Cloud Laboratories

A cloud laboratory is a physical laboratory with an application programming interface in front of it. The user writes a protocol as code, submits it, and a scheduler assigns instruments and robotic handling to execute it; samples arrive and leave by courier; the raw instrument output comes back as data. It is pursued in three distinguishable forms: commercial remote-execution platforms, university-owned cloud laboratories, and closed-loop self-driving laboratories where the protocol is chosen by a model rather than a person.

Established Three briefs on this map hold pieces of this subject and none holds the instrument layer. Artificial Scientists owns the discovery loop and its verification problem — who or what decides whether an output is true. Science of Science owns the production function and the replication record. Scientific Funding Models owns how the money is allocated. This brief owns the machine: what a remotely programmable instrument can and cannot do, what it costs to run, whether it delivers the reproducibility claimed for it, and who gets to use it.

Established The strongest claim made for cloud laboratories is not speed. It is reproducibility, and that claim has never been tested by the one experiment that would test it. The argument runs: a protocol expressed as code is deterministic, a machine executes it identically every time, therefore the experiment is reproducible. That is an argument from mechanism, not a measurement. Nobody has published a run of the same machine-executable protocol on two independent cloud laboratories with a blinded comparison of results. Until someone does, the field’s central selling point rests on an inference about scripts rather than evidence about experiments.

Frontier The one case where an automated laboratory’s output was independently re-examined suggests automation moved the error rather than removing it. An autonomous inorganic-synthesis laboratory reported dozens of new compounds; an independent group re-examined the reported products and concluded that no new materials had been discovered, naming unreliable automated crystallographic refinement as a root cause. Artificial Scientists uses that episode to argue about verification and cognition. This brief uses it for a different purpose: it is the only audit of an automated instrument stack in the literature, and what it found was a failure in the layer that converts raw instrument output into a claim — the layer cloud laboratories automate hardest and check least.

Established A note on sourcing. This brief was commissioned in September 2026 from the Institute’s research base. Reading-list entries without links are cited from the bibliographic record rather than re-fetched, and claims are dated no later than early 2026 unless carried by a linked source.

Frontier The commercial record of the pure-play cloud laboratory is poor, and the surviving instances are institutionally owned. The two best-known remote-execution platforms of the 2010s both ceased independent operation; the most prominent continuing cloud laboratory is a university facility built on one of those platforms’ technology. That pattern — the technology survives, the standalone business does not — is the most informative fact about the sector’s economics, and it is reported here from trade coverage and company communications rather than filings, so the dates should be checked before they are quoted.

Established The public sector has begun funding the missing layer directly, and the framing of that funding is the tell. A United States national funder has solicited a test bed for network-programmable cloud laboratories — a shared facility to work out interoperability, remote programmability and access rather than to run any particular science. A funder paying for a test bed is a funder stating that the interoperability problem is unsolved, which is more candid than the field usually is about itself.

Established The best-documented autonomous laboratory result and its refutation should be read as one item. The original report describes a closed-loop inorganic-synthesis system producing 36 compounds from 57 targets over 17 days. The independent re-analysis re-examined 43 reported products, concluded that no new materials were discovered in that work, and identified unreliable automated Rietveld refinement among the causes. The two accounts do not even share a denominator, which is itself a reproducibility observation: the number of things an autonomous laboratory made is not a settled quantity in the published record.

Established The founding closed-loop system claimed efficiency, not discovery, and the distinction has been lost in transmission. The 2004 robot-scientist paper reports experiment-selection efficiency competitive with human performance and cost decreases of threefold and a hundredfold on its own task, and makes no claim to have discovered new knowledge; Artificial Scientists establishes this in detail. What matters for the instrument layer is that the measured benefit in the founding case was cost per experiment — exactly the quantity a cloud laboratory sells and almost never publishes.

Established Language models can now drive real instruments, and the chemistry they have driven is not new chemistry. A 2023 system executed six demonstrations autonomously, including palladium-catalysed cross-couplings, planning and running them on a cloud-connected liquid-handling platform. A companion system augmented a model with expert-designed tools and executed syntheses of known compounds. Both are genuine instrument-control results, and in neither case was the target compound novel: the effector layer works while the question of what to point it at remains open.

Frontier Automation has been turned on the literature itself, which is the most under-rated use of the technology. A robot system was applied to testing the reproducibility and robustness of published cancer-biology results rather than to generating new ones. That is the application where the economics are clearest — replication is high-volume, protocol-defined and chronically unfunded — and the one with the least commercial pull, which is why it has one paper rather than a field.

Established Where reproduction has been benchmarked systematically, the numbers are low and consistent. A computational-reproducibility benchmark built from 270 tasks across 90 papers reports about 21% on its hardest tier. A benchmark requiring agents to replicate twenty machine-learning papers across 8,316 gradable sub-tasks reports 21.0% for the best system, with doctoral researchers ahead. Both measure computational work, with no pipetting, no reagent lot and no instrument drift. They are the optimistic bound on automated reproduction of wet-lab work, and the bound is about one in five.

Established The only randomised measurement of an artificial-intelligence tool on expert technical work found a slowdown that its users experienced as a speedup. Experienced developers working on their own repositories were measured 19% slower with the tool while believing themselves about 20% faster; Science of Science owns the result. It belongs here because every cloud-laboratory value proposition contains a claim about researcher time, and this is the only adjacent place where researcher time was measured rather than asserted.

Frontier Published throughput figures for self-driving laboratories are almost all acceleration factors against counterfactuals the authors constructed. “Ten times faster” is a ratio whose denominator is an estimate of how long a hypothetical human would have taken, produced by the same group reporting the numerator. There is no convention for constructing that denominator, no requirement to publish it, and no case in this literature where a matched human laboratory ran the same campaign. An acceleration factor with an unpublished denominator is not a measurement.

Established The time budget automation claims to attack is not mostly at the bench. A survey of 11,167 principal investigators at 111 institutions found 44.3% of federally funded research time going to administrative and compliance tasks, rising to 54.7% for those submitting twenty or more proposals, with proposal preparation alone at 16.0%. A cloud laboratory removes hands-on execution time. The largest measured non-research load in the working week is paperwork, and no instrument addresses it.

Established The access baseline against which democratisation would have to be measured is published, and the measurement has not been made. Federal science and engineering funding across 1,110 United States institutions rose from $31.2bn to $49.0bn between fiscal years 2014 and 2023, with the top 100 taking 78.4% in FY2023 — the concentration a cloud laboratory is supposed to relieve by decoupling capability from campus infrastructure. No study has reported who actually used commercial or academic cloud laboratories, by institution type or country.

Frontier Does automated execution actually raise reproducibility? The mechanism is plausible and the evidence is one audit that went the other way. Machine execution removes operator-to-operator variance in the steps the protocol specifies. It does nothing about unspecified steps, the reagent lot, calibration history, or the interpretation software that converts a spectrum into a claim — the layer where the one published audit found the failure. The net effect could be positive, negative or nil, and it is unmeasured.

Speculative Remote execution may select the science, not just serve it. If an experiment must be expressible in a platform’s instruction set to be runnable, that instruction set becomes a filter on what gets attempted. High-replicate parameter-sweep work fits; improvisational work, work requiring a judgement call mid-run, and work needing an instrument the platform lacks, does not. Whether this raises aggregate discovery yield by concentrating effort or lowers it by pruning the unusual is an open empirical question nobody has framed as one.

Frontier How much of laboratory skill is in the protocol and how much is in the hands is genuinely contested. The optimistic reading is that tacit knowledge is mostly under-documentation, and that forcing protocols into executable form surfaces it. The pessimistic reading is that much of it is perceptual and corrective — noticing that a precipitate looks wrong — and cannot be written down because the person doing it cannot articulate it. Both readings predict the same first outcome, an initial burst of failed transfers, and they differ only in whether the failure rate falls with effort.

Established The binding constraint is the absence of a vendor-neutral instrument-control standard with majority adoption. Candidate standards exist for device control, analytical data interchange and protocol description, alongside vendor-specific and open-source control libraries, and none has won. A protocol written for one platform is therefore not executable on another without re-implementation, which is why the cross-platform reproducibility test at the centre of this brief cannot currently be run cheaply.

Established Physical logistics sets a latency floor that no amount of software removes. Samples, reagents and materials move by courier under temperature and custody constraints. A remote experiment that would take an hour at the bench takes days door to door, and iteration speed — the thing a closed loop is supposed to buy — is bounded by shipping rather than by execution for any protocol whose inputs are not already on site.

Frontier Automated interpretation of instrument output is the least examined and most consequential layer. Converting a diffraction pattern, a spectrum or an image into an identification is a fitting problem with free parameters and failure modes that a human specialist recognises and a batch script does not. This is the named root cause in the one published audit of an autonomous laboratory. It belongs to crystallography, spectroscopy and image analysis rather than to robotics, and it has no institutional home in the cloud-laboratory field at all.

Frontier Reagent and consumable provenance is the classical reproducibility failure and the protocol record usually drops it. Lot-to-lot variation in biological reagents, resin batches, cell lines and solvent purity is a documented source of irreproducibility. A machine-executable protocol records what was called for, not necessarily what was in the bottle, and until lot identifiers are carried through the execution record an executed protocol is not a complete description of what happened.

Frontier Exception handling is where automation and bench practice diverge most sharply. A skilled experimenter who sees something unexpected improvises, repeats, or stops and thinks. An automated system either halts, or continues and produces a clean-looking dataset describing a run that went wrong. The second failure mode is the dangerous one, because it produces well-formatted data with no flag on it, and it is the mode least visible in published results.

Established This brief depends on a verification layer it does not own. Artificial Scientists establishes that autonomous discovery results survived scrutiny where a cheap exact verifier existed and failed where the verifier was a noisy physical instrument. A cloud laboratory is precisely a noisy physical instrument offered as a service, so every claim about its output inherits that finding whole. Nothing in the instrument layer fixes it.

Established It depends on an outcome measure nobody has validated. Science of Science records that the field lacks a validated measure of discovery yield that is not citation-based, and that the output instruments in use are under active methodological dispute. Any claim that a cloud laboratory raised discovery yield therefore has no agreed denominator, which is the same structural problem the acceleration factors have, one level up.

Established It depends on an allocation decision that is not a research result. Scientific Funding Models records that every alternative funding mechanism is running with real money and none has a published outcome comparison against what it replaced. Cloud-laboratory access is an allocation mechanism of exactly that kind — who gets instrument time, chosen how — and it will inherit the same evaluation vacuum unless somebody decides otherwise in advance.

Established The decisive experiment is the cross-platform replication, and nobody has run it. Take one non-trivial protocol, implement it on two independent cloud laboratories, run it blinded on both with identical inputs, and compare the results against each other and against a skilled bench laboratory running the same protocol from its written form. That would convert the field’s central claim from an argument about scripts into a measurement, in either direction: agreement establishes reproducibility-through-automation as a real property, and disagreement locates where the variance lives. It requires no new technology and no new instrument, and it has not been done.

Established The second experiment is the matched-pair cost measurement the sector sells on and never publishes. Run the same campaign remotely and at a competent bench laboratory, and report cost per completed experiment, wall-clock time from submission to usable data, and failure rate, on both. The founding robot-scientist paper reported cost decreases; no cloud-laboratory operator has published the equivalent. Until one does, every economic claim in this field is a quotation rather than a result.

Frontier The third is to institutionalise the audit that has been performed exactly once. The independent re-examination of an autonomous laboratory’s reported products is a method, not a one-off: sample a fraction of automated outputs, re-characterise them by hand, and publish the discrepancy rate. Run routinely, it would give the field a continuously measured error rate on its instrument stack, which is the thing no automated laboratory currently reports.

Frontier The fourth is a randomised trial of access, and the design already exists in an adjacent field. The one randomised measurement of an artificial-intelligence tool on expert work used a within-subject design on the subjects’ own tasks and found a slowdown against a believed speedup. The same design applied to cloud-laboratory access — randomising which of a researcher’s planned experiments go remote — would measure the thing the sector claims and would take one funding cycle.

Speculative The experiment that cannot be run on a useful timescale is the one about yield. Whether a decade of cheap remote experimentation produces more important science than a decade of the same money spent on people and benches is a question about outcomes that take a decade to score, using an outcome measure that does not yet exist. Nothing here is an argument for waiting; it is an argument for starting the measurement now rather than asserting the answer.

Established A remotely executable protocol needs four things a written protocol does not. A device abstraction naming capabilities rather than machines, so that “centrifuge at 3,000 g for 5 minutes” binds to whatever rotor is free; an explicit state model for samples and containers; error semantics saying what happens when a step fails; and provenance capture recording what was done rather than what was requested. Most published protocols supply none of the four, which is why converting one into an executable protocol is authoring, not translation.

Frontier Liquid handling is solved to a useful standard and sample handling is not. Pipetting, plate movement and standard analytical workflows automate well. Anything involving unstructured solids, heterogeneous material, live organisms outside standard formats, or an assembly step defined by feel remains manual or bespoke. The boundary of what a cloud laboratory can offer is roughly the boundary of what fits a microplate.

Established The unit of reproducibility is raw instrument output, not the processed result. If a platform returns a processed identification, a later re-analysis is impossible and the audit described above cannot be performed. Capturing and retaining raw output with its instrument and calibration metadata is cheap, is not universal, and is the single most valuable engineering requirement in this brief because it is what makes every other check possible afterwards.

Established The three commissioning neighbours are used here rather than repeated. Artificial Scientists owns the discovery loop, the verifier dichotomy and the A-Lab episode as a verification case. Scientific Funding Models owns funder design and the finding that no alternative mechanism has a published outcome comparison. Scientific Institutions Through History owns the record of how epistemic machinery accretes and is rationalised afterwards, which is the best available prior on how this technology will be described in retrospect.

Established Two further briefs supply the measurement frame. Science of Science owns the production function, the disputed output instruments and the randomised tool trial this brief leans on. Innovation History owns the instrument-validity rule that governs any long-run series used to argue that research is speeding up or slowing down.

Frontier Outside this map the adjacent disciplines are not the ones the field talks to. Crystallography, spectroscopy and quantitative image analysis own the interpretation layer where the one published audit found the failure; industrial process control owns exception handling; logistics owns the latency floor; and metrology owns the calibration-traceability practice a networked instrument fleet will need and does not have.

Frontier The ownership question decides the durability question. A standalone operator must earn its capital cost from external demand; a university facility can carry it as infrastructure, the way a synchrotron or a sequencing core is carried. The observed survivals so far are institutional, which suggests that the cloud laboratory is a shared-facility category rather than a market category — and shared facilities have a well-developed governance literature, allocation practice and cost-recovery model that the sector has largely not borrowed.

Established No journal requires deposit of a machine-executable protocol alongside a paper. Written methods are required and known to be incomplete; executable protocols exist for a growing share of published work and are collected nowhere systematically. It is the cheapest intervention in the subject: a deposit requirement would build the corpus on which cross-platform reproducibility could be tested at scale, and it needs an editorial decision rather than a technology.

Frontier Funding instruments are shaped for capital, and a cloud laboratory is an operating expense. Equipment grants buy instruments that sit in a named laboratory; remote execution is bought per run, as a subscription or a service line. Institutions account for the two differently, indirect-cost treatment differs, and a researcher with equipment money may find they cannot spend it on instrument time. Nobody has documented how often this actually blocks use, which is a gap that could be closed with a survey.

Frontier A remotely programmable wet laboratory is a dual-use surface, and the screening question is not solved by the existing model. Synthesis-order screening is built around sequences and compounds ordered from suppliers; a cloud laboratory takes a protocol, not an order, and a harmful outcome can be assembled from individually unremarkable steps. Safety evaluation for models operating on molecular, protein and genomic tasks exists and is early. The right unit of screening for an executable protocol has not been defined by anyone.

Frontier The democratisation claim is the field’s main public justification and its least examined one. Decoupling capability from campus infrastructure is a real mechanism, and per-run pricing genuinely lets a small group buy an experiment it could not buy a building for. Whether that changed who does science is measurable from operator records and has never been reported. Against a baseline in which the top 100 institutions take 78.4% of federal science funding, an unmeasured equity claim is marketing until somebody publishes the user distribution. Research technicians, who hold much of the tacit knowledge this brief argues is load-bearing, are the group displaced, and nobody counts them either.

Speculative If the model works, the experiment becomes a utility and the unit of scientific capacity stops being the building. A world in which any group can buy a characterised measurement at a published price is a different research economy: capability follows budget rather than affiliation, and the fixed cost of entering an empirical field collapses. That is the strong form of the case; the evidence for it is the mechanism plus the failure of the businesses that tried it.

Speculative If it does not work, the reason will be the tacit-knowledge floor, and that has a consequence for a much larger argument. Autonomous discovery systems need a physical arm. If remote execution turns out to be reliable only for protocol-shaped work, then the ceiling on machine science is set by the fraction of science that is protocol-shaped, which is a bound nobody has estimated. It would also mean the binding constraint on automated discovery is instrumentation rather than cognition — the conclusion Artificial Scientists reaches from the other direction.

Frontier Concentration is the risk the optimistic case creates. A small number of operators executing a large share of the world’s experiments is a chokepoint of a kind science has not previously had: a private scheduling decision, a pricing change or a business failure would propagate into the empirical record itself. The mitigations are ordinary — interoperability, protocol portability, more than one operator — and are the same requirements this brief types as constraints, which is a reason to treat them as public-interest infrastructure rather than product features.

These horizons track when the evidence about cloud laboratories changes, not when a facility opens. Each entry names something that could be shown to be wrong.

• 10 yr: Frontier Either the cross-platform replication is published and the reproducibility claim acquires a number, or it is not and the claim remains an argument from mechanism a decade after the technology matured. Expect published acceleration factors to keep carrying unpublished denominators unless a journal requires otherwise, and expect the surviving facilities to be institutional.
• 25 yr: Speculative The question decided in this window is whether protocol portability becomes normal, in which case remote execution becomes a routine option like sequencing, or whether it does not, in which case cloud laboratories remain a specialist service for high-replicate work. The determinant is a standards outcome, not a robotics one, and standards outcomes are decided by adoption rather than by merit.
• 50 yr: Speculative If the interpretation layer becomes reliable and auditable, the interesting consequence is not faster science but a continuously measured error rate on the empirical record — something no scientific institution has ever had. If it does not, automated facilities will have accumulated a large volume of well-formatted results whose error rate is unknown, which is a worse position than the one the field started in.
• 100 / 250+ yr: Handwave Claims at this range concern whether human hands remain part of experimental science at all. The candidate answers all depend on the unestimated fraction of science that is protocol-shaped, and any figure attached to this row is invented rather than derived.

• Depends on This brief depends on Artificial Scientists for the verification layer, and the dependency is load-bearing rather than decorative: that brief’s finding is that autonomous results survived audit where a cheap exact verifier existed and failed where the verifier was a noisy physical instrument, and a cloud laboratory is a noisy physical instrument sold as a service. It additionally inherits, without claiming edges, the missing non-citation measure of discovery yield from Science of Science and the none-compared finding of Scientific Funding Models.
• Requires (not on this map) Five constraints, in the order of the tokens. First, a cross-platform replication of one protocol on two independent cloud laboratories, blinded and compared: the field’s central claim is that machine execution makes an experiment reproducible, and that claim is currently an inference about scripts rather than a measurement about experiments. Second, a vendor-neutral instrument-control interface with majority adoption. Candidate standards exist for device control, analytical data interchange and protocol description and none has won, so a protocol written for one platform must be re-implemented for another — which is why the first constraint is expensive rather than routine. Third, a journal requirement to deposit machine-executable protocols with the paper: written methods are known to be incomplete, executable protocols exist for a growing share of published work, and nobody collects them, so the corpus needed for reproducibility testing at scale is discarded as it is generated. Fourth, lot-level reagent provenance carried through the execution record as first-class data: lot variation is a classical source of irreproducibility, and an executed protocol records what was called for rather than what was in the bottle. Fifth, and market-shaped: a repeat paying user base deep enough to keep expensive instruments utilised. Utilisation decides whether the model works, it is set by demand composition rather than engineering, no operator publishes it, and the standalone businesses that tried did not survive.
• Enables What this brief would enable if its first constraint were cleared is a measured error rate for automated experimentation, which is simultaneously the missing input to autonomous-discovery claims, to automated replication at scale, and to any economic comparison between remote and bench execution. No typed enabling edge is claimed, because every downstream use waits on a measurement this brief argues has never been made rather than on a result it could supply.
• Adjacent Within this map: Artificial Scientists, Science of Science, Scientific Funding Models, Scientific Institutions Through History and Innovation History. Outside it: crystallography and spectroscopy for the interpretation layer, industrial process control for exception handling, metrology for calibration traceability, and logistics for the latency floor.

Frontier “Automated execution makes experiments reproducible.” This is an argument from mechanism that has never been measured. Machine execution removes operator variance in specified steps and does nothing about unspecified steps, reagent lots, calibration history or interpretation software. The one published audit of an automated laboratory found the failure in the interpretation layer. The decisive test — the same protocol on two independent platforms — has not been published by anyone.

Established “An autonomous laboratory discovered dozens of new materials.” An independent re-examination of the reported products concluded that no new materials were discovered in that work. The original report gives 36 compounds from 57 targets over 17 days; the re-analysis re-examined 43 reported products and named unreliable automated crystallographic refinement among the causes. Citing the first without the second misreports the state of the record.

Frontier “Self-driving laboratories are an order of magnitude faster.” The denominator is usually unpublished. Acceleration factors are ratios against an estimate of how long a hypothetical human would have taken, produced by the group reporting the numerator, and no case in this literature has a matched human laboratory running the same campaign. The claim may be true; as published it is not checkable.

Established “Automating the bench gives researchers their time back.” The largest measured time sink is not the bench. Across 11,167 principal investigators at 111 institutions, 44.3% of federally funded research time went to administrative and compliance work, with proposal preparation at 16.0%. And the only randomised measurement of an artificial-intelligence tool on expert technical work found a 19% slowdown against a believed 20% speedup.

Frontier “Cloud laboratories democratise access to experimentation.” The mechanism is real and the outcome is unreported. Per-run pricing genuinely substitutes for a building. Whether it changed the distribution of who does empirical work is knowable from operator records and has never been published, against a baseline in which the top 100 United States institutions take 78.4% of federal science and engineering funding.

Speculative “A protocol written as code is a complete description of an experiment.” It is a complete description of the instructions. It does not carry the reagent lot, the instrument’s calibration history, the ambient conditions, or the operator judgement that the written protocol never contained either. Handwave The version of this claim that treats executability as equivalent to completeness is the step where the whole reproducibility argument works by assertion.

Frontier “The cloud laboratory sector is scaling.” The two best-known pure-play platforms ceased independent operation and the prominent survivor is a university facility. That is compatible with the technology being sound and the standalone business model being wrong, which is the reading this brief takes. It is not compatible with describing the sector as a growth market, and the reporting here comes from trade coverage and company communications rather than filings.