Somewhere in a lab at the National University of Singapore there is a server rack that needs to be fed. Not powered down and rebooted. Fed. It runs on 16 million living human neurons, grown from stem cells, sitting on chips, and if you stop delivering nutrients to them they die, the way any tissue dies. That is not a metaphor for a fragile supply chain. It is the literal maintenance requirement of a computer that NUS and the Australian startup Cortical Labs unveiled in August 2026, and it is, per The Straits Times, the world’s first independently operated biological server rack.
I want to be careful with the word “computer” here, because it is doing a lot of work. This is not a chip inspired by the brain, the way a neuromorphic processor borrows the architecture of neurons without the biology. This is actual neural tissue, wired into silicon, doing the computing itself. The distinction matters because it flips a premise that biotech and hardware people have chased for decades: using bioelectricity to power machines. Cortical Labs did the opposite. They made living cells the machine.
Twenty CL1 units, 800,000 neurons each, one very literal feeding schedule
The hardware, per The Straits Times reporting, consists of 20 biological computers called CL1 units. Each contains roughly 800,000 live human neurons cultured directly onto a silicon chip studded with micro-electrodes. Multiply it out and you get the headline number: about 16 million neurons working across the rack. Cortical Labs already sells the CL1 as a standalone commercial biological computer, and the company built its reputation on an earlier demonstration in which cultured neurons learned to play Pong, adjusting their firing patterns in response to a paddle and a ball rendered as electrical stimulation.
The mechanism is the same at rack scale. A program fires tiny electrical pulses into the neurons. The neurons fire back with their own bioelectric signals, and, crucially, they physically restructure their synaptic connections in response. That rewiring is the wetware version of training a model. Where a GPU adjusts numerical weights in a neural network simulated in software, these neurons adjust actual physical connections between actual cells. It is slower to characterize, harder to control, and, its backers argue, radically cheaper to run once you count the electricity bill.
The neurons don’t simulate learning. They rewire themselves and call that the answer.
The real target isn’t intelligence, it’s the wattage
Every story about AI’s power appetite eventually lands on the same comparison, and it holds up here too: the human brain runs on approximately 20 watts, about what you’d get from a dim incandescent bulb, and it handles vision, language, and motor control simultaneously. A modern AI training cluster can draw megawatts. Hyperscale data centers now strain regional power grids badly enough that utilities in Virginia, Ireland, and Singapore itself have had to negotiate new terms with operators just to keep the lights on for everyone else. Singapore, a small island nation with limited land and no domestic energy reserves to speak of, has particular reasons to care about compute that doesn’t scale linearly with power draw.
That is the actual bet behind the NUS and Cortical Labs project. Not that living neurons will out-think a transformer model anytime soon. They will not, not on anything like the tasks GPT-class systems handle. The bet is narrower and more practical: that for certain classes of problems, biological neural tissue can compute at a small fraction of silicon’s energy cost, because it already does, inside every skull on the planet.
Twenty watts runs a human brain. A single AI training run can draw megawatts for weeks.
This is not the only lab growing computers instead of building them
Cortical Labs isn’t working in isolation, and the company’s approach isn’t the only biological bet on the table. FinalSpark, a Swiss startup, has run something called the Neuroplatform since 2024, a cloud service that lets researchers rent remote access to living brain organoids, lab-grown clusters of neural tissue sometimes called mini-brains, for roughly $500 a month, to experiment with biological data processing over the internet. In November 2025, researchers at Ohio State University took the concept somewhere stranger, using the electrical spiking patterns of shiitake and button mushrooms to build memristors, circuit components that retain a memory of past electrical states. None of these are competing to replace your laptop. They are competing to answer a much stranger question: which substrates, biological or fungal or otherwise, can hold and process a signal more efficiently than a transistor.
| Project | Substrate | Access model | Scale |
|---|---|---|---|
| NUS / Cortical Labs biological data center | Human neurons on silicon microelectrode chips | Institutional, on-site rack | 16 million neurons across 20 CL1 units |
| FinalSpark Neuroplatform | Cultured brain organoids | Cloud subscription, about $500/month | Not disclosed at rack scale |
| Ohio State mushroom memristors | Shiitake and button mushroom electrical spiking | Research prototype, component level | Single memristor components |
What “keeping it alive” actually costs you
Here is where I think the excitement needs a counterweight, and it is a short one, so I will state it plainly: living tissue is a maintenance liability that silicon is not. A GPU sits on a shelf indefinitely and works when you plug it back in. A rack of human neurons needs continuous nutrient perfusion, temperature control, sterility, and presumably a plan for what happens when a batch of cells degrades or dies, since neurons cultured outside a body do not last forever under current techniques. None of that appears in the wattage comparisons, and it should. The honest framing isn’t “biological computers use less power than silicon.” It’s “biological computers may use less power per computation, in exchange for a fundamentally different, currently unsolved set of operational costs.” Cortical Labs and NUS have not, per the available reporting, published numbers on the total energy or infrastructure overhead of keeping the rack alive versus what it saves computationally. That gap matters more than the neuron count.
What I keep coming back to is that this project isn’t really competing with Nvidia. It’s running a parallel experiment on a question data center operators have never had to ask before: what if the cheapest way to compute something is to grow the part that does the computing, and just remember to feed it. Singapore, of all places, land constrained and power constrained, is a plausible spot for that question to get asked seriously. Whether the answer holds up outside a demonstration rack, at the scale where it would actually dent a national power grid, is the experiment nobody has run yet.
Source: The Straits Times