Research analysis · Platforms

The wetware cloud, on the vendor's own terms

Cortical Labs now offers two ways to run code against living human neurons: buy the CL1 device, or sign up for the Cortical Cloud from a browser. The company's public pages are the closest thing the field has to primary documentation of commercial wetware access, and they reward a careful read. The closed-loop mechanism underneath is real and published; the capability language wrapped around it runs well ahead of the evidence the vendor itself links.

Source: Cortical Labs public platform documentation (home, CL1, Cloud, and Research pages), Cortical Labs Pte Ltd, accessed 2026-08-07. Primary source. Read all four pages in full as served on the access date; the papers the vendor links were checked as citations and abstracts, not re-read in full except where stated.

What the work claims

This is vendor documentation, not a study, and it should be weighed as such: it is the authoritative record of what a commercial operator says its platform does and who may use it, and it is evidence of nothing else. The claims are unusually concrete for marketing copy. The CL1 is described as "the world's first code deployable biological computer," a self-contained unit in which lab-grown neurons on a silicon chip are kept alive "for up to 6 months" with no external compute required.1 The Cortical Cloud extends this to remote users: "Deploy code to real neurons, no lab or device required," from a browser, through Jupyter notebooks and a Python SDK, onto "an array of CL1's."1

Around that core sit stronger assertions: the platform's neural systems "require minimal energy and training data to master complex tasks," exhibit "generalization and learning efficiency" that artificial intelligence "can only simulate," and will unlock "breakthroughs across industries from customer service to security protocols and beyond."1 A site banner links a Bloomberg report, whose URL headline describes human brain cells running new data centers in Singapore and Melbourne, and a video of the game Doom running on a CL1.3 Separately, the Research page indexes roughly two dozen papers, preprints, and commentaries co-authored by or associated with the company, including the 2022 Neuron paper in which in vitro neurons learned a Pong-like game.2 The peer-reviewed corpus is the capability floor. The pages themselves are the claim ceiling. The distance between them is the subject of this analysis.

How it works

The mechanism that makes any of this more than metaphor is the closed loop. Neurons cultured on a multielectrode array are embedded in what the company calls the Biological Intelligence Operating System (biOS), which "runs a simulated world and sends information directly to the neurons about their environment. As the neurons react, their impulses affect their simulated world."1 Concretely: the state of a virtual environment is encoded as patterned electrical stimulation across the array; the culture's evoked activity is recorded, decoded as an action; the action updates the environment; the loop closes, at millisecond latency. "Deploying code" to neurons means writing software that defines this environment and its feedback contingencies through the vendor's API.

The published anchor for the loop is the 2022 Neuron study, in which cultures of human iPSC-derived and mouse cortical neurons, embodied in a Pong-like game through exactly this kind of stimulation-and-recording cycle, showed within-session performance improvements when closed-loop feedback was delivered, relative to control conditions without meaningful feedback.2 A later vendor-linked comparison study measured sample efficiency of such cultures against deep reinforcement learning agents in the same game-world.6 The CL1 packages the loop into a desktop unit with integrated life support; the Cloud stacks remote scheduling, monitoring, and multi-tenancy on racks of those units. One sentence on the Cloud page deserves more attention than it invites: "We continuously monitor neural health and performance, ensuring optimal conditions and continuous access to an always-on network of living neurons."1 Health monitoring of living tissue is here an uptime feature, run by the seller.

Where a skeptic should push

The load-bearing assumption in the vendor's framing is that the demonstrated mechanism licenses the general capability language. It does not, and the gap is measurable on the vendor's own Research page. Everything linked there that constitutes a primary capability result operates in the Pong-class regime: embodied game-worlds, minutes-to-hours timescales, performance measured statistically against controls. Nothing linked demonstrates "mastering complex tasks," cross-task generalization, or any workload resembling customer service or security. Doom running on a CL1 is a demo whose division of labor between tissue and conventional silicon is not documented on the pages read.

The efficiency claims deserve the same treatment. "Minimal energy" is asserted without any linked like-for-like benchmark against a silicon system doing the same task at the same accuracy; the training-data claim traces to game-world sample-efficiency comparisons, which are real but narrow.6 "Up to 6 months" is a life-support ceiling, not a demonstrated duration of useful learning. And the vendor's strongest published claim is terminologically unstable by its own admission: the Neuron paper's title says the cultures "exhibit sentience," language that drew debate, and the same group co-authored both a call for nomenclature consensus on intelligent systems and an ethics paper asking whether embodied cultures provide evidence for organoid ethics.245 When the flagship result's own vocabulary is under revision by its authors, the marketing built on it inherits the uncertainty. Finally, these are undated, editable web pages: this analysis is a snapshot, checked against an earlier library capture from June 2026, and claims quoted here may silently change.

Who governs the vendor that governs the tissue

For platform access, the Cloud is the significant object, not the CL1. A device purchase carries incidental gates: capital, bench space, aseptic technique, an institution around the buyer. The Cloud removes them. What remains between an anonymous developer and experimentation on living human neural tissue is a signup flow and the vendor's terms of service. None of the existing oversight regimes attaches: animal research law governs no animal here, human-subjects regulation governs no person, and institutional review attaches to institutions the cloud user is not required to have. Browser access to cultured neurons is not illegal; it is unclassified. The last incidental gate, owning the instrument, was removed deliberately, because removing it is the product.

There is also a jurisdictional relocation. Per the vendor-linked Bloomberg headline, the tissue racks sit in Singapore and Melbourne; the users can be anywhere.3 The applicable law is wherever the neurons are, while the experimental intent forms wherever the developer is. That is a familiar structure from data governance, now applied to living tissue, and no tissue-specific instrument exists to meet it.

For vendor capability, the honest reading is that Cortical Labs has vertically integrated not just the stack but the governance around the stack. The same entity makes the capability claims, allocates access, operates the only welfare-relevant telemetry in existence for this tissue, and authors a substantial share of the ethics literature about it, while its pages cite "our established ethical frameworks" without linking a framework document.15 The non-obvious implication sits in the health-monitoring sentence: the continuous measurements a future welfare standard for cultured neural tissue would need are already being collected, as private service telemetry, by a party with commercial exposure to what the numbers say. The opportunity is genuine on both fronts: this is real democratization of an instrument class that was recently confined to a handful of labs, and a vendor that publishes in Neuron and hosts its critics' vocabulary debates is auditable in a way a secretive one is not. The threat is equally specific: terms of service become the de facto access-governance instrument for living neural tissue, a private allowlist standing in for review, while the marketing operates in both calibration directions at once, inflating capability where it sells ("sentience," intuition, industry breakthroughs) and deflating moral concern where it reassures ("animal-free," "ethically superior"). Both cannot be right about the same tissue.

The bottom line

Established: a commercial closed-loop wetware platform exists, its core mechanism is peer-reviewed, and remote browser access to it is being sold now. Hypothesis, at best: the energy and data-efficiency advantages at real workloads, generalization, and every industry application named. The claims to watch are not the capability ones, which independent benchmarks on non-game tasks would settle, but the governance ones: whether the Cloud's user-vetting and terms are ever published, whether neural-health telemetry becomes an auditable standard or stays proprietary, and whether any regulator notices that the jurisdiction of record for experiments on human neural cultures is now a data-center address. A vendor page is a weak source for what a platform can do, and a strong source for what its operator intends. Read this one as intent.

Frequently asked questions

What is the Cortical Labs CL1?

A commercial desktop unit that grows neurons on a silicon multielectrode chip, keeps them alive with onboard life support for up to six months by the vendor's specification, and exposes them to software through a closed-loop stimulation and recording system the company calls biOS.

What is the Cortical Cloud?

A remote-access service that lets users run code against racks of CL1 units from a browser, using Jupyter notebooks and a Python SDK, with no laboratory or device of their own. The vendor operates the tissue and monitors its health as part of the service.

Did neurons really learn to play Pong?

A 2022 peer-reviewed study in Neuron reported that cultures of human and mouse cortical neurons, embodied in a Pong-like game through closed-loop electrical stimulation and recording, improved their performance within a session when given feedback, relative to control conditions. The improvement is statistical and task-specific; the paper's use of the word sentience remains debated.

What law governs who may run code on cultured human neurons?

No regime squarely applies. Animal research law involves no animal, human-subjects regulation involves no enrolled person, and institutional review boards attach to institutions that a cloud user is not required to have. Access is currently governed by the vendor's own terms of service.

Is biological computing actually more energy-efficient than silicon AI?

It is claimed, and it is plausible for specific regimes, but the vendor's linked research contains no like-for-like benchmark of energy per task at matched accuracy. The published comparisons concern sample efficiency in a single game-world, which is a much narrower claim.

References

  1. Cortical Labs Pte Ltd. Platform documentation: home, CL1, Cloud, and Research pages. corticallabs.com. 2026. https://corticallabs.com/. Accessed 2026-08-07.
  2. Kagan BJ, Kitchen AC, Tran NT, et al. In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron. 2022. doi:10.1016/j.neuron.2022.09.001. Accessed 2026-08-07.
  3. Bloomberg News. Human brain cells run new data centers in Singapore, Melbourne (headline as slugged in the vendor-linked URL; article paywalled and not read). Bloomberg. 2026. bloomberg.com. Accessed 2026-08-07.
  4. Kagan BJ, et al. Toward a nomenclature consensus for diverse intelligent systems: Call for collaboration. The Innovation. 2024. doi:10.1016/j.xinn.2024.100658. Accessed 2026-08-07.
  5. Kagan BJ, et al. Neurons embodied in a virtual world: evidence for organoid ethics? AJOB Neuroscience. 2022. doi:10.1080/21507740.2022.2048731. Accessed 2026-08-07.
  6. Khajehnejad M, et al. Biological neurons compete with deep reinforcement learning in sample efficiency in a simulated gameworld. arXiv. 2024. arXiv:2405.16946. Accessed 2026-08-07.