Cortical Labs

Cortical Labs is a Melbourne biotechnology company that develops experimental biological-computing systems. Its platform joins flat, lab-grown networks of living neurons to a silicon multielectrode array and conventional software. Electrical stimulation supplies input, recorded neural activity supplies output, and a feedback loop changes the neurons' simulated environment. The result is a hybrid research instrument, not a cultured human brain or a stand-alone artificial intelligence model.^[1][2][4][7]

The company was founded in 2019 by Hon Weng Chong. Its best-known research system is DishBrain, used in a peer-reviewed 2022 study of a simplified Pong task. Cortical Labs introduced the self-contained CL1 device publicly in March 2025 and later opened remote access through Cortical Cloud. By Aug. 22, 2026, published evidence supported goal-directed adaptation in the constrained DishBrain experiment, but no independent common-workload benchmark had validated broad performance or energy claims for CL1, the company's Doom demonstration, Cortical Cloud, or its multi-unit prototype racks.^{[2][4][7][9][14][17]}

Key facts

Scroll sideways for more →

History

Chong started Cortical Labs in Melbourne in 2019. The current company page lists him as chief executive, Brett Kagan as chief scientific and operating officer, David Hogan as chief technology officer, and Andrew Doherty as chief hardware officer. Independent reporting traced the project to Chong's interest in whether living neurons could become a useful computing substrate.^[1][2]

In April 2023, the company raised a reported US$10 million round led by Horizons Ventures, with participation from LifeX Ventures, Blackbird Ventures, Radar Ventures, and In-Q-Tel. That financing followed the publication of the DishBrain work and supported further hardware development. It does not by itself establish current revenue, valuation, customer numbers, or product performance.^[3]

ABC reported that Cortical Labs launched CL1 at an international technology conference in Barcelona on March 5, 2025. The company presented it for neuroscience, drug-testing, and biological-computing research and offered a path to remote use. The launch marked a shift from a laboratory prototype to a packaged platform, but did not convert the earlier DishBrain result into a general benchmark for every CL1 culture.^[7][9]

DishBrain research

Closed-loop Pong experiment

The 2022 Neuron paper connected human induced-pluripotent-stem-cell-derived neurons or embryonic mouse cortical neurons to a high-density multielectrode array. Stimulation patterns encoded the position of a ball in a simplified Pong-like environment, while activity recorded from designated regions moved a virtual paddle. A miss produced unpredictable stimulation; a hit produced a structured signal. The study compared this closed loop with rest, no-feedback, non-neural, and in-silico controls.^[4][5]

The authors reported apparent learning within five minutes, increasing performance over time, and no comparable learning in the control conditions. They called the system "synthetic biological intelligence" and used "sentience" in the narrow sense of responsiveness to sensory impressions. Those terms were an operational framing of the experiment, not evidence that the cultures were conscious, had subjective experiences, understood Pong, or possessed general intelligence.^[4][5][22]

Performance was modest. ABC reported that DishBrain hit only slightly more balls than it missed, although it performed better than a stimulated system without feedback. An outside stem-cell researcher told ABC that simple Pong learning did not establish complex decision-making. The peer-reviewed result is therefore evidence of task-specific, closed-loop adaptation in a controlled setup, not ordinary video-game skill or a general machine learning capability.^[7]

DishBrain also did not have one universal neuron count. The paper's methods describe about 800,000 mouse cortical cells per array, about one million cells for one human preparation, and 100,000 induced neurons plus 25,000 astrocytes in another preparation. Its electrophysiology, gameplay data, and Python and Matlab analysis code were deposited publicly on OSF.^[5][6]

Sample-efficiency preprint

A 2024 preprint from Cortical-affiliated researchers compared DishBrain cultures with DQN, A2C, and PPO reinforcement learning agents. With training limited to about 70 episodes, corresponding to a 20-minute culture session, the authors reported better sample-limited Pong metrics for the biological cultures. The paper also showed that the digital agents improved when allowed much longer training and acknowledged the difficulty of comparing biological and artificial systems.^[8]

That study remained an arXiv preprint at the research cutoff. It measured samples and task metrics within one simplified game; it did not compare energy, lifecycle cost, accuracy on other tasks, CL1 hardware, cloud reliability, or rack performance. It cannot support a broad claim that biological cultures outperform current digital AI.^[8]

CL1 platform

Hardware and cell lifecycle

CL1 packages a living neural network, electrode interface, software, and life support in one device. Neurons grow as a two-dimensional culture over the silicon chip in nutrient solution. The hardware stimulates selected channels, records electrical spikes, runs applications, and manages temperature, nutrients, and waste. This is related to neuromorphic computing, but it uses biological cells rather than an electronic circuit that merely imitates them.^[7][9][18]

Cortical Labs says its environment can keep a culture alive for up to six months. "Up to" describes the platform's intended maximum support period, not an independently audited mean lifetime or a guarantee for each cell batch. Cultures can differ, age, become contaminated, or need replacement, while pumps, tubing, electronics, media, and human cell-culture work remain part of the operating system.^[9][17]

Neuron counts likewise vary by source and culture. IEEE Spectrum reported 800,000 neurons per CL1 in 2025, whereas Cortical's 2026 Doom demonstration and Melbourne deployment reporting described about 200,000 neurons in a unit. The current product page publishes no fixed standard count. These figures should be read as dated descriptions of particular configurations, not as interchangeable specifications for DishBrain, every CL1, or a rack.^{[9][16][17][18]}

Software and licensing

The Python CL API controls stimulation, recording, and timed feedback. Official documentation says it is preinstalled on CL1 and limits individual stimulation on a physical channel to 200 Hz to protect the cells. A company-authored February 2026 preprint describes sub-millisecond closed-loop operation and deterministic timing contracts. It documents the interface design, but is not an independent hardware-performance study.^[10][11]

Cortical also publishes a local CL SDK Simulator. The simulator generates random data or replays a recording and does not learn in response to stimulation; the documentation says it should not be used as experimental evidence. The public repository is licensed under Creative Commons Attribution-NonCommercial 4.0. That license applies to the repository material and does not establish that CL1 hardware, the biological operating system, Cortical Cloud, or the entire on-device stack is open source.^[10][12][13]

Physical and cloud access

The company maintains a purchase route for physical CL1 devices and says buyers need appropriate laboratory capability. Cortical Cloud provides remote browser, Jupyter notebook, and Python-SDK access to real CL1 cultures without a local device. The official page described the service as available by the cutoff, distinguishing it from the non-learning simulator.^[14][15][18]

Public product pages do not list a stable current physical-device price, cloud price, service-level agreement, independently measured uptime, or guaranteed number and health of cultures. Launch-era prices reported elsewhere should not be assumed to remain current. Cloud access provides a path to experimentation, not a guarantee that every workload will learn or outperform a digital system.^[10][14][15]

Demonstrations and deployments

Doom demonstration

In early 2026, Cortical Labs released a video of a CL1 culture interacting with Doom. Game information was encoded into electrode stimulation and neural activity was mapped back to actions. The video described about 200,000 human neurons, and company representatives characterized the play as beginner-like rather than expert. The ordinary game engine and interface software remained conventional digital components; the neurons formed one adaptive component of the closed loop.^[16][17]

The Doom work was a company demonstration arising from Cortical Cloud early access, not a peer-reviewed replication of the DishBrain protocol. By the cutoff, Cortical had not published a blinded repeated-trial study, standard score, matched digital baseline, or independently reproduced CL1 result for Doom. The video therefore shows an implemented interface and observed behavior, not proof that CL1 solved Doom or achieved general-purpose intelligence.^[16][17]

Melbourne prototype

Information Age reported that Cortical's Melbourne prototype contained 120 CL1 devices and supported Cortical Cloud. The same article attributed roughly 200,000 cells per device, 30 watts per CL1, five-to-six-month tubing replacement, and a 500-day culture result to company representatives. These are useful descriptions of that operation, but they are not independent measurements or fixed specifications for all units.^[17]

Bloomberg described the Melbourne installation and the then-planned Singapore installation as small experimental data centres, and reported that the approach remained years or decades from challenging mainstream technology. A rack of CL1 devices is a collection of hybrid biological instruments with conventional infrastructure, not a replacement equivalent to a rack of GPUs merely because both are installed in a data-centre setting.^[19]

Singapore research prototype

NUS Medicine announced the Singapore collaboration with DayOne and Cortical Labs in March 2026. The planned initial validation phase was one rack of 20 Cortical Cloud units at the NUS Life Sciences Institute, with a later transition envisioned for a DayOne commercial facility. The larger Singapore data-centre concept was a future plan, not a completed deployment at that stage.^[20]

An Aug. 17 NUS update documented a completed Aug. 6 showcase and a 20-unit CL1 system in a live NUS research environment. NUS said the cells would be cultured under Professor Rickie Patani and the infrastructure was designed and supported by DayOne. It called the rack the first independently operated biologically integrated server rack. The release documented live neural activity and system integration, not a 1,000-unit installation or independent proof of commercial-scale computing.^[21]

The NUS release did not publish a workload, throughput, accuracy, uptime, cost, culture-survival distribution, or work-per-watt result. Its language about future efficiency and scale came from project partners. The completed evidence as of Aug. 22 was the 20-unit research prototype; the proposed large facility remained an ambition.^[20][21]

Performance and energy limits

Cortical's pages say CL1 needs a fraction of the energy and training data of conventional systems. Reported power figures have different scopes: ABC relayed a company statement of a few watts, Information Age reported 30 watts per unit, and IEEE Spectrum reported 850 to 1,000 watts for a rack. None defines a common workload, output quality, cell-culture overhead, rack boundary, or equivalent digital comparator.^{[7][9][17][18]}

A defensible comparison would include the silicon electronics, pumps, temperature control, nutrient and waste handling, culture preparation and replacement, cloud or rack overhead, time to train, and quality of the result. No independent benchmark covering those factors was located. The Pong paper and sample-efficiency preprint do not fill that gap because they tested constrained learning behavior, not complete-system energy or cost.^[4][8]

Ethics and governance

The DishBrain paper says its experiments followed the Australian National Statement on Ethical Conduct in Human Research and the Australian Code for the Care and Use of Animals. Mouse work used approval E/1876/2019/M, and human-cell work used material-transfer agreements. Those approvals applied to that study; they are not a blanket ethical certification for every commercial culture, customer, or application.^[5]

ABC reported that Cortical's current human neurons were derived from induced pluripotent stem cells generated from volunteer blood samples. IEEE Spectrum reported that customers seeking to generate cell lines must secure ethical approval and that physical-unit buyers require suitable cell-culture facilities. This system is not a clinical brain-computer interface, but it still raises questions about donor consent, stewardship, culture disposition, biosafety, customer use, and any future evidence of morally relevant experience.^[7][18][22]

A 2023 peer-reviewed ethics analysis concluded that DishBrain did not provide clear evidence of artificial suffering, while arguing for precaution as systems become more capable. ABC's external researchers similarly described current cultures as too primitive to feel or understand, but urged continued evaluation. Cortical says it works within ethical frameworks; the public pages reviewed did not provide one comprehensive protocol covering donor governance, welfare thresholds, customer oversight, incident reporting, and end-of-life handling.^[1][7][22]

Claims that CL1 is conscious, sentient in the everyday phenomenal sense, a miniature brain, or guaranteed incapable of future moral relevance exceed the available evidence. The more limited conclusion is that living neural cultures can show adaptive electrical behavior in particular closed-loop tasks, while their technical capability and moral status require separate evidence and continuing review.^[4][7][22]

References

1. ^Cortical Labs. "Company." Accessed Aug. 22, 2026. corticallabs.com/company
2. ^Bill Goodwin. "How lab grown neurons could power the future of AI." Computer Weekly, June 6, 2023. computerweekly.com/...could-power-the-future-of-AI
3. ^Mike Butcher. "Cortical Labs raises $10M for its Pong-playing stem cells that eventually could power AI." TechCrunch, April 19, 2023. techcrunch.com/...-which-eventually-could-power-ai
4. ^Brett J. Kagan et al. "In vitro neurons learn and exhibit sentience when embodied in a simulated game-world." Neuron 110, 2022. discovery.ucl.ac.uk/...10158064
5. ^Brett J. Kagan et al. "In vitro neurons learn and exhibit sentience when embodied in a simulated game-world," open-access published version. discovery.ucl.ac.uk/...-S0896627322008066-main.pdf
6. ^Cortical Labs. "In vitro neurons learn and exhibit sentience when embodied in a simulated game-world," data and code. Open Science Framework. osf.io/5u6qv
7. ^Jacinta Bowler. "Melbourne start-up launches 'biological computer' made of human brain cells." ABC News, March 5, 2025. abc.net.au/...104996484
8. ^Moein Khajehnejad et al. "Biological Neurons Compete with Deep Reinforcement Learning in Sample Efficiency in a Simulated Gameworld." arXiv, May 27, 2024. arxiv.org/...2405.16946
9. ^Cortical Labs. "CL1." Accessed Aug. 22, 2026. corticallabs.com/cl1
10. ^Cortical Labs. "Developer Guide." Accessed Aug. 22, 2026. docs.corticallabs.com
11. ^David Hogan et al. "CL API: Real-Time Closed-Loop Interactions with Biological Neural Networks." arXiv, Feb. 12, 2026. arxiv.org/...2602.11632
12. ^Cortical Labs. "CL SDK Simulator." GitHub. Accessed Aug. 22, 2026. github.com/...cl-sdk
13. ^Cortical Labs. "CL SDK Simulator license." GitHub. Accessed Aug. 22, 2026. github.com/...LICENSE
14. ^Cortical Labs. "Cortical Cloud." Accessed Aug. 22, 2026. corticallabs.com/cloud
15. ^Cortical Labs. "Purchase." Accessed Aug. 22, 2026. corticallabs.com/purchase
16. ^Cortical Labs. "Living Human Brain Cells Play DOOM on a CL1." YouTube, 2026. youtube.com/watch
17. ^Tom Williams. "This Melbourne data centre runs on human brain cells." Information Age, March 11, 2026. ia.acs.org.au/...-centre-runs-on-human-brain-cells
18. ^Shannon Cuthrell. "Human Brain Cells on a Chip for Sale." IEEE Spectrum, May 29, 2025. spectrum.ieee.org/biological-computer-for-sale
19. ^Olivia Poh. "Human Brain Cells Run New Data Centers in Singapore, Melbourne." Bloomberg Law, March 9, 2026. news.bloomberglaw.com/...rs-in-singapore-melbourne
20. ^NUS Medicine. "Biological Data Centre prototype established at NUS Medicine." March 11, 2026. medicine.nus.edu.sg/...established-at-nus-medicine
21. ^NUS Medicine. "NUS Medicine, DayOne and Cortical Labs Unveil Biological Data Center Prototype in Singapore." Aug. 17, 2026. medicine.nus.edu.sg/...nter-prototype-in-singapore
22. ^Stephen R. Milford, David Shaw, and Georg Starke. "Playing Brains: The Ethical Challenges Posed by Silicon Sentience and Hybrid Intelligence in DishBrain." Science and Engineering Ethics 29, 2023. link.springer.com/...s11948-023-00457-x