Singapore has switched on a computer built from living human neurons. The engineering is remarkable, the energy maths is seductive, but the hardest questions have barely been asked.
Somewhere inside the Life Sciences Institute at the National University of Singapore, a technician arrives every third day to feed a computer.
This is not a metaphor. The machine in question is a 20-unit server rack holding Cortical Labs CL1 biological computing units, each one containing lab-grown human neurons living on a silicon chip. The cells need a nutrient solution. They need their temperature held steady, their gas mixture regulated, their waste filtered away. Left alone, they die. Managed properly, they survive for about six months, after which they are replaced.
The system went live on July 16, 2026. On August 6, more than 80 guests from industry, government and academia watched a live demonstration of the units running, complete with microelectrode array integration and real-time neural network activity on screen.
The formal unveiling followed on August 17. NUS Medicine, which built the prototype with Singapore data centre operator DayOne and Melbourne-based biotechnology firm Cortical Labs, describes it as the world’s first independently operated biologically integrated server rack.
What is actually in the box
Every CL1 unit is a self-contained life support system with a computer attached. Human neurons, grown from induced pluripotent stem cells that were themselves reprogrammed from adult donor skin or blood samples, are cultured across a planar electrode array. The array is essentially metal and glass. Electrodes send electrical impulses into the neural tissue and read the responses back out.
Cortical Labs wraps this in what it calls biOS – Biological Intelligence Operating System. The software runs a simulated environment and feeds information about that environment directly into the culture.
The neurons fire in response, and their firing changes the simulated world. Read, act, write, repeat, in loops that close in under a millisecond. Developers can deploy code to the unit the way they would to any other machine. There is a touchscreen showing the cells’ vital signs, and USB ports for cameras or actuators.
The neuron count is worth pausing over, because the public numbers do not agree. NUS and several outlets have described each unit as holding at least 200,000 neurons.
Cortical Labs’ own published specification for the CL1, going back to its commercial launch in 2025, puts the figure at roughly 800,000 per unit, which is also the number implied by the widely reported total of 16 million living neurons across the 20-unit rack.
The gap may reflect a conservative floor rather than a contradiction. It has not been formally reconciled, and anyone modelling the technology should treat the per-unit figure as unsettled.
Either way, the scale is modest. A human brain holds something in the region of 86 billion neurons. Sixteen million is a rounding error against that, and no one involved is claiming otherwise.
Why this is not a neuromorphic chip
The distinction that matters here is easy to miss. Neuromorphic computing has existed for years. Intel’s Loihi, IBM’s TrueNorth, and a growing cluster of European research programmes all build chips that imitate the way neurons behave, using spiking architectures and event-driven design to cut power draw. The Netherlands is assembling a neuromorphic hub on exactly this principle. Every one of those systems is made of transistors. They mimic biology. They are not biology.
The CL1 inverts that. It uses actual human neurons as the computing substrate, and treats the silicon as the interface rather than the processor. That is the uniqueness of the concept, and it produces properties that no transistor design can replicate.
Biological neurons store and process information in the same place, as opposed to the memory and compute separation that has bottlenecked conventional computer architecture since the 1940s.
They rewire themselves in response to stimuli, which means they adapt rather than being trained in the way a neural network is trained. They operate on chemistry rather than switching voltage, which is why the power figures look the way they do.
They also die, which no transistor does, and that single fact reshapes the entire commercial proposition.
The energy arithmetic
Cortical Labs says a single CL1 draws roughly 25 to 30 watts. The full 20-unit rack in Singapore consumes between 850 and 1,000 watts.
For comparison, one Nvidia H100 accelerator draws around 700 watts on its own, and a conventional AI server rack runs into the tens of kilowatts.
Set that against the macro picture and the appeal is obvious. Global data centre electricity consumption reached 415 terawatt hours in 2024, about 1.5% of world demand, and the International Energy Agency projects it will roughly double to around 945 terawatt hours by 2030, close to Japan’s entire annual electricity use.
Consumption from AI-focused facilities is set to triple over the same period. Capital expenditure by the largest technology companies passed $400 billion in 2025, and is expected to rise by another three-quarters this year, which is more than global investment in oil and gas production. Against that backdrop, a rack that sips a kilowatt looks like an escape route.
It is not, or not yet. The comparison flatters the biology because the workloads are not equivalent. An H100 is doing matrix multiplication at industrial scale for a specific and enormously valuable set of tasks.
Sixteen million neurons in a nutrient bath are not doing that, and will not be doing that soon. The honest version of the claim is that biological computing might eventually be efficient at a different category of problem, not that it is efficient at the same problems. Nobody is replacing a training cluster with a fish tank.
There is also a hidden energy cost that rarely appears in the comparisons. Growing neurons from stem cells is expensive, laborious, and takes place in laboratories that consume power of their own.
Replacing every culture twice a year across a large deployment is a recurring biological and financial burden with no silicon equivalent. A rack that draws a kilowatt but needs a molecular biology facility standing behind it has a total cost profile that no wattage figure captures.
Why Singapore, and why now
The location is not incidental. Singapore has spent seven years managing a collision between digital ambition and physical constraint.
In 2019, the government imposed a de facto moratorium on new large-scale data centre approvals, driven by concern over energy, water, and land. The pause held until 2022, when a pilot Data Centre Call for Applications reopened the door on strictly selective terms.
Roughly 80 megawatts went to four operators in 2023, among them Equinix, Microsoft, GDS, and an AirTrunk-ByteDance consortium. In May 2024, the Infocomm Media Development Authority published its Green Data Centre Roadmap, promising at least 300 megawatts of additional near-term capacity, and tying it explicitly to efficiency and green energy conditions.
A second call, DC-CFA2, launched on December 1, 2025, offered at least 200 megawatts. Applications had to befiled on or before March 31, 2026. Applicants had to demonstrate best-in-class efficiency, a power usage effectiveness ceiling of 1.25 at full load, and at least half their power drawn from approved green sources.
Singapore now hosts more than 70 data centres totalling roughly 1.4 gigawatts, in a country of 730 square kilometres with no domestic energy resources to speak of. Vacancy rates have fallen close to one per cent. Every additional megawatt is rationed.
That is precisely the environment in which a technology promising radical efficiency gets a hearing it would not receive in Virginia or Johor.
Singapore cannot build its way out of the constraint. It has to compute its way out, and that makes it unusually willing to host experiments that larger markets would leave in the laboratory.
The commercial calculation
DayOne’s involvement is the part that turns a neuroscience project into a business story. The company was carved out of Chinese operator GDS Holdings in 2022 to hold assets outside the mainland, rebranded as DayOne in early 2025, and has since become one of Asia’s most aggressively capitalised infrastructure platforms.
It closed a $4.5 billion Series C in June 2026, led by Coatue and Hillhouse with participation from Indonesia’s sovereign wealth fund, at a reported valuation near $20 billion.
It has secured more than 1.5 gigawatts of customer bookings across Asia Pacific and Europe, is negotiating a corporate loan facility of up to $7 billion, and has confidentially filed for a US listing that could raise $5 billion.
A company at that stage of its life does not attach itself to a university biology project for the science. The NUS rack is a validation phase, structured to transition into a live deployment inside a commercial DayOne facility in Singapore.
The stated ambition is a large-scale biological data centre, the first outside Australia, eventually housing as many as 1,000 CL1 units subject to regulatory approval and safety testing.
For an operator heading into public markets during an AI infrastructure boom, being the only listed name with a credible biological computing asset is worth something regardless of whether the technology works at scale.
Investors have shown a consistent willingness to pay a premium for optionality on the next architecture. That is not a criticism of DayOne. It is simply the commercial logic that makes a project like this fundable at all.
The pricing tells a similar story. A CL1 sells outright for around $35,000, falling to roughly $20,000 per unit when bought in 30-unit racks. Cortical Labs also runs a cloud model it calls Wetware-as-a-Service, originally priced at $300 per week.
Reporting around the Singapore launch has put access at about $2,200 a month, roughly half what major cloud platforms charge for comparable high-end AI chip access.
Undercutting the hyperscalers on price is a recognisable go-to-market strategy. It also implies that the company expects to be compared with them, which is a bolder claim than the science currently supports.
What it can actually do
Cortical Labs made its name in 2022 with DishBrain, a peer-reviewed study in which human and rodent neurons were connected to a simulated game of Pong, and appeared to improve under closed-loop feedback. The work was genuinely significant. It also demonstrated learning in a very narrow sense, not general intelligence, and certainly not readiness for enterprise workloads.
Four years on, the capability question remains the weakest link in the story. Founder and chief executive Hon Weng Chong says the prototype shifts the conversation from research to commercial application, and points to drug discovery, humanoid robotics, cybersecurity, and fraud detection as the promising areas.
The common thread is that these are domains where data is scarce, unpredictable or expensive to simulate, which is where adaptive biological systems might plausibly outperform statistical models trained on enormous corpora.
NUS Medicine’s own interest is more concrete and arguably more defensible. Professor Rickie Patani, who directs the Neurobiology Programme at the Life Sciences Institute and supervises the cultures, has emphasised the platform’s value for studying learning and adaptation at their biological source, for modelling neurological disease, and for testing compounds on human neurons rather than animal tissue. That is a real and immediate use case with a clear ethical advantage over animal testing.
Whether it justifies calling the installation a data centre is a separate question. A rack of 20 units in a university laboratory is a research instrument. The data centre framing belongs to the commercial roadmap, not to what exists today.
The question nobody has answered
The ethics are unresolved, and the unresolved parts are not the ones that get the headlines. Public debate fixates on consciousness. Could a sufficiently large culture of human neurons, given sensory feedback and a simulated world, experience something?
Bioethicists have argued that if biocomputers become conscious, they acquire moral status, which would place hard limits on permissible research.
The trouble is that consciousness has no agreed definition and, therefore, no agreed test. Sixteen million neurons is almost certainly nowhere near any plausible threshold. Nobody can say where the threshold is.
The more immediate problem is consent. A comment published in Nature in July 2026 pointed out that donors whose skin or blood cells were reprogrammed into these neural cultures were not, in most cases, told that their tissue might end up as the computing substrate of a commercial machine.
Existing consent frameworks were written for medical research, not for biocomputing. That gap is administrative rather than philosophical, and it is fixable, but it has not been fixed.
Meanwhile the scientists who built the brain organoid field are increasingly uneasy. At a meeting at Asilomar in late 2025, researchers, ethicists and legal experts warned that inflated commercial claims about organoid intelligence risk a public backlash that damages legitimate medical research.
Human neural organoids fall outside the regulatory structures governing both human and animal research. A March 2026 paper in Science called for international oversight specifically covering biocomputing applications, on the grounds that this use case sits in territory research ethics committees were never designed to police.
Singapore, which has built its reputation on regulating emerging technology early and precisely, now hosts the most advanced deployment of a technology with no regulatory category. That is either an opportunity to write the rules first, or an oversight waiting to be noticed.
The honest assessment
What has been switched on in Singapore is a demonstration of extraordinary engineering discipline that answers a question nobody has yet posed clearly.
The energy comparison is real but not yet meaningful, because the workloads are not comparable. The commercial model is coherent but depends on capabilities that have not been shown.
The scientific value, particularly for disease modelling and replacing animal testing, is the most solid part of the proposition, but receives the least attention.
The rack does one thing unambiguously well. It forces a question that the AI industry has spent a decade avoiding, which is whether the path to more capable machines runs through more silicon or through a different substrate entirely.
Feeding a computer every three days is an absurd way to run infrastructure. It is also the only computing architecture anyone has built that mimics how the human brain actually learns.