All postsScience explained

Wetware as a service: is neuron computing efficient?

APJ-RAISE team · · 7 min read

Wetware as a service (WaaS) is cloud access to living neurons grown on a chip: you send signals from code on your laptop, real brain cells in a lab respond, and software reads their reply.

A researcher in a lab coat looks into a microscope next to coloured petri dishes
Photo: Edward Jenner, Pexels

The selling point is energy. Brains run on very little power, so the hope is that computing with neurons will too. This post asks whether that is true today, and how you would even measure it. The short answer is that nobody has done the fair test yet, and the numbers people quote measure the wrong thing.

The topic went mainstream on 20 October 2026, when Gartner placed wetware as a service at number 10 in its top strategic technology trends for 2027. Gartner describes it as an emerging model for remote, cloud access to living or synthetic biological systems for computation.

What is a wetware computer?

Wetware is an old word for the living brain, set against hardware and software. A wetware computer grows neurons on a multi-electrode array (MEA), a chip with many tiny electrodes. Each electrode can send a small pulse into nearby cells (input) or pick up the spikes they fire (output). Software turns your data into pulses and turns the spikes back into an answer.

Hon Weng Chong, CEO of the Australian firm Cortical Labs, made the phrase Wetware-as-a-Service popular when he launched the CL1 at MWC Barcelona in March 2025. We could not confirm who used it first.

Two related names show up in papers. Biocomputing means computing with living cells. Organoid intelligence, named in a 2023 paper by Lena Smirnova and colleagues, uses organoids, which are tiny 3D clumps of brain tissue grown from stem cells.

What can neurons on a chip actually do?

So far, simple tasks. Two experiments set the bar.

DishBrain (Brett Kagan and colleagues, Neuron, October 2022) connected human or mouse neurons to the game Pong. Eight electrodes told the cells where the ball was, and spikes in two motor areas moved the paddle. A hit earned a short, predictable pulse (100 Hz for 100 milliseconds). A miss earned 4 seconds of random noise. The team reported apparent learning within five minutes of play, which did not appear in controls.

Brainoware (Feng Guo's team at Indiana University, Nature Electronics, December 2023) used a brain organoid as a reservoir. In reservoir computing, a fixed, messy network turns input into rich patterns, and only a simple output layer is trained. Fed 240 audio clips of 8 Japanese speakers, the system named the speaker from one vowel 78% of the time after about two days of training. It also predicted the Hénon map, a known chaotic system.

An ordinary computer solves both tasks with ease. These were proofs that tissue can compute, not speed records.

Is biological computing more energy efficient?

It is a claim, not a measured result. Three numbers get quoted, and none of them is energy per task:

  • The brain uses about 20 watts. That is a textbook fact about brains, not about any product.
  • Cortical Labs says a CL1 runs on about 30 watts, against up to 6,000 watts for a GPU, as Capital Brief reported. That is the power of the box.
  • FinalSpark's 2024 press release claimed a million times less power than digital processors. Its own peer-reviewed paper does not measure this.

Let us work out what the CL1 figure means. This is our arithmetic on the company's numbers:

  1. 1A CL1 at 30 W running for a week: 30 × 24 × 7 = 5,040 watt-hours, or about 5 kWh.
  2. 2A GPU at 6,000 W for the same week: 6,000 × 168 = 1,008 kWh.
  3. 3Ratio: 6,000 ÷ 30 = 200. So the CL1 draws about 200 times less power.

That ratio sounds huge until you ask what each machine finishes in that week. A GPU can run millions of speaker checks in the time a culture learns one game. If the GPU does more than 200 times the useful work, it wins on energy per task. Brett Kagan, Cortical Labs' chief scientist, put it honestly to Scientific American: a pocket calculator beats him at long division.

How would you measure wetware efficiency fairly?

A fair test compares energy per correct answer on the same task. We would run it like this:

  1. 1Pick one task both systems can do, such as the Japanese vowels speaker task from Brainoware.
  2. 2Fix a target, for example 78% accuracy, and stop each system when it gets there.
  3. 3Measure total energy for training and for answering, in watt-hours, with a meter at the wall.
  4. 4Count everything the system needs. For neurons that means the incubator, pumps and life support, not only the chip. For the GPU, include the host computer.
  5. 5Divide energy by the number of correct answers. Report both training energy and per-answer energy.
  6. 6Repeat on several cultures, because results vary from dish to dish.

Step 4 is where most claims fall apart. A 30 W figure may or may not include life support; the reports we read do not say. A digital baseline also does not need to be a big GPU. A small echo state network on a laptop CPU is the fairer rival for a vowel task.

What most wetware coverage gets wrong

Headlines mix power with efficiency. Power is how fast a device uses energy. Efficiency is how much useful work you get per unit of energy. Only the second matters, and it has not been published for any wetware product.

Learning credit is also blurry. Some of the learning in these systems may sit in the software around the cells, not in the cells. A good test would freeze the software decoder and show the culture still improves.

Finally, most results come from the companies themselves, and there are no standard benchmarks yet.

Cortical Labs CL1 price and other platforms

Two companies sell access as of November 2026.

  • Cortical Labs CL1: human neurons on a silicon chip with built-in life support, kept alive for up to 6 months, says the company. Reports say a unit holds about 800,000 neurons. No price on its website; TechRadar reported US$35,000 to buy or US$300 a week to rent remotely in June 2025.
  • Cortical Cloud: remote CL1 access with a Python API. A proof-of-concept data centre opened in Melbourne in March 2026, with about 120 units by the company's count, per Euronews.
  • FinalSpark Neuroplatform: 24/7 remote access to brain organoids through Python. Its 2024 paper reports over 1,000 organoids used in three years and 18 TB of data. Pricing is on request.

In February 2026, an independent developer used the Cortical Labs API to teach about 200,000 neurons to play Doom in about a week; reports say better than random but far below good players, and not peer-reviewed. In August 2026, DayOne, Cortical Labs and NUS Medicine reportedly unveiled a 20-unit prototype in Singapore.

At US$300 a week, a rented CL1 costs about US$43 a day. That is our division, and it is within reach of a funded student project, though not of most students alone.

Is wetware ethical?

The open questions are consciousness, consent and oversight. The DishBrain title used the word sentience, which drew criticism; Kagan now says the cells are nowhere near as complex as an insect. Alysson Muotri and others argue that more complex organoids may need new rules. A 2024 paper by Kataoka and colleagues says donors should consent to the specific project, and the 2023 Baltimore Declaration asks for ethics work alongside the science. Harvard bioethicist Insoo Hyun argues the science does not yet support the consciousness worry, and that consent matters more today.

In India, BioSpectrum India reports organoid work for medicine at IISc Bengaluru's Centre for Neuroscience and at CSIR-CCMB in Hyderabad. We found no Indian lab publishing on computing with neurons on chips as of November 2026.

Research you can do without a wet lab

  • DishBrain data: the paper says its spike and gameplay data are on the Open Science Framework. Try to reproduce one figure.
  • Tampere University MEA data: 2 TB of recordings from human stem cell neuron networks, good for practising spike detection.
  • A digital rival: the reservoirpy Python library includes the Japanese vowels task and the Hénon map. Build an echo state network, log its energy with a power meter, and compare with Brainoware, noting the setups differ.

Our course on reading and reviewing papers helps you read DishBrain and Brainoware critically. For a different way to put the physical world into AI, see physics-informed neural networks, and for another 2027 trend, world models. To find a team, join APJ-RAISE.

Questions people ask

What is wetware as a service?
It is remote, cloud access to living neurons grown on a chip, used for computing. Cortical Labs (Cortical Cloud) and FinalSpark (Neuroplatform) offer it as of November 2026.
What is the Cortical Labs CL1 price?
Cortical Labs lists no price. In June 2025, TechRadar reported US$35,000 to buy a CL1, or US$300 a week to rent one remotely.
Is wetware AI real?
Yes, as early research. Neurons and organoids have done simple tasks such as Pong and telling speakers apart, but nothing close to general intelligence.

Sources

  1. 1.B. J. Kagan et al. In vitro neurons learn and exhibit sentience when embodied in a simulated game-world. Neuron 110(23), 2022. DOI 10.1016/j.neuron.2022.09.001.
  2. 2.H. Cai et al. Brain organoid reservoir computing for artificial intelligence. Nature Electronics, 11 December 2023. DOI 10.1038/s41928-023-01069-w.
  3. 3.L. Smirnova et al. Organoid intelligence (OI): the new frontier in biocomputing and intelligence-in-a-dish. Frontiers in Science, 28 February 2023.
  4. 4.T. Hartung et al. The Baltimore declaration toward the exploration of organoid intelligence. Frontiers in Science, 28 February 2023.
  5. 5.F. D. Jordan et al. Open and remotely accessible Neuroplatform for research in wetware computing. Frontiers in Artificial Intelligence 7, 2 May 2024.
  6. 6.M. Kataoka, C. Gyngell, J. Savulescu, T. Sawai. The Donation of Human Biological Material for Brain Organoid Research: The Problems of Consciousness and Consent. Science and Engineering Ethics, 2024.
  7. 7.Ann Bednarz. Gartner's top tech trends for 2027 expand, and constrain, AI capabilities. Network World, 20 October 2026.
  8. 8.David Hollingworth. Crystal ball time: Wetware-as-a-service among key 2027 trends, Gartner says. CyberDaily, 21 October 2026.
  9. 9.Biotech Dispatch. Australian startup launches the world's first commercial biological computer. 4 March 2025.
  10. 10.Cortical Labs. CL1.
  11. 11.Wayne Williams. Cortical Labs offers weekly access to real neuron computing for under the cost of a Nintendo Switch 2. TechRadar, 14 June 2025.
  12. 12.Capital Brief. Cortical Labs launches first human brain cell data centre. March 2026.
  13. 13.Roselyne Min. Report on Cortical Labs' Melbourne data centre using human neurons. Euronews, 4 April 2026.
  14. 14.Deni Ellis Béchard. How human neurons on a chip learned to play Doom. Scientific American, 28 March 2026.
  15. 15.BioSpectrum Asia. Report on DayOne's Singapore biological data centre prototype with Cortical Labs and NUS Medicine. August 2026.
  16. 16.FinalSpark. Neuroplatform.
  17. 17.Andrew Paul. A 'brain organoid' biochip displayed serious voice recognition and math skills. Popular Science, 12 December 2023.
  18. 18.Harvard Magazine (2023). Article on the bioethics of brain organoids, with Insoo Hyun.
  19. 19.Rishika Agarwal. How Are Indian Institutions Advancing Organoid Research? BioSpectrum India, 31 July 2025.
  20. 20.Tampere University. Neuro Group shared a large microelectrode array dataset for public use. 1 June 2022.
  21. 21.reservoirpy: a Python library for reservoir computing (GitHub).
All posts