Lab Automation & Closed-Loop Drug Discovery
Robotics, not AI, is the bottleneck. The build step has not kept pace with the design step.
BLUF Summary
The advancement of tertiary protein folding predictors in tandem with AI have accelerated lead hit identification through in silico SAR studies, but robots lack the dexterity that innovative drug discovery demands across a lab automation market that reached $6.36B in 2025 and is projected to exceed $24B by 2035. Excelsior Sciences is addressing this challenge by redesigning chemistry itself to be machine-executable, while Recursion has proven that automated biology generates defensible data moats, but the companies that solve physical execution will own the next decade of drug discovery infrastructure.
Background & historical context
How small molecules are made, and how that is changing
Small molecule drugs have historically been synthesized by trained medicinal chemists using solution-phase and solid-phase techniques refined over decades. Solid-phase peptide synthesis, pioneered by Merrifield in 1963, was the first meaningful step toward automation; anchoring a growing peptide chain to a resin support allowed iterative coupling steps to be performed without isolating intermediates. Solution-phase synthesis of non-peptide small molecules remained largely artisanal: a chemist designs a retrosynthetic route, executes each step at the bench, purifies intermediates by column chromatography, and confirms structure by NMR and mass spectrometry. Combinatorial chemistry in the 1990s attempted to industrialize this process but produced quantity over quality. Modern medicinal chemistry has returned to a quality-first model guided by structure-activity relationship data at each iteration.
Automation has progressed unevenly. Liquid handling platforms from Tecan and Hamilton automated dispensing but cannot execute the bespoke synthetic steps that novel drug candidates require. Flow chemistry platforms like Chemspeed advanced parallel synthesis for well-characterized chemistries. The self-driving lab concept, demonstrated by Cooper's group at Liverpool and Cronin's at Glasgow, extended automation into closed-loop experimental design, but within narrow chemical domains. Excelsior Sciences' smart bloccs platform redesigns the chemistry itself to be machine-executable rather than adapting robots to perform human chemistry.
The first true automation milestone. Merrifield built a machine that could synthesize peptides step-by-step without a chemist's hands at each stage. Nobel Prize 1984. This was the conceptual seed of everything that followed.
Geysen's multi-pin technology and Houghten's tea-bag method allowed synthesis of thousands of peptide variants in parallel. The pharmaceutical industry could now generate compound libraries rather than synthesizing molecules one at a time.
Big pharma invested billions in robot-controlled screening platforms capable of testing hundreds of thousands of compounds against biological targets. The promise was to screen everything and find something. In reality, massive attrition rates revealed that volume without quality is insufficient.
Tecan, Hamilton, and Labcyte (acoustic dispensing) industrialized liquid handling. Robots became extremely good at pipetting, diluting, and dispensing in fixed, pre-programmed workflows. But they remained brittle: good at tasks they were programmed for, unable to adapt.
Companies like Schrödinger, Atomwise, and early Recursion began applying ML to molecular design and target identification. AI could propose molecules. But testing them still required human-paced labs. The design-execution gap opened.
The Cronin group at Glasgow and the Aspuru-Guzik group at Harvard demonstrated true closed-loop discovery: a robot proposes, executes, measures, and redesigns experiments without human intervention. Liverpool's mobile robotic chemist ran 688 experiments autonomously, a proof of concept for the entire industry.
Ginkgo Bioworks launches its Autonomous Lab offering with Reconfigurable Automation Cart (RAC) systems sold directly to pharma and national labs. Strateos offers robotic cloud labs accessible remotely. The market for autonomous lab infrastructure becomes real.
Arctoris acquires Eli Lilly's San Diego automated research platform, significantly expanding its automated wet-lab capabilities. Recursion merges with Exscientia ($688M deal), adding generative chemistry AI to its phenomics platform. The Lilly divestiture signals big pharma working through build-vs-buy on lab automation.
Chemistry and biology foundation models proliferate rapidly. Isomorphic Labs, Insilico, and open-source alternatives make molecular design increasingly accessible. The question shifts: if design is cheap, who owns execution? The answer matters enormously, because the physical validation step has not gotten proportionally faster. The design-execution gap widens.
Excelsior Sciences raises $95M (a $70M Series A led by Deerfield, Khosla Ventures, and Sofinnova, plus a $25M Empire State Development grant) in December 2025 to commercialize 'smart bloccs', modular, machine-readable chemical building blocks. Instead of making robots do human chemistry, redesign the chemistry for what machines can reliably execute. Synthesis cycles compress from 4 months to ~2 weeks.
With AI design largely commoditized, proprietary wet-lab throughput is now the clearest source of defensibility. Full-stack players (Recursion, Lila Sciences) compete against chemistry-first approaches (Excelsior Sciences) for the same wedge: who can generate validated experimental data faster and cheaper than anyone else. Big pharma accelerates automation investment internally while also acquiring or partnering externally. The autonomous lab is no longer a research curiosity. It is a strategic asset.
OpenAI and Molecule.one connected GPT-5.4 to Maria, an agentic chemistry system layered on high-throughput experimentation. Given a goal, it proposed TEMPO as an additive to improve Chan-Lam coupling of primary sulfonamides and validated the idea across 10,080 reactions, raising mean yield from 16.6% to 25.2%. The achievement is real but narrow. It optimizes the yield of a single known reaction class rather than exploring broad chemical space, and the underlying automation is high-throughput experimentation within a fixed, well-defined reaction type. The intelligence layer continues to mature; the physical capability to build novel molecules across broad chemical space does not advance with it.
Market size & opportunity
The numbers behind the bottleneck
The lab automation market reached $6.36B in 2025 and is projected to reach $24.02B by 2035, an implied CAGR of roughly 14%. This top-line figure understates the opportunity because it combines mature point-solution hardware, such as liquid handlers and plate readers, with the genuinely unsolved layer of flexible, closed-loop synthesis. The market is bifurcating. Incumbents including Tecan, Hamilton, and Beckman Coulter dominate a commoditizing hardware segment optimized for fixed, repetitive tasks; a thin layer of well-capitalized entrants competes to own the autonomous discovery loop itself, and capital is concentrating there. Strategic validation has been unambiguous. Recursion completed a $688M all-stock merger with Exscientia in 2024, Eli Lilly divested its automated facility to Arctoris, and Excelsior raised $95M in December 2025 (a $70M Series A led by Deerfield, Khosla Ventures, and Sofinnova, plus a $25M Empire State Development grant). Each transaction signals that acquirers and specialist investors believe the build-and-test layer, rather than molecular design, is where durable value accrues.
The investment logic follows from the bottleneck thesis. AI design tools have commoditized; AlphaFold3 and generative chemistry models are now broadly accessible, collapsing the moat at the design layer. As that step cheapens, the binding constraint shifts downstream to physical execution, where throughput remains gated by human chemists and bespoke synthesis. The companies positioned to capture value are therefore not those generating more molecular hypotheses, but those compressing the cost and time of building and testing them. This is a picks-and-shovels thesis; the most defensible returns come from owning the infrastructure layer, including proprietary machine-executable chemistry, automated synthesis platforms, and the data those platforms generate, rather than from betting on any single therapeutic asset.
Competitive landscape
Drug discovery automation: Design · Make · Test
Hover any company for funding stage and description.
Deep Dives
Excelsior is the purest expression of the thesis. Rather than teaching robots to perform human chemistry, it redesigns chemistry to be natively machine-executable. The smart bloccs platform builds on iterative carbon-carbon bond formation using designer building blocks from Marty Burke's University of Illinois lab, converting synthesis into a modular, predictable sequence that robots can run end to end. The platform compresses a synthesis cycle that traditionally takes four months across US, Asian, and contract labs into roughly two weeks in a single automated facility. The strategic significance lies in the data layer; smart bloccs function as a modular chemical language that AI can read and optimise within a closed-loop system, integrating discovery and scale-up chemistry in one platform. Founded in 2024, the company raised $95M in December 2025, with lead investor Deerfield framing it as the missing piece that makes AI-designed drugs physically realisable. The reshoring thesis adds a strategic tailwind; the platform positions itself to rebuild domestic discovery and manufacturing capacity as the industry commits roughly $250B to supply chain security. The principal risk is coverage. Smart bloccs chemistry addresses a meaningful but bounded region of drug-like chemical space, and the breadth of that coverage is the central diligence question.
Two routes to machine-executable chemistry
Excelsior Sciences
Core bet
Redesign the chemistry so machines can execute it reliably.
Mechanism
Smart bloccs, modular building blocks advanced by iterative Matteson homologation with designer B-TIDA reagents, descended from the Burke lab boronate lineage.
Claim
High-fidelity automated synthesis within a defined region of drug-like chemical space. Cycle time compressed from roughly four months to roughly two weeks.
Origin
Founded 2024. $95M raised December 2025, a $70M Series A led by Deerfield, Khosla Ventures, and Sofinnova plus a $25M Empire State Development grant.
Principal risk
Coverage. Smart bloccs address a bounded region of chemical space, and the breadth of that coverage is the central diligence question.
Chemify
Core bet
Keep conventional chemistry and build a universal language that machines can execute.
Mechanism
Chemputation. The chi-DL chemical programming language compiles synthetic routes into hardware-independent code executed on a ChemPU.
Claim
Universality. Cronin has formalized chemputation as a Chemical Synthesis Turing Machine, arguing a chemputer with sufficient reagents and dynamic error correction can produce any stable, isolable molecule in finite time. Validated against more than 100 chi-DL programs.
Origin
University of Glasgow spinout from the Cronin Group. Series B of $50M-plus led by Wing Venture Capital with Insight Partners. First Chemifarm opened June 2025.
Principal risk
Abstraction does not remove physics. A programming language standardizes description, but it inherits every physical limitation of conventional chemistry, including powder handling, air-sensitive transfer, and workup. Universality is a formal claim before it is an operational one.
The two companies are solving adjacent problems that are often conflated. Chemify addresses the translation problem, which is how to describe a synthesis unambiguously enough that a machine can execute it reproducibly. Excelsior addresses the execution problem, which is what chemistry a machine can reliably perform at all. Chemify's universality claim is formal and rests on error correction holding across long step counts; the Cronin group has extended the platform toward air-sensitive organometallic work, but a chemical programming language does not by itself teach a robot to weigh a solid or scrape a vial. Excelsior accepts a narrower chemical space in exchange for higher automation fidelity within it. For the thesis, both validate the central claim that the Make step is the constraint, and they represent the two credible architectures for relieving it. The diligence question that separates them is empirical rather than philosophical: whether Chemify's error correction holds across long multi-step routes on physically awkward operations, and whether Excelsior's accessible chemical space is large enough to matter commercially.
Recursion produces world-class work and operates the strongest data-generation engine in the space, but it optimises around the build bottleneck rather than solving it. The model is to predict more so that less is made; machine learning on massive phenomic datasets, drawn from up to millions of wet-lab experiments weekly, simulates which molecules are worth synthesising, so the physical lab only builds the compounds the system already favours. The results are notable. Recursion synthesises roughly 330 compounds to advance one candidate against an industry norm of 2,500 to 5,000, and does so in roughly 17 months against an industry average above 40. The 2024 Exscientia merger added an automated small-molecule synthesis capability. The distinction that matters for this thesis is narrow but important; Recursion minimises the build step through prediction, but it does not transform the underlying chemistry. When synthesis is required, the company relies on conventional automated chemistry and precision-design workflows, failing cheaply in silico rather than making the physical build step itself fundamentally faster or more flexible. That approach is defensible and impressive, yet it leaves the core synthesis bottleneck intact for any molecule the models cannot confidently pre-screen, which is precisely the novel, first-in-class chemical space where the bottleneck binds hardest.
Ginkgo is the cautionary financial case in the category, and its pivot is instructive. After years operating as a royalty-and-services cell programming platform, the company is restructuring toward an Autonomous Lab and tools-and-automation model, anchored by its Reconfigurable Automation Cart system sold as lab-as-a-service. The financials explain the urgency. FY2025 revenue fell to roughly $170M against a net loss of roughly $313M, and the company has cut aggressively, achieving over $250M in annualised cost savings and reducing quarterly cash burn roughly 75% year-over-year by Q3 2025. Cell engineering revenue dropped sharply, falling roughly 61% year-over-year in Q3 2025 on restructuring and a shift toward larger clients, and the stock trades far below its SPAC-era highs at a market capitalisation near $550M. Year-end liquidity stood at roughly $422M against a stated target of adjusted EBITDA breakeven by mid-2026. The implication for this thesis is twofold. Ginkgo validates that demand for autonomous lab infrastructure is real enough to redirect an entire company toward it, but it also demonstrates that owning general-purpose automation infrastructure is capital-intensive and margin-challenged. The lesson favours specificity; machine-native chemistry, as pursued by Excelsior, over general-purpose automation services. Whether the breakeven target holds is the figure to monitor, since a liquidity crunch remains a live risk.
Potato is the orchestration-layer bet, the Scientific OS that sits above the hardware. Its multi-agent system, Tater, converts plain-language research intent into lab-ready, robot-executable protocols grounded in peer-reviewed literature, analyses raw data, and connects steps into end-to-end workflows. The capability most relevant to this thesis is its ability to bridge the gap between intent and robot-ready scripts, translating protocol steps into code that robots can run. The company is seed-stage, having raised roughly $4.5M, and sits in a genuinely new category that does not yet have a settled name. The strategic question is durability. Orchestration is necessary but is the layer most exposed to absorption; full-stack players such as Lila and Recursion and instrument vendors both have incentives to own it, and an orchestration layer can only schedule the physical tasks that robots are already capable of performing. Potato makes the existing automatable surface more usable, but it does not expand what chemistry is physically automatable. It is valuable as connective tissue and vulnerable as a standalone moat.
The connective tissue layer
Between the chemistry and the hardware sits a software layer that determines whether an automated lab functions as a system or as a collection of instruments. This layer covers instrument connectivity, data harmonization, protocol translation, and workflow orchestration. It is distinct from the synthesis platforms and from the hardware vendors, and it is where a meaningful share of automation spend now goes. The category has no settled name; participants variously describe it as a scientific data cloud, a lab digital core, or a scientific operating system.
The clearest venture play
Potato is the most direct expression of the orchestration thesis. Its multi-agent system, Tater, converts plain-language research intent into lab-ready, robot-executable protocols grounded in peer-reviewed literature, then analyzes the resulting data and chains steps into end-to-end workflows. The distinguishing capability is that it bridges intent and robot-ready scripts rather than moving data between systems that already exist. The company is seed stage with roughly $4.5M raised and connects to more than 100 instruments. It is early, and the category is new enough that its boundaries are still forming.
Incumbent data infrastructure
TetraScience
Tetra Scientific Data and AI Cloud. Positions on the largest integration network of lab instruments and informatics applications, harmonizing scientific data into standardized formats to make it usable by AI. Enterprise scale. Implementation is resource-intensive and requires sustained IT support.
Scitara
Scitara DLX, a vendor-neutral connectivity and orchestration platform using proprietary connectors for plug-and-play integration across instruments, ELNs, LIMS, and web services, with bi-directional data exchange. Enterprise-focused with implementation cycles commonly exceeding six months.
Ganymede
Lab data cloud connecting instruments and informatics systems into a unified pipeline. Competes directly with Scitara and TetraScience on integration breadth, with similarly long deployment cycles.
Scispot
Lighter-weight alternative positioning on rapid deployment. GLUE integration engine for connecting instruments and applications, with an AI assistant layer. Targets labs unwilling to absorb enterprise implementation timelines.
Benchling
The dominant system of record. ELN and LIMS infrastructure across most automated pharma and biotech labs, functioning as the default data layer rather than an orchestration engine.
Synthace
Experiment design and automation orchestration, strongest in structured design-of-experiments workflows. Bridges the handoff between planning and execution.
What Recursion built instead
Recursion did not buy this layer. It built the Recursion Operating System, a proprietary platform spanning target identification through clinical trial enrollment, backed by tens of petabytes of internal phenomics, transcriptomics, proteomics, and ADME data. In January 2024 it added LOWE, an LLM-Orchestrated Workflow Engine that gives scientists a natural-language interface to orchestrate both wet-lab and dry-lab operations, chaining steps from querying its internal Maps of Biology and Chemistry through generating novel compounds and scheduling them for synthesis and experimentation. LOWE was demonstrated alongside NVIDIA, the same partnership behind the BioHive supercomputers.
Investment read
Recursion's decision to build rather than buy is the most informative fact in this category. The most sophisticated operator in automated drug discovery concluded that no external vendor could supply orchestration at the fidelity its loop required. That cuts both ways. It confirms the layer is strategically important, and it identifies the structural risk for standalone vendors, which is that the buyers with the most acute need are also the most capable of building internally, while instrument manufacturers and full-stack platforms both have incentives to absorb the function. The deeper limitation is architectural. Connective tissue raises utilization of existing physical capability; it does not expand what chemistry a laboratory can perform. Orchestration can only schedule tasks the hardware can already execute, which is why this layer is necessary for the thesis to resolve and insufficient to resolve it.
Physical AI & the dexterity frontier
The most severe limitation identified in this thesis is dexterous manipulation, the inability of laboratory robots to perform the fine physical operations a trained chemist executes without thought. The most credible path to relieving that limitation is not coming from laboratory automation vendors. It is coming from general-purpose robotics, where foundation models are being trained to generalize manipulation across tasks and objects rather than to execute fixed programmed motions.
NVIDIA's stack
NVIDIA has assembled the reference architecture for this category. Cosmos provides world foundation models trained on video, robotics data, and simulation, enabling closed-loop simulated environments at a scale that would be prohibitive to collect physically. Isaac GR00T provides vision-language-action models purpose-built for robots, with GR00T N1.6 released at CES in January 2026 as an open reasoning VLA supporting full body control and using Cosmos Reason for contextual understanding. Isaac Lab and Isaac Sim provide the training and evaluation environment, with Isaac Lab-Arena added for large-scale benchmarking and Isaac Lab 3.0 introducing multiphysics simulation and improved support for complex dexterous manipulation. At GTC in March 2026, GR00T N1.7 entered early access with commercial licensing and explicit support for advanced dexterous control, and NVIDIA announced ecosystem partners including ABB Robotics, Agility, FANUC, Figure, KUKA, Medtronic, Skild AI, Universal Robots, and YASKAWA. Cosmos 3 followed on May 31, 2026, shipping open variants at 16B and 64B parameters. The stated position from NVIDIA's chief executive is that physical AI has arrived and every industrial company will become a robotics company.
Two readings
The accelerating case
If GR00T-class models deliver general dexterous manipulation, the physical bottleneck erodes faster than this thesis assumes. A robot that learns manipulation in simulation and transfers those skills to unfamiliar objects is precisely the capability laboratory automation has lacked, and NVIDIA is commoditizing the robot control layer on the same pattern by which it commoditized the model training stack. Under that scenario, value migrates away from bespoke automation hardware and toward whoever holds the chemistry, the protocols, and the proprietary data.
The thesis-consistent case
None of this stack is chemistry-specific. General manipulation benchmarks do not address air-sensitive transfer, accurate mass measurement of solids, solvent handling, glassware, or the hazard constraints that define synthetic operations. A capable general robot brain does not supply chemistry-specific end effectors, validated protocols, or instruments designed for robotic rather than human operation. Physical AI is best understood as compressing the timeline on one of several critical limitations rather than removing the constraint, and the compression applies to the layer where no chemistry-specific player has yet established a position.
Where the white space sits
No company has emerged as the clear leader in applying physical AI foundation models specifically to chemical synthesis. Laboratory automation vendors build fixed-function hardware; physical AI developers target humanoids, manufacturing, and surgical robotics. Medra's instrument-agnostic approach, in which general-purpose robots operate existing laboratory equipment through its physical controls, is the nearest available analogue and remains early. A company combining GR00T-class manipulation models with chemistry-specific tooling and protocol data would sit directly on the thesis. That position is currently unoccupied.
The bottleneck in detail
Tasks like sample scraping, handling of solid powders, weighing of small quantities, or operating equipment designed for human hands remain very difficult. Researchers at Liverpool are still publishing papers on how to teach robots to scrape vial walls in a circular motion, a task a chemist does without thinking.
Most synthesis workflows include at least one step that requires human dexterity. Until this is solved, fully unattended synthesis is not possible.
Liquid handling is mature. Solid handling is not. Weighing powders, loading solid reagents, and handling heterogeneous reaction mixtures all require physical manipulation that fluidic systems cannot perform. This is a hard constraint, since most drugs begin as solid-state molecules.
The majority of small-molecule pharmaceutical synthesis involves solids at some stage. This is the single largest gap between what robots can do and what medicinal chemists do daily.
Most lab instruments were designed for humans. A robot must operate equipment not built for robots. This requires custom engineering for every instrument integration. 'Integration of human-centered equipment is one of the biggest challenges in SDL hardware development,' per Cell/Matter (2024).
Every new instrument in a lab workflow represents a new integration project. This dramatically increases the engineering cost of building a fully automated workflow.
Complex synthetic routes involving non-standard conditions (high heat, inert atmospheres, reactive reagents, Schlenk techniques) are not yet automatable. Most robot platforms work in narrow chemical space. Excelsior's 'smart bloccs' is a direct attempt to redesign chemistry to fit within what machines can reliably execute.
First-in-class drugs often require novel synthetic routes with conditions outside the operational envelope of current robotic systems.
Human chemists recognize when something has gone wrong, such as a color change, unexpected precipitate, or equipment malfunction, and adapt. Robots generally cannot. They execute their scripts even when the chemistry has failed. Embedding adaptive sensing and decision-making into physical lab workflows is an open problem.
Without anomaly detection, robotic systems produce large quantities of failed experiments without flagging the failure. Yield and quality suffer.
Early autonomous labs relied on a single characterization technique. Liverpool's 2024 paper was notable for combining UPLC-MS and benchtop NMR with a mobile robot — but human chemists draw on far more instruments and use contextual judgment. Scaling multi-modal analysis into an autonomous decision loop remains hard.
Limited analytical feedback means the autonomous system has incomplete information about reaction outcomes, reducing its ability to generate useful experimental data.
Field note: reading the 2026 OpenAI chemist result
In June 2026, OpenAI and Molecule.one published a result in which GPT-5.4 was connected to Maria, an agentic chemistry system integrated with a high-throughput laboratory. Given an open-ended objective, the system proposed TEMPO as an additive to improve Chan-Lam coupling of primary sulfonamides and validated the hypothesis across 10,080 reactions in two cycles, raising mean yield from 16.6 percent to 25.2 percent, with bench-scale confirmation by human chemists for 11 of 14 substrate pairs. The engineering is real and the closed-loop infrastructure functions. The question worth asking is what the result actually demonstrates.
The chemistry is not novel
TEMPO as an oxidant and additive in copper-mediated coupling is established chemistry, and the authors cite the prior reports themselves. The system did not discover an unknown reagent effect; it rediscovered a known one through exhaustive search. That is a meaningful engineering demonstration and a weak scientific one.
The effect is not general
The work itself reports substrate pairs where the TEMPO additive is harmful at bench scale. An additive that improves some couplings and degrades others has not been established as a general method, and the framing of a broadly improved reaction is stronger than the underlying data supports.
The yields remain poor and the dominant pathway is unchanged
A mean yield improvement from 16.6 percent to 25.2 percent leaves the transformation firmly in the low-yielding regime. Oxidative deboronation remains a major product pathway, which means the principal competing reaction was not suppressed. Optimizing around a dominant side reaction is not the same as solving it.
The target compound class is trivially accessible by classical means
The optimized reaction using 4-hydroxy-TEMPO produces N-aryl sulfonamides. That compound class is routinely accessible by sulfonylation of the corresponding aniline, a first-year transformation requiring a sulfonyl chloride and a base. The value of optimizing a low-yielding Chan-Lam route toward products obtainable in one straightforward step is not established by the work.
Throughput substituted for insight
More than 10,000 reactions were consumed to reach a conclusion that a trained chemist could have proposed from the existing literature. The reasoning output is at the level of an undergraduate mechanistic argument, and the scale of experimentation reflects brute-force search rather than hypothesis quality. High-throughput experimentation with a planning layer is a legitimate tool, but the result should be read as a demonstration of experimental throughput, not of scientific reasoning.
Implication for the thesis
The result is the clearest available illustration of the distinction this thesis rests on. The system searched exhaustively within a single known reaction class and returned a known additive with a modest effect. It optimized a reaction; it did not explore chemical space, and it did not expand the set of molecules that can be physically made. The automation involved is high-throughput experimentation within a fixed, well-defined reaction type, which is precisely the category of automation that has existed for decades and remains brittle outside its programmed envelope. The intelligence layer continues to advance and now reaches into optimization. The physical capability to construct novel molecules across broad chemical space did not move. The secondary risk is narrative rather than technical: demonstrations of this profile can create an impression that synthesis is being solved while the binding constraint remains where it was.
Steelmanning the bears
“AI will make the physical execution step irrelevant by designing better molecules computationally.”
strengthensAI models require experimental validation to generate the training data they depend on. Even perfect computational proposals must be synthesized and tested. The loop cannot be closed without physical execution. Better AI makes the bottleneck more acute, not less.
“Big pharma will build this internally rather than buy from startups.”
complementaryArctoris's acquisition of Lilly's automated facility and the Recursion/Exscientia merger are evidence of the opposite dynamic. The infrastructure layer is consolidating and scaling. Internal builds in pharma are notoriously slow; Lilly divesting its automated lab validates that the specialist CRO model is winning.
“The robotics hardware cost is prohibitive for broad adoption.”
strengthensCommercial industrial robot costs have been decreasing for 20 years and continue to fall, while chemistry hardware has barely changed in real terms. The cost asymmetry is resolving in robotics' favor, as Liverpool's researchers noted: robots are used everywhere while chemistry instrumentation remains a specialty market.
“This is a services business, not a platform — margins will be thin.”
monitorThe most defensible plays (Recursion, Lila, Excelsior) hold proprietary data or chemistry IP, not services. The services layer (Arctoris, Strateos) is indeed margin-constrained, which is why the thesis centers on companies with platform characteristics.
“The ordering of the loop matters — Bessemer and others frame it as Design-Test-Make-Analyze, not Design-Make-Test. Does this undermine the Make bottleneck thesis?”
strengthensThe loop ordering is workflow-dependent and actively debated. In hit-to-lead optimisation, Test can precede Make. In de novo synthesis programs, Make precedes Test. Regardless of ordering, the thesis is about the Make step being structurally undersupplied relative to Design and Test. Bessemer's own language confirms this: the Make step requires skilled human labor at nearly every stage and is one of the biggest bottlenecks in shortening timelines. The ordering debate does not change the supply gap.
Quotes & signal
Published BVP Atlas piece naming closed-loop lab automation as one of three core principles for the next generation of drug discovery companies. Explicitly identifies wet lab execution as the primary bottleneck. Features Medra, Automata, Dash Bio, and Lila Sciences as key companies in the space.
Wet lab experiments are slow, prone to failure, and dependent on skilled human labor, making them one of the biggest bottlenecks in shortening drug development timelines.https://www.bvp.com/atlas/building-biology-native-data-infrastructure-for-the-ai-era
Though Hamilton robots for liquid handling and Chemspeed platforms for automated synthesis have existed in labs for decades, they're optimized for the high throughput of specific point tasks rather than the automation and integration of entire experimental workflows.
Commercial industrial robot costs have been decreasing for the last 20 years and they are still going down, whereas chemistry hardware has basically not changed at all. It's just gone up with inflation. The reason for that is that chemistry is a very specific discipline, whereas robots are used for everything.
Integration of human-centered equipment is one of the biggest challenges in SDL hardware development.
We are not trying to make robots do human chemistry. We are redesigning chemistry for what machines can actually do.
References
- 1.MarketsandMarkets, 2025 · lab automation market size, 2025
- 2.SNS Insider, 2026 · projected by 2035, ~14% CAGR
- 3.Acta Pharm Sin B, 2022 · of drug candidates fail in clinical trials
- 4.Bessemer Venture Partners: Building biology-native data infrastructure for the AI era, BVP Atlas, April 2026
- 5.Prof. Andrew Cooper, University of Liverpool: On the cost asymmetry between robotics and chemistry instrumentation, 2024
- 6.Cell/Matter, 2024: On the cross-instrument integration problem in self-driving labs
- 7.Excelsior Sciences: On the 'smart bloccs' design philosophy and the core insight behind the company
- 8.OpenAI and Molecule.one, June 2026: A near-autonomous AI chemist improves a challenging reaction in medicinal chemistry
- 9.NVIDIA, January 2026: NVIDIA Releases New Physical AI Models as Global Partners Unveil Next-Generation Robots
- 10.NVIDIA, March 2026: NVIDIA and Global Robotics Leaders Take Physical AI to the Real World
- 11.Cronin, Pagel and Sharma: Chemputer and chemputation, a universal chemical compound synthesis machine, PNAS
- 12.Recursion: LOWE, LLM-Orchestrated Workflow Engine
Elliot Silk · silkelliot@gmail.com · linkedin.com/in/elliotsilk
Confidential · 9/3/2026