In February 2026, a video of human neurons “playing Doom” went around the internet. Half the comments said we were living in a horror film. The other half said it was a party trick.
Both camps skipped the part worth knowing: the Doom clip is the fourth in a series of experiments, each with checkable numbers, each showing a different strength and a different weakness of the idea. This article goes through them in order, tells you which claims are peer-reviewed and which are marketing, and ends with a way to size up the next headline yourself.
Short answer. Living human neurons on chips are real, commercially available, and now installed in data-center racks. They learn simple tasks quickly from little data and draw very little power. They are also tiny (frog-scale at best), short-lived (roughly 3 to 6 months), and unproven against ordinary AI on the same job. Useful lab instrument today; not a GPU replacement.
What “merging brain cells with silicon” actually means
Nobody is connecting a person’s brain to a server. The process is more mundane:
- Take ordinary cells (for example, from blood or skin) and reprogram them into stem cells (iPSCs).
- Coax those into neurons, either as a flat layer or as a 3D clump called an organoid.
- Grow them on a chip covered in electrodes, in a life-support system that feeds them.
- Send information in as electrical pulses, and read their firing back out.
Software does the translating. A game or a sound is turned into stimulation patterns, and the cells’ firing is turned into an action. The neurons never “see” anything. That closed loop is the whole technology.
Case 1: DishBrain learns Pong (2022)
Setup. Brett Kagan’s team at Cortical Labs grew roughly 800,000 neurons (a mix of mouse and human-derived cells) on a high-density electrode array. Ball position went in as stimulation. Activity in designated regions moved the paddle. A hit produced predictable feedback; a miss produced random noise. The cells reorganized to avoid the noise, and performance improved within about five minutes.
Source. Kagan et al., Neuron 110, 2022. The study was co-authored by neuroscientist Karl Friston, whose free-energy principle inspired the feedback design.
What it proves. Cultured neurons can change their behavior in a closed loop, fast, from very little data.
What it doesn’t. The paper’s title used “sentience” in a narrow technical sense (responding to sensory input). Plenty of readers heard “conscious,” and the backlash produced ethics papers, including one specifically about DishBrain. Cortical Labs now says it works with bioethicists. Also keep in mind that Pong is a modest task; the interesting part was the speed of adaptation, not the score.
Case 2: Brainoware recognizes voices (2023)
Setup. Feng Guo’s group at Indiana University Bloomington used a 3D brain organoid as a “reservoir” in reservoir computing. The organoid scrambles inputs in a rich, nonlinear way, and a small conventional model reads the output. The test was 240 audio clips of eight people saying Japanese vowels; the system had to identify the speaker.
Result. Accuracy rose from about 51% to 78% over two days of training. The system also predicted a chaotic equation. Published in Nature Electronics.
The honest part. Guo told MIT Technology Review the accuracy was low, and the study itself reported it trailing artificial neural networks. Lena Smirnova of Johns Hopkins pointed out that organoids don’t truly hear speech; they react to electrical pulses. Guo’s own description was “proof-of-concept.”
Case 3: FinalSpark rents out organoids (2023 to now)
Setup. The Swiss company FinalSpark skipped the consumer product and built a research service. Its Neuroplatform lets scientists run experiments on human brain organoids remotely, through a Python library or Jupyter notebooks. A microfluidic system automates feeding, and the hardware records activity around the clock.
What the peer-reviewed paper says. In Frontiers in Artificial Intelligence (2024), the team reports organoid lifetimes of more than 100 days, use over three years with more than 1,000 organoids, and over 18 terabytes of data collected. At the 2024 launch, it offered 16 organoids and free access to nine institutions.
What the paper doesn’t establish. The headline “a million times less power than silicon” appears in the company’s press release, not as a measured benchmark. Nobody has published the same task run on both systems under the same rules. Treat it as an estimate of potential, not a result.
Case 4: CL1 and the Doom clip (2025 to 2026)
Product. In March 2025, Cortical Labs launched the CL1, described as the first commercially available, code-deployable biological computer, with a reported starting price of $35,000 and a cloud option the company calls Wetware-as-a-Service. The CL1 later won a 2025 Good Design Award in the medical and scientific category.
Doom. In February 2026, the company released a video of roughly 200,000 neurons playing Doom, reported by Scientific American and others. The first working version was built by independent developer Sean Cole, who used the company’s cloud platform and Python API, reportedly in under a week. Kagan’s own verdict in the video: the cells play “like a beginner who’s never seen a computer.”
Read this carefully. The Doom demo is a company video, not a peer-reviewed result. What it does show is accessibility: an outside developer programmed living neurons in days. That is the story, more than the gameplay.
The rack. In August 2026, NUS, data-center operator DayOne and Cortical Labs unveiled a rack of 20 CL1 units in Singapore, about 16 million neurons in total, comparable to a frog’s brain according to the report. The same report says the neurons last about six months before replacement, and that Cortical Labs puts each CL1 at about 25 watts.
The scale problem, in numbers
| System | Approx. neurons | How it compares |
|---|---|---|
| FinalSpark organoid | ~10,000 | smaller than a speck of tissue |
| DishBrain (2022) | ~800,000 | the Pong experiment |
| CL1 Doom demo (2026) | ~200,000 | per company video coverage |
| Singapore rack (20 CL1) | ~16 million | frog-brain scale, per ioplus |
| Human brain | ~86 billion | roughly 5,000 times the rack |
Neuron counts differ between reports (the ioplus article cites about 800,000 per CL1; Doom coverage says about 200,000 for that run), so confirm per-unit specs with the manufacturer before quoting a number.
Claim checker: how much to trust each statement
| Claim | Status | Why |
|---|---|---|
| Neurons on a chip can learn Pong in minutes | Peer-reviewed | Neuron, 2022 |
| Organoid reservoir reaches 78% speaker ID | Peer-reviewed, and below conventional AI | Nature Electronics, 2023 |
| Organoids survive 100+ days; 1,000+ used | Peer-reviewed (company authors) | Frontiers, 2024 |
| Neurons play Doom | Company video, no paper found | Treat as a demo |
| “A million times less energy” | Company claim | No same-task benchmark published |
| A CL1 uses about 25 to 30 W | Company figure | Excludes some support hardware; verify |
| This replaces GPUs | Unsupported | Scale, lifespan and I/O all fall short |
Where the energy story gets slippery
The brain runs on roughly 20 watts, and AI’s electricity appetite is real; the International Energy Agency has a report on it. That is the strongest argument for biocomputing.
But watch the accounting. Reports on Cortical Labs’ Melbourne data center mention 120 CL1 units at around 30 watts each (per a secondary source), which is about 3.6 kilowatts. That’s not nothing, and it is spent on keeping cells alive, not on computing. A chip’s watts and a neuron’s watts aren’t the same measurement. Until someone publishes “same task, same accuracy, total system power,” the efficiency claim is a promising hypothesis.
Limits nobody should gloss over
- Lifespan. Roughly six months for CL1 neurons and 100+ days for FinalSpark organoids, against years for silicon.
- Bandwidth. One secondary source reports the CL1 chip has a 59-electrode array, a narrow pipe for hundreds of thousands of cells. Check the manufacturer’s current spec sheet.
- Reproducibility. Organoids are always firing, so a clean baseline is hard, and no two cultures are identical.
- Benchmarks. Very few head-to-head comparisons against standard AI exist.
The ethics, without the drama
There is no evidence that these cultures are conscious, and at this size most researchers consider it very unlikely. The questions the field is asking are still fair:
- Where did the donor cells come from, and what did the donors consent to? Per reporting on the company, labs buying a CL1 must hold ethical approval for generating their cell lines.
- Who owns something built from a person’s cells?
- What would change at larger scales, and who decides where the line is?
The Johns Hopkins group that coined “organoid intelligence” in 2023 argues for building ethics in parallel with the technology. Their Nature Electronics commentary ran alongside Brainoware. For the philosophical side, see Milford, Shaw and Starke, “Playing Brains: The Ethical Challenges Posed by Silicon Sentience and Hybrid Intelligence in DishBrain.”
What to do with this information
- Researchers: the proven use is disease modeling, drug screening and learning studies, and cloud access (FinalSpark’s Neuroplatform, Cortical Cloud) avoids a hardware purchase. Check what each platform’s license and cell-line ethics requirements are first.
- Investors and journalists: ask four questions. How many neurons? How long do they live? Compared against what, on which task? Is it peer-reviewed or a company video? A missing answer to any of the four is information.
- AI engineers: read biocomputing as a source of ideas for low-power, data-efficient learning. Don’t plan deployments around it this decade. Kagan himself talks in terms of a decade or more in a 2026 JMIR feature.
- Curious readers: the fair summary is impressive, tiny and early, and the two halves of that sentence are both true.
My read
What changed my mind wasn’t Doom. It was the pace: in four years the field went from a single lab’s Pong study to remote platforms, a commercial product, and a rack in a data center. Learning from small amounts of noisy data is a real weak spot of today’s AI, and neurons do it natively. But a frog-sized brain that you replace twice a year is a laboratory instrument, and a good one. The useful habit is to keep checking each new claim against the same four questions above.
FAQ
Are the neurons from real human brains? No. They are grown from stem cells that were reprogrammed from ordinary adult cells, per Cortical Labs’ description of the CL1.
Can a CL1 or organoid be conscious? There is no evidence of it, and the cultures are orders of magnitude smaller than a human brain. The ethical debate is about future scale and consent, not about current devices.
How long do the neurons live? About six months in CL1, and 100+ days for FinalSpark organoids as published.
Can I use one? Yes, through purchase (reported from $35,000) or cloud access, and FinalSpark offers research access for institutions. Both require research use and, for cell lines, ethical approval.
Is biocomputing more energy-efficient than AI chips? Plausibly, but unproven for equal tasks. The one-million-times figure is a company estimate.
Sources and further reading
- Kagan et al. (2022), Neuron: doi:10.1016/j.neuron.2022.09.001
- Cai et al. (2023), Nature Electronics: doi:10.1038/s41928-023-01069-w; Smirnova et al. commentary: doi:10.1038/s41928-023-01096-7
- Jordan et al. (2024), Frontiers in Artificial Intelligence: doi:10.3389/frai.2024.1376042
- MIT Technology Review on Brainoware
- Scientific American on the Doom demo
- ioplus: Singapore neuron rack, Aug 2026
- BioPharmaTrend: CL1 launch
- IEA: Energy and AI
Figures in this field change quickly. Verify against the primary sources before citing, and update the “Updated” date when you do.