Why AI for Science May Need Experimental Infrastructure More Than Bigger Models

AI models can propose molecules, targets, and biological hypotheses at a pace that conventional laboratories struggle to match. MegaRobo’s thesis is that this imbalance makes experimental infrastructure—not simply more compute or more sophisticated algorithms—the decisive constraint for AI for science and drug discovery. 59

The company’s proposed answer is a lab-in-the-loop: an integrated system in which instruments and robots generate reliable evidence, AI interprets that evidence, and the results automatically shape the next experiment.

The real bottleneck is verified biological data

A model’s prediction about protein structure, binding, efficacy, toxicity, or clinical response is not yet a drug-development result. It is a proposition that must be tested in the physical world.

That distinction matters because biological evidence is difficult to standardize and difficult to transfer between stages. A result from a simplified assay may not predict what happens in a more complex biological system, an animal, or a patient. The valuable dataset is therefore not just a collection of molecular structures. It is a connected record linking experimental conditions and measurements across screening, dose response, efficacy, safety, pharmacology, and clinical outcomes.

MegaRobo frames the problem as a data-and-workflow challenge: experiments must produce trustworthy, structured information that models can use without losing context through manual transcription or inconsistent procedures. 59

Why more compute cannot solve the physical-validation gap

More capable models can generate or rank far more candidates than a traditional lab can synthesize and assay. But every candidate still requires physical validation, with controls, reproducible sample handling, calibrated instruments, quality checks, and traceable metadata.

Human researchers remain essential for scientific judgment and handling exceptions, but human-operated workflows are constrained by labor, scheduling, throughput, and variation. Automation changes the scale of the problem. One robotic quantitative high-throughput screening platform, for example, generated more than 6 million concentration–response curves from over 120 assays in three years. 33

That example does not mean every experiment can be automated or that quantity alone creates useful knowledge. It shows why the physical layer is becoming as important as the model layer: AI can only learn from experiments that are executed consistently and recorded in a form suitable for analysis.

MegaRobo’s Perception–Conception–Execution architecture

MegaRobo describes its system as a three-part loop:

• Perception: instruments, imaging, machine vision, sample identification, and sensor systems observe laboratory state and capture results.
• Conception: AI models or agents interpret results, select experiments, and reason about what to try next.
• Execution: robotic arms, liquid handlers, incubators, assay systems, and process controls carry out the chosen procedures.

The important feature is not any individual component. It is the connection between them. In a closed-loop laboratory, experimental results are machine-readable and available to the planning layer quickly enough to influence the next cycle. MegaRobo’s Kunpeng Laboratory is described as combining robotics, automation, and AI while also accumulating data for AI-for-science applications. 9

Recent descriptions of MegaRobo facilities similarly portray AI agents coordinating robotic arms, incubators, and liquid-handling workstations, with physical test results feeding subsequent experimental decisions. 4

From laboratory automation to machine-native instruments

MegaRobo’s development story can be understood as a progression from automating tasks to redesigning the laboratory around autonomous machine operation.

Stage 1.0: automate repeatable workflows

The first step is conventional laboratory automation: standardizing and automating repetitive procedures to improve throughput, consistency, and data capture. This creates the operational foundation for more autonomous systems.

Stage 2.0: prove reliability in demanding industrial settings

The second step is to demonstrate that perception, control, and execution can meet stringent industrial requirements. MegaRobo applies capabilities across life sciences and manufacturing, including machine-vision and automation systems for industrial environments. Its public materials describe a broader business spanning laboratory automation, life sciences, and advanced manufacturing. 68

The strategic logic is straightforward: an autonomous scientific system must do more than produce an impressive demonstration. It must identify samples correctly, execute procedures consistently, detect failures, and preserve a reliable record of what happened.

Stage 3.0: build tools for machines rather than people

The final step is to make instruments natively usable by AI agents. Instead of treating robotics as an add-on to human-designed equipment, MegaRobo’s vision is to create tools that expose machine-readable state, accept direct instructions, recover from routine exceptions, and operate as part of a continuous experimental loop. 5

This is a deeper change than replacing a manual pipetting step with a robot. It means redesigning the laboratory’s interfaces, protocols, and data architecture around autonomous operation.

Why the laboratory becomes “experimental compute”

In this model, a laboratory is not just a place where scientists perform procedures. It is a system that converts computational proposals into evidence and evidence into better proposals.

That makes physical infrastructure a potential strategic bottleneck and a potential source of defensibility. MegaRobo’s proposed advantage is not necessarily ownership of every scientific model. It is control of the interface through which models can:

1. observe laboratory conditions;
2. translate scientific goals into executable procedures;
3. obtain consistent, validated results; and
4. accumulate proprietary, high-quality experimental data.

The more tightly those steps are integrated, the more useful each experimental cycle becomes. A disconnected model can suggest what to test, but a connected system can test it, interpret the result, and launch the next cycle with less delay and less information loss.

What the AI-drug-discovery investment cycle reveals

The financial momentum around AI-driven drug discovery demonstrates investor interest, but it does not remove the distinction between a promising candidate and a clinically successful medicine. Insilico Medicine’s Hong Kong IPO in December 2025 raised about $293 million, while company listing materials stated that none of its drug candidates had been commercialized as of the relevant cutoff. 1719

That combination is not proof that AI drug discovery has failed. It is evidence that capital, computational design, and clinical commercialization occupy different points on a long development path. MegaRobo uses this gap to support its infrastructure argument: generating candidates is only one part of the process, while reliable testing and translational validation remain indispensable.

The limits of autonomous laboratories

A faster closed-loop lab does not eliminate the hardest uncertainties in translational medicine. It cannot by itself guarantee that a model is biologically valid, that an animal result will translate to humans, or that a candidate will succeed in clinical trials.

Autonomous systems also introduce their own requirements. They need robust safety controls, anomaly handling, audit trails, and human oversight. A scientific discovery laboratory is not fully autonomous merely because it can run a protocol without a person standing beside the instrument; higher levels of autonomy require systems to complete scientific cycles and interpret routine results while escalating anomalies. 1

MegaRobo’s argument should therefore be read as a claim about enabling infrastructure, not a promise of automatic drug approval. If AI-generated hypotheses continue to outpace experimental verification, the organizations that can produce reproducible, connected, machine-readable evidence may determine which models become scientifically useful.

The bottom line

MegaRobo’s thesis is that the next competitive layer in AI for science will sit between the model and the physical world. Compute can expand the search space, and algorithms can improve prioritization, but neither can replace the experiments needed to establish whether a hypothesis is true.

The “lab-in-the-loop” is the company’s proposed solution: combine perception, AI-driven conception, robotic execution, and continuous data capture into one feedback system. If that infrastructure works at sufficient scale and reliability, the laboratory becomes a form of experimental compute—and control of that layer could matter as much as control of the model generating the hypotheses.