In a nondescript laboratory—its exact location isn't really the point—what once demanded 10,000 painstaking manual experiments now takes 24. The kind of catalyst research that would have consumed months of a scientist's career runs autonomously, around the clock, with AI agents planning each test, tweaking parameters on the fly, and learning from every outcome.
This isn't a pilot project tucked away in some corporate showcase. It's happening, and the scramble to deploy these "AI co-scientists" is quietly reshaping everything from drug discovery to battery materials research.
The terminology shifts depending on who's talking—self-driving labs, autonomous experimentation systems, robot scientists—but the architecture keeps circling back to the same core idea: AI systems that close the loop on hypothesis, execution, and analysis without requiring human handholding at every turn. Market analysts peg the AI-in-drug-discovery segment alone somewhere between $2.6 billion and $6.9 billion this year, with forecasts climbing to $10.3 billion by 2031 at a 25.9% compound annual growth rate, according to Mordor Intelligence. Broader lab automation adds another $8 to $9 billion, trending toward $14.7 to $18.4 billion by 2034.
That growth isn't just spreadsheet optimism. Billions in committed deals and venture checks underscore an industry betting—hard—that autonomous research platforms will compress timelines and costs in ways manual or semi-automated labs simply cannot match.
Consider Isomorphic Labs, the Alphabet spinout helmed by DeepMind co-founder Demis Hassabis. In January 2024, it inked roughly $3 billion in pharma collaborations with Eli Lilly and Novartis. Then, in April 2025, it reportedly closed a $600 million external funding round led by Thrive Capital. Recursion Pharmaceuticals, meanwhile, completed BioHive-2, a 504-GPU NVIDIA H100 supercomputer that ranked #35 on the TOP500 list as of May 2024—the fastest wholly owned system in biopharma, they claim.
And Chemify, a Scottish startup automating chemical synthesis with its "Chemputation" language (yes, that's what they call it), raised $43 million in 2023 and another $50 million in October 2025, opening a dedicated "Chemifarm" facility by mid-year.
From Narrow Tools to Generalist Agents
Early "robot scientists" like Adam—published in Science back in 2009—demonstrated hypothesis-driven automation in yeast genetics. But those systems were narrow, brittle, and required extensive human setup. What's different now is the fusion of large language models, multi-agent reasoning, active learning algorithms, and workflow orchestration software with physical robotics and cloud-based lab infrastructure. It's a stack, essentially, that didn't exist five years ago.
Berkeley's A-Lab, detailed in Nature in 2023, autonomously synthesized 36 novel inorganic materials in 17 days using closed-loop solid-state synthesis. The system integrated machine learning, density functional theory calculations, robotic sample prep, and real-time literature extraction. Over 1.5 years, it processed more than 3,500 samples using AlabOS, a reconfigurable workflow framework released in May 2024.
Newer architectures go further—perhaps further than some researchers expected. AutoLabs, introduced in September 2025, employs multi-agent, self-correcting logic to translate natural-language instructions into high-throughput protocols, marrying tool-use and reasoning to improve accuracy. Companies like Synthetic Sciences, a YC Winter 2026 startup, are building what they call "AI co-scientists" that triage literature from PubMed and bioRxiv, design experimental methods, orchestrate GPU training runs, and auto-generate LaTeX manuscripts. They're positioning the software as an orchestration layer above both compute and wet-lab execution.
The trend is unmistakable: from narrow optimizers to generalist agents that read, plan, execute, and write. Academic reviews in ACS Chemical Reviews (August 2024) and Nature Synthesis (January 2023) map this progression in detail, noting that the degree of autonomy now spans experiment selection, execution, analysis, and hypothesis updating.
Which raises an obvious question: How much faster are we talking?
Compression, Not Just Acceleration
The efficiency gains, when systems actually work, are staggering. The University of Liverpool's mobile robot chemist, first unveiled in 2020, was followed by multi-robot team demonstrations in 2024. By December 2025, researchers reported reducing catalyst screening from roughly 10,000 manual experiments to 24 autonomous runs, with labs operating continuously.
Google DeepMind's GNoME model predicted 2.2 million crystal structures, identifying 380,000 stable candidates—an order-of-magnitude expansion of known materials. Microsoft's AI-accelerated screening reportedly discovered a new lithium-alternative battery material. EvolutionaryScale's ESM3 protein model generated esmGFP, a fluorescent protein roughly 58% identical to the nearest natural GFP, demonstrating generation far outside known biology.
LabGenius, which uses its EVA robotic platform for antibody discovery, raised a £35 million Series B in May 2024 to scale its multispecific antibody pipeline. Profluent launched OpenCRISPR-1, the first AI-designed, open-source genome editor, in April 2024, publishing validation data in Nature in 2025 and securing $106 million in November 2025. Kebotix, partnering with the NIH's National Center for Advancing Translational Sciences, reported 5× cost reduction and 5× speed increase in high-throughput experimentation versus traditional factorial design-of-experiments.
A self-driving enzyme engineering platform delivered greater than 12°C thermostability improvements while exploring less than 2% of the design landscape, according to a Nature Chemical Engineering paper.
These aren't incremental improvements—the kind you celebrate with a modest press release. They represent a compression of research cycles that, if sustained, could fundamentally alter the economics of R&D-intensive industries. The pharmaceutical sector, notoriously slow and capital-intensive, is paying close attention.
Three Camps, One Race
The field divides roughly into three camps, though the boundaries blur. Large tech AI labs anchor the first group—Isomorphic and DeepMind, deploying foundation models like AlphaFold 3 (announced May 2024 with expanded biomolecular interactions) and leveraging Alphabet's compute resources.
Recursion, Insitro, Absci, and LabGenius represent the biotech cohort, integrating proprietary datasets with in-house or partnered automation.
Chemify, IBM's RoboRXN, Citrine Informatics, and Kebotix target chemistry and materials, often with closer ties to manufacturing.
Then there are the cloud lab infrastructure providers—Emerald Cloud Lab, which relocated and expanded to Austin; Strateos, pivoting to on-site deployments with its LodeStar OS. They're enabling remote access to instrumentation, a model the National Science Foundation endorsed in August 2025 with up to $100 million for a national network of AI-programmable cloud labs.
Synthetic Sciences enters as a software layer, offering agent orchestration across compute and literature without owning physical labs. They're betting that researchers want an operating system for autonomous science rather than vertically integrated facilities. Whether that bet pays off remains to be seen.
The Credibility Gap
Not everything works as advertised, of course.
Berkeley's A-Lab synthesis claims sparked debate over novelty and validation. Critics questioned whether the materials were genuinely new or rediscoveries, and whether synthesis alone constitutes meaningful innovation without property characterization. A 2024 BCG study found that only 26% of companies showed tangible AI value, and a 2025 follow-up noted just 5% derive material value at scale. The gap between pilot hype and production ROI remains wide, particularly in organizations lacking pristine data infrastructure or the cultural readiness to trust autonomous systems.
Replicability concerns plague the field. Nature coverage in December 2023 highlighted reproducibility challenges, and industry insiders acknowledge that wet-lab bottlenecks—throughput limits, multi-modal measurement constraints, instrument calibration drift—still throttle genuine "end-to-end" autonomy.
Publishing teams that share benchmarks, error analyses, and user studies (as detailed in Nature Communications performance metrics papers from February 2024) gain credibility. Those making broad claims without transparent validation face skepticism, and rightly so.
Regulatory Fog
The FDA released draft guidance in late 2024 on AI model credibility for drug and biologic submissions, focusing on risk-based assessments of AI-generated data in regulatory decisions. Notably, the guidance excludes "drug discovery" per legal analyses, concentrating instead on models supporting submission evidence. The EU AI Act, staged for implementation from February 2025 through August 2027, imposes conformity assessments and documentation requirements for high-risk AI, with research-use carve-outs still being clarified.
Biosecurity looms larger, though. The Biden administration's 2023 AI executive order emphasized nucleic acid screening and biosafety; it was rescinded in January 2025, replaced by a new biosafety order in May 2025 with tighter gain-of-function controls. As agent systems become more capable—able to design experiments and access synthesis protocols—concerns grow that bio or chemical misuse risks could escalate. Model developers face calls for evaluations and access gating, a tension between open science and responsible deployment that won't resolve easily.
What Remains Unsolved
Several technical frontiers remain stubbornly unsolved. Theory-in-the-loop discovery—integrating mechanistic models with empirical search—is one. Standardized experiment ontologies for cross-lab interoperability, another. Multi-agent reliability with formal verification, end-to-end data provenance for regulatory traceability—all appear in 2025 research roadmaps, none yet solved at scale.
OpenAI's FrontierScience benchmark, released in December 2025, tests expert-level scientific reasoning in physics, chemistry, and biology. It reveals progress, yes, but also significant limitations on open-ended problem-solving.
Industry leaders are converging on a playbook: scaled compute plus curated proprietary datasets plus tight lab integration. Daphne Koller of Insitro frames it as a paradigm requiring re-architected data and fast feedback loops. Recursion's leadership emphasizes "scaled data plus scaled compute" driving foundation models to cut wet-lab workload while maintaining throughput. Demis Hassabis points to Isomorphic's multi-billion-dollar pharma deals as proof that AI-first design can command enterprise budgets.
Adoption is spreading beyond early movers. The NSF's programmable cloud lab initiative signals federal backing for shared research infrastructure. More domain-specific foundation models—proteins, materials, small molecules—are expected via server-gated APIs with enterprise tiers, following the AlphaFold 3 model. Pharma mega-collaborations akin to Isomorphic's deals with Lilly and Novartis will likely multiply as platforms demonstrate validated drug candidates entering clinical trials.
The Real Question
The question isn't whether AI co-scientists will transform research—early results suggest they already are, in pockets at least. The question is how fast credible, replicable systems displace traditional workflows, and whether the organizations deploying them can navigate data governance, regulatory uncertainty, and the cultural shift of trusting machines to run experiments humans once controlled.
For those who solve it—and solve it with transparency, not just press releases—the payoff is measured in years saved and billions redirected from failed hypotheses to validated discoveries. The labs running 24/7 now aren't waiting for permission. They're compressing decades into months, and the rest of the industry is scrambling to keep up.
