Reading time: 6 min
An independent developer with no background in biological computing sat down with a Python API, connected to a chip carrying 200,000 living human neurons in a Melbourne laboratory, and had them controlling the 1993 first-person shooter Doom within seven days. The neurons aren’t good at it. They play like someone who has never seen a keyboard. But they are learning. And the company that built the system has already shipped 115 units at $35,000 each.
The company is Cortical Labs, founded by Dr. Hon Weng Chong and headquartered in Melbourne. The product is the CL1, which the company calls the world’s first commercially available biological computer. It fuses lab-grown human neurons with a silicon chip to create what Cortical Labs describes as Synthetic Biological Intelligence, or SBI. The Doom demonstration, announced via YouTube in late February and picked up this week by Popular Science, Tom’s Hardware, The Register and roughly every tech outlet with a pulse, is the moment the technology crossed from lab curiosity into something developers can actually interact with. The source code is on GitHub. The API is in Python. The neurons are alive in a nutrient bath in Australia.
How the Thing Actually Works
The CL1 grows human neurons on a 59-electrode multielectrode array built on a chip of metal and glass. The neurons are not extracted from brains. They begin as cells taken from adult skin or blood donors, which are reprogrammed into induced pluripotent stem cells and then differentiated into cortical neurons, per a detailed technical breakdown published by PerfScience in March 2026. The cells are kept alive inside a sealed life-support chamber that regulates temperature, gas exchange and nutrient flow. Cortical Labs says the neurons can survive for up to six months.
The system runs on what the company calls biOS, a Biological Intelligence Operating System. It creates a simulated environment and sends information about that environment to the neurons as patterns of electrical stimulation through the electrodes. The neurons fire back their own electrical spikes. Those spikes are interpreted as actions. In Pong, the earlier demonstration from 2022, the relationship was direct: ball goes up, paddle goes up. In Doom, the problem is orders of magnitude harder. The game is three-dimensional. There are enemies. The player has to explore, navigate, aim and shoot.
To bridge that gap, Sean Cole, the independent developer who built the Doom interface, translated the game’s video feed into electrical stimulation patterns that the neurons could process. Per Tom’s Hardware, CTO David Hogan explained the mapping: “If the neurons fire in a specific pattern, the Doom guy shoots. If they fire in another pattern, he moves right.” The neurons receive reinforcement feedback, positive and negative signals that shape their firing patterns over time. It is, in the most literal sense, reinforcement learning running on wetware instead of silicon.
What It Can and Cannot Do
Brett Kagan, Cortical Labs’ Chief Scientific Officer, was direct about the current performance. “Is it an eSports champion? Absolutely not,” he said in the announcement video, per Tom’s Hardware. The system plays Doom better than a random input generator but worse than any human who has spent five minutes with the game. The significance, Kagan argued, is not the score. It is that the neurons demonstrated “adaptive, real-time, goal-directed learning” in a complex three-dimensional environment. That is a different class of problem from batting a Pong ball back and forth.
Put that in context. In 2022, Cortical Labs’ earlier prototype, DishBrain, learned to play Pong in roughly five minutes. A standard deep reinforcement learning system takes approximately 90 minutes to reach comparable performance, per PerfScience. The neurons are not faster at computing. They are more data-efficient. They learn from fewer examples. That property, if it scales, matters enormously for drug discovery and disease modelling, where the cost of generating training data is often the binding constraint.
The Business Model
The CL1 launched at Mobile World Congress in Barcelona in March 2025 and began shipping in the second half of last year. Each unit costs $35,000, or $20,000 per unit in a 30-unit server rack. A full rack draws between 850 and 1,000 watts, which, as several outlets noted, is less than some high-end gaming PCs. Cortical Labs has shipped 115 units to date, per PerfScience, generating roughly $4 million in revenue at list price. The company also offers Wetware-as-a-Service, or WaaS, allowing researchers to access neurons remotely through the Cortical Cloud.
Per IEEE Spectrum, Kagan said the company has seen strong interest from universities, startups and government groups exploring applications in drug discovery, neurocomputation, AI acceleration and, somewhat unexpectedly, Bitcoin mining. The core commercial pitch is that the CL1 is an ethically superior alternative to animal testing for pharmaceutical research. Since it uses human neurons rather than rodent models, it can capture genetic differences between donors and model disease-specific responses in ways animal testing cannot. Cortical Labs positions the technology as a platform for studying epilepsy, Alzheimer’s and other neurological conditions.
The competitive landscape is small but growing. Swiss company FinalSpark already offers remote access to neural organoids starting at $1,000 per month. At Indiana University, researchers built Brainoware, a system that achieved 78% accuracy in speaker identification after just two days of training. A team at UC San Diego has proposed using organoid-based systems for environmental modelling. In China, Tianjin University introduced MetaBOC, a brain-on-chip platform. None of these are as far along commercially as Cortical Labs, but the venture capital flowing into anything adjacent to AI has made speculative bets in biocomputing suddenly fundable.
The Question Nobody Wants to Answer
The ethical dimension is unavoidable and Cortical Labs has not tried to avoid it. The neurons on the CL1 are not conscious. They are too few and too simple to produce anything resembling awareness. Current neuroscience consensus holds that 200,000 neurons on an electrode array cannot generate subjective experience. A human brain contains roughly 86 billion neurons connected by trillions of synapses. The gap is not incremental. It is civilisational.
But the trajectory raises questions that the industry has not yet answered. As RealClearScience noted in a January 2026 analysis, claims of intelligence or consciousness in these systems are unsupported today. The systems display simple capacity to respond and adapt, not higher cognition. The question is what happens when the neuron count scales, when the electrode arrays get denser, when the feedback loops get more sophisticated. Cortical Labs has talked about building a “Minimal Viable Brain.” That phrase, borrowed from startup culture and applied to biological tissue, carries implications the company may not fully control.
For now, though, the CL1 is a $35,000 box that keeps human neurons alive for six months, lets anyone with Python skills talk to them, and has just been taught to shoot demons in a game from 1993. The fact that the Doom interface was built by a single developer in a week, using a public API, with no neuroscience training, is the detail that should hold your attention. It means the barrier to building on this platform is already low enough for the developer community to start experimenting. What they build next is the story worth watching.
