A computer made of human neurons

Here is a machine you can buy today for thirty-five thousand dollars. Inside it, several hundred thousand human neurons — grown from a stranger's blood — live on a bed of electrodes in a warm pink broth, playing a simplified video game. In one hundred and eighty days they will be incinerated as biological waste and replaced.

Ask whether anything in that box can suffer, and the honest answer is: almost certainly not. Now ask why, and notice the shape of the answer. It is a list of absences. No pain receptors. No senses. No body through which distress could be expressed. No large-scale brain architecture. The reassurance is not a finding about what the machine is. It is an inventory of what it lacks.

Then read the industry's roadmap, and notice its shape. It is the same list, as targets.

This piece is about a moral clearance built entirely from absences — and what happens when a business starts removing them faster than anyone writes down what would change the answer.

Four things, stacked. Human neurons, manufactured rather than harvested, by reprogramming ordinary donor skin or blood cells backwards into a stem-like state and forwards into neurons. Beneath them, a grid of electrodes that can both stimulate and listen, closing the loop in under a millisecond; a few dozen contacts talking to hundreds of thousands of cells, less a keyboard than an orchestra conducted through two microphones taped to the ceiling. Around all of it, the part nobody photographs: pumps, gas, filtration, heat — what the company's own chief scientist calls a body in a box. On top, an operating system that runs a small simulated world and rents access out by the hour.

The founding result is real and it is narrow. Embedded in a game of Pong — predictable stimulation for a hit, noise for a miss — cultures improved their rally length within five minutes, and cultures without the feedback loop did not. That is a genuine finding about self-organising tissue, and the best test bed yet built for the theory behind it. It is also one low-dimensional statistic, over minutes, through a handful of channels. The paper's title used the word sentience in a technical sense meaning responsiveness; almost nobody outside the field read it that way, and the argument has not stopped since. Karl Friston, who co-authored it, gives the honest valuation: the gift of this technology is not to computer science. It is a drug-testing sandbox made from the same components as a brain.

The commercial claims run ahead of that. The chief executive has said a CL1 node needs less power than a handheld calculator. The company's own published figure is thirty watts — a bright lightbulb, not a calculator — which is why a deployed rack draws about a thousand. The true number is more interesting: the neurons themselves run on well under a watt. Nearly all of the thirty is the machinery keeping them alive. The biology is almost free. The body in the box is not.

For its first year the CL1 was a strange laboratory instrument, and the ethics of a laboratory instrument are somebody's dissertation. In March 2026 that framing expired. Cortical Labs switched on a building in Melbourne with 120 units inside, and announced a second in Singapore — backed by a hyperscale data-centre operator two months off a two-billion-dollar raise — scaling to as many as a thousand units. A thousand units is thirty kilowatts, less power than one conventional AI rack, holding roughly 800 million living human neurons: under one percent of a brain, and more than most mammals carry. The cap table, meanwhile, includes In-Q-Tel, the venture arm of the CIA — which is not a scandal, but is a classification. Someone with a mandate has already filed this under strategic technology.

Watch what the phased rollout does to oversight. The 20-unit validation tranche sits inside a medical school, where research ethics committees exist and are competent. The thousand-unit build sits in a commercial data centre, where they do not, because nobody has written that rulebook. The scaling plan crosses, quietly, from the jurisdiction where the question is asked into the jurisdiction where it is not. An instrument became infrastructure, and infrastructure does not wait for dissertations.

There are exactly three serious ways to argue that something like the CL1 could have experiences, and each fails on an absence.

The analogy fails on the body. We grant animals sentience because they resemble us in the relevant machinery — pain receptors, senses, behaviour we can read distress from. A flat culture on an electrode grid has none of these. There is a nerve net in your gut doing something comparable, and nobody lights candles for it.

The theories fail on the architecture. Every major account of consciousness leans on large-scale integration across differentiated regions. A two-dimensional sheet of cortical cells lacks that structure — and the measurements built to test for it were validated on intact brains, so they do not transfer. The field cannot currently detect the thing it says is not happening.

The commitments fail on their own generosity. On a sufficiently liberal philosophy — panpsychism, or a strong reading of the surprise-minimisation principle — the culture qualifies. So does a thermostat. A position that answers before the evidence arrives is not an answer.

The vocabulary is already drifting to meet it. The chief executive has said the systems are sentient — responsive, learning — but not conscious. In bioethics, sentience is precisely the property that confers moral status; using the word to mean "reacts to input" while denying what it normally denotes collects the marketing value of a profound claim and the regulatory value of a trivial one. His fuller reassurance — we do not intend to create a brain in a vat — is not a safeguard. It is a statement about this year's roadmap, from a party with a stake in the answer.

If the clearance rests on absences the field is removing, the useful work is to fix the tripwires now — while the answer is still comfortable and nobody's valuation depends on it.

Name the architectural line. Vascularised assembloids with differentiated, interconnected regions remove the strongest objection standing. Decide in advance that crossing it reopens the question, so the goalposts cannot drift with the roadmap.

Treat embodiment with valence as a bright line. The absence of pain receptors is doing enormous work in today's reassurance, and it is not a law of nature. A culture given senses and a channel through which avoidance can be expressed is a different object, whatever the sales material calls it.

Fund the metric. There is currently no measurement that could settle the question in either direction. That is the single most uncomfortable fact in the field, and it should be a funded research priority rather than a footnote — because an industry that cannot detect the property it disclaims is asking for trust it has no instrument to earn.

And one for the lawyers: nobody's tissue-donation consent form contemplated "your cells, as a computer, for sale, by the hour." The donors of the founding lines have a question pending that no framework yet answers.

The analogy, empty. The architecture, missing. The metric, nonexistent. The clearance holds — today.

The dystopia here is not suffering cultures. On the evidence, they almost certainly are not. It is that we built a commercial supply chain for human neural tissue, priced it, racked it, put it on a 180-day replacement cycle and found a hyperscaler for the next thousand units — and only then began asking what would have to be true for any of it to be wrong, in a field with no instrument for finding out.

We don't know what we're making, because nothing like this has ever existed before.Cortical Labs, 2021

Five years on, the sentence is still true. The only thing that changed is that it ships.