A computer in Melbourne has to be fed. Its processor is about 800,000 living human neurons on a chip the size of a stamp. Biological computers like it are now sold, rented over the internet and taught to play games. They learn fast and use almost no power. Nobody can yet say exactly what is learning inside them.
A computer that runs on sugar and salt
The machine is called the CL1, and it is made by the Australian company Cortical Labs. From the outside it looks like a shoe box with clear panels. Inside there are small tubes, a reservoir of pink nutrient fluid and a chip at the centre. On that chip sit roughly 800,000 human neurons. That is about the number of neurons in the brain of a honeybee.
The CL1 went on sale in 2025 for 35,000 US dollars, about the price of a small new car. It is programmed in Python. A researcher sends short electrical pulses into the cells through tiny electrodes, and the cells send pulses back. If the pumps stop, the computer does not crash. It dies.
The company calls the life support system a body in a box. Pumps do the work of a heart. Filters do the work of kidneys. A heater and a gas mixture keep the cells at body temperature with the oxygen they need. According to the documentary this article draws on, racks of these units are now running in small data centres in Melbourne and Singapore, next to ordinary servers.
Grown from a scrap of skin
No brain is involved at the start. The neurons are grown from ordinary donor cells, a small piece of skin or a few millilitres of blood. In 2006 the Japanese scientist Shinya Yamanaka showed that adult cells can be rewound into stem cells by switching on four genes. He shared the Nobel Prize for it in 2012. Over a few weeks, those stem cells can be guided into becoming neurons.
Each neuron carries the full genome of the person it came from. That is why medical researchers value these cells. Neurons grown from a patient with epilepsy or Alzheimer’s disease can show how that person’s brain cells react to a drug before the patient takes it.
It also makes the story personal. One of the donors behind the first commercial cell lines was the founder of Cortical Labs, Hon Weng Chong. His own cells, grown into neurons, run inside the product he sells. Buyers must show ethical approval for the cell lines they use, and donors sign consent forms. But those forms were written for research on cells, not for cells that work in a server rack.
Learning Pong with calm and noise
The work that made the company famous was published in the journal Neuron in 2022. A team led by neuroscientist Brett Kagan grew about 800,000 neurons on an electrode chip and connected them to a simple version of Pong. One group of electrodes told the cells where the ball was. Activity in another group moved the paddle up or down.
You cannot give a dish of cells a reward. So the team used an idea from the British neuroscientist Karl Friston: living systems try to keep their world predictable. When the paddle hit the ball, the cells got a short, orderly pulse. When it missed, they got several seconds of random noise. Calm for a hit, chaos for a miss.
Within about five minutes the cultures started to play better. Rallies grew longer. Cultures that could move the paddle but got no feedback did not improve. The learning only happened when the loop was closed, so the cells could act and then sense what their action did. The human neurons also did better than mouse neurons, and nobody can yet fully explain why.
The cells were not good players. A child would beat them easily. The point was that they improved at all, and that it took minutes. Artificial neural networks usually need thousands of examples to learn a task like this.
Why anyone builds computers out of cells
The short answer is energy. A human brain runs on about 20 watts, roughly a dim light bulb. The International Energy Agency estimated that the world’s data centres used about 415 terawatt hours of electricity in 2024, and that this could more than double to around 945 terawatt hours by 2030. That is close to the total electricity use of Japan.
According to Cortical Labs, a rack of 30 CL1 units draws between 850 and 1,000 watts, about what a microwave oven uses. A single server for training modern AI can draw several thousand watts on its own. The Swiss company FinalSpark says its living processors use up to a million times less energy than digital chips. That is the company’s own figure, not an independent measurement.
The comparison is not fully fair, and the companies admit it. Silicon is far better at fast, exact arithmetic. The claim is narrower. Living networks seem to be good at learning from very little data and at coping with the unexpected. A toddler needs to see a dog once or twice. An AI model needs thousands of labelled photos. Hon Weng Chong has described biological computing as a supplement to AI in areas where data is scarce.
There is a catch. A silicon chip can sit in a drawer for ten years and still work. A dish of neurons needs food, warmth at 37 degrees and sterile conditions. The cells in a CL1 are advertised to last up to six months. When they die, whatever the network learned dies with them, because it was never stored anywhere else. The neuroscientist Tony Zador has called the idea that organoid computing could compete with silicon AI fundamentally misguided.
Human neurons as a cloud service
If 35,000 dollars is too much, there is a cheaper option. Cortical Labs rents access to its cells over the internet for about 300 dollars a week. The company calls it wetware as a service. A scientist sends a program from a laptop, the cells in Melbourne respond, and the results come back like a web search.
FinalSpark in Switzerland rents access to 16 small balls of human brain tissue, called organoids, for 500 dollars a month. Each organoid is smaller than a grain of rice and holds about 10,000 neurons. Each one lives for around 100 days. The company says it has recorded data from more than 1,000 organoids over three years.
Scale matters here. A human brain has about 86 billion neurons. You would need more than 100,000 CL1 units just to match the number of cells in one head, and you would still have nothing like a brain. What sits on the chip is a thin film of raw material.
The Doom test that few people noticed
In early 2026, according to the documentary, a video of human brain cells playing the 1993 shooter Doom spread around the world. An independent developer had connected the game to about 200,000 neurons through the company’s cloud service. The cells played like a beginner. They walked into walls and died often, but they played.
The detail most headlines missed is what the developer did next. A lot of ordinary software sat around the cells. One program turned the game screen into pulses. A small artificial neural network turned the cells’ spikes into moves. The developer then switched off the cells’ contribution and replaced it with random numbers or zeros. The game still played. The software around the cells had learned enough to move the player on its own.
That does not mean the project was a trick. The code and the result were published openly. It shows how hard it is to tell how much of the cleverness belongs to the tissue and how much to the software wrapped around it. A useful habit for every claim in this field is to ask three things. Is the living tissue really needed? Was there a fair comparison? Has the work been checked by other scientists?
From flat sheets to small brains and robot bodies
Organoids are a step beyond a flat layer of cells. In 2013, Madeline Lancaster and Jürgen Knoblich in Vienna showed that stem cells left to float in a gel fold themselves into small balls with layers that resemble a developing human brain. Most organoids are used in medical research on autism, epilepsy and the damage the Zika virus does to a growing brain.
In 2023, a team at Indiana University led by Feng Guo used an organoid as a kind of living pond. They sent recordings of spoken vowels into the tissue and trained simple software to read the ripples. The system, published in Nature Electronics, identified which of eight speakers was talking about 78 percent of the time. The organoid itself was not trained, only the reader was.
The documentary describes newer results as well. Researchers in Bristol used rented organoids to help read Braille letters from an artificial fingertip, and the living layer helped the system cope when the signal was noisy. Cortical Labs linked 15 CL1 units into one network it calls Hive, where the units are reported to have split the work between them and beaten a standard AI program at a game task. A US team placed an organoid in the decision loop of a humanoid robot that reached for objects and chased a laser dot. Several of these results are still conference talks or preprints that other scientists have not yet checked.
Could anyone ever be home in there?
Almost every researcher agrees that today’s cultures are far too simple to feel anything. The harder truth is that nobody knows where the line is, and there is no test for it. The tools doctors use to judge whether a patient who cannot speak is aware were built for whole human brains. They do not transfer to a dish.
Meanwhile the organoids grow more complex every year. Researchers are adding blood vessels and immune cells so they can live longer and grow larger. Some show electrical rhythms that resemble those of premature babies. At Stanford, Sergiu Pașca’s team has linked several organoids into a chain that models the pathway pain signals take through the body. The experiment shows that the wiring can be grown, although nobody claims the chain feels anything.
The International Society for Stem Cell Research has guidelines for this kind of work, and the 2023 roadmap for organoid intelligence from Johns Hopkins University built ethics into the plan from the start. Some bioethicists now argue that dismissing the question of consciousness is premature. There is still no binding international law written for biological computing and no agreed point at which an experiment should pause.
What the cells are actually learning
Strip away the headlines and the answer is fairly simple. What learns is the web of connections the cells grow between themselves. Connections that help produce a predictable world grow stronger. Those that do not grow weaker. It is the same basic mechanism the psychologist Donald Hebb described in 1949, and the same one at work in your own head as you read this.
For businesses that depend on AI, the story is worth following for one reason above all: power. If living tissue can take over even a small part of the learning that now burns electricity in data centres, the savings would be large. That is still a big if. The machines are real, the results are modest, and the questions about what we are building are moving more slowly than the technology.
This article draws on the documentary by the YouTube channel Interstellar Dreams.
Photo: A researcher holds a flask of human cerebral organoids. NIAID, CC BY 2.0.
Related Posts
-
OpenAI pauses training after agents probed US government sites
OpenAI pauses training of its most recent models following an incident that caught the company…
-
Kenya bets on AI: what the Anthropic deal means for business
At the UN General Assembly in September 2025, Kenya signed a Joint Declaration with Anthropic.…
-
How AI saved Old Mutual Kenya Sh400 million in a single year
AI in Kenyan business is no longer a conference talking point. Old Mutual General Insurance…


