largest and most capable CRO/CDMO platforms in the world, deeply embedded in Western pharma supply chains. They now carry a specific regulatory overhang: the proposed U.S. BIOSECURE Act and the Department of Defense's 1260H list, which target certain Chinese biotech vendors. This cuts both ways and investors should hold both edges. A durable decoupling is a tailwind for Western CROs that absorb displaced work; a policy reversal or a US-China thaw unwinds part of that reallocation. The regulatory discount on the WuXi complex is not a simple bear or bull, it is a live, bidirectional catalyst.

How to Own It

The disciplined conclusion is that the most exciting version of this layer, the genuinely autonomous, full-loop-as-a-service lab, is mostly private, mostly early, and mostly proven in materials rather than medicine. Public exposure runs overwhelmingly through the tech-enabled CROs and, as the single pure-play, Ginkgo. That means the layer should be sized for what it is today, not what the pitch decks promise: a real and growing market for renting experimentation, wrapped in an "autonomous" narrative that is still, outside a few labs, more aspiration than product. The right posture is to own the incumbents whose data and scale compound regardless of how fast autonomy arrives, and to watch for the one signal that would change everything, a commercial cloud lab demonstrating genuine closed-loop, AI-chosen drug discovery, not just materials. Until that shows up, treat "self-driving lab" as a direction, not a destination.

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