Singapore's Biological Data Center Has 16 Million Living Neurons. It Would Need 5,375 Racks to Match One Brain.

Sixteen million. That is the number of living human neurons now processing information inside a prototype data center at the National University of Singapore's Life Sciences Institute. The facility, unveiled on August 17 by NUS Medicine, Singapore-based data center operator DayOne, and Melbourne startup Cortical Labs, represents the world's first independently operated biological server rack. Twenty CL1 biological computing units, each containing approximately 800,000 lab-grown cortical neurons on silicon-based microelectrode arrays, went live on July 16 and drew more than 80 guests to a demonstration on August 6.

Energy numbers alone could sell the concept. All 20 units together consume between 800 and 1,000 watts. A single NVIDIA H100 GPU draws 700 watts under full load. Cortical Labs CEO Hon Weng Chong frames the pitch in sustainability terms: "Biological computing supplements AI in areas where data is sparse, learning from far less and adapting as conditions change." DayOne, which achieved a $20 billion valuation after raising $4.5 billion in June, is already considering scaling to 1,000 units.

Before anyone writes the eulogy for silicon, though, the math deserves a harder look. Much harder.

What 16 Million Neurons Actually Do

Each CL1 unit houses 800,000 human cortical neurons derived from induced pluripotent stem cells, sitting on a planar array of 59 electrodes. The neurons fire, form connections, and respond to electrical stimulation through Cortical Labs' proprietary biOS operating system, which creates real-time feedback loops at sub-millisecond latency. That speed was itself a significant engineering achievement over the earlier DishBrain prototype, the one that first demonstrated neurons playing Pong and was published in the journal Neuron. Researchers can deploy code directly to the neurons via a Python API, and independent researcher Sean Cole taught roughly 200,000 cortical cells to play Freedoom in a week using the same interface.

Playing a 1993 first-person shooter is genuinely impressive for a clump of human tissue on a chip, demonstrating adaptive, self-organizing computation that no silicon system achieves the same way. But it is also nowhere close to useful computation at commercial scale. Steve Furber, professor of computer engineering at the University of Manchester, put it plainly to New Scientist: "It's a very big step from a small network playing a computer game to a large language model."

The Power Efficiency Illusion

Here is the headline comparison: 1,000 watts for 16 million neurons versus 700 watts for one GPU. Seductive. But watts per neuron tells a story that Cortical Labs would rather you not calculate.

A CL1 unit draws 25 to 30 watts, including the life-support system that maintains precise gas mixtures, temperature, and fluid balance for the living cells. That works out to 37.5 microwatts per neuron. A human brain runs 86 billion neurons on roughly 20 watts total, or about 0.23 nanowatts per neuron. CL1 neurons are approximately 163,000 times less power-efficient than the organ they are modeled on, because the overwhelming majority of the CL1's power budget goes not to computation but to keeping cells alive: maintaining incubation temperature, circulating nutrient media, monitoring gas composition.

This gap matters enormously, because the efficiency argument is the primary business case for biological computing at scale, the entire reason a data center operator would buy into a technology that can barely play Doom. If the neurons themselves are not where the power goes, then adding more neurons does not automatically deliver the efficiency gains that justify the investment.

The 1,000-Unit Scaling Math

DayOne's reported interest in a 1,000-unit deployment makes the economics concrete enough to test.

Metric	1,000 CL1 Units	Equivalent Power: 43 H100 GPUs
Hardware cost	$20M (at $20K/unit rack price)	~$1.3M (at ~$30K/GPU)
Power draw	30 kW	30 kW
Annual electricity ($0.10/kWh)	$26,000	$26,000
Computational units	800M neurons (0.93% of a brain)	~3.4 trillion transistors
Demonstrated capability	Playing Doom, drug response modeling	Inference on 100B+ parameter models
Annual bio-operating cost	$5M–$10M (cells, nutrients, techs)	$0

One thousand CL1 units at the rack-volume price of $20,000 each would cost $20 million in hardware. They would contain 800 million neurons and draw approximately 30 kilowatts. That same 30 kilowatts could power roughly 43 NVIDIA H100 GPUs, a cluster capable of running inference on models with hundreds of billions of parameters and training substantial neural networks from scratch. The 800 million neurons represent 0.93 percent of a human brain's neuron count.

But the hardware sticker price understates the true cost by an order of magnitude. CL1 neurons remain viable for up to six months before requiring replacement, which means 1,000 units need 2,000 fresh batches of differentiated cortical neurons per year, each requiring induced pluripotent stem cell culture, differentiation over 30 to 90 days, and quality validation. At a conservatively estimated $1,000 per batch at scale, annual cell replacement alone runs $2 million. Add nutrient media replenished every three days, lab technicians, gas mixture maintenance, and quality control, and realistic biological operating expenses likely land between $5 million and $10 million annually for 1,000 units. By contrast, 43 H100s running on the same power draw cost about $26,000 per year in electricity at $0.10 per kilowatt-hour. That is a 200-to-1 operating cost ratio.

The Real Bottleneck: 59 Electrodes for 800,000 Neurons

Cost and power, though, are not even the deepest constraint. Bandwidth is.

Each CL1 reads from and writes to its 800,000 neurons through 59 electrodes. That ratio means the system can directly observe or stimulate 0.007 percent of its own neural network at any given moment. Every remaining neuron is doing something: forming connections, propagating signals, reorganizing in ways that might be computationally profound. The silicon interface just has no way to know.

Scale this to brain parity and the bandwidth deficit becomes surreal. Matching the human brain's 86 billion neurons would require 107,500 CL1 units at a cost of $2.15 billion in hardware, drawing 3.2 megawatts across 5,375 server racks. Those 107,500 units would provide a combined 6.3 million electrode channels. The human brain has approximately 100 trillion synaptic connections. The electrodes would be observing 0.000006 percent of the network's communication traffic. That is not a computing system you can program. That is a stethoscope pressed against a stadium full of conversations, picking up murmurs from six seats near one exit.

Princeton University's 3D-MIND platform, published in May 2026, embedded a flexible three-dimensional electronic mesh directly inside lab-grown neural networks of 70,000 neurons, achieving stable interaction tracking over six months with dozens of electrodes throughout the tissue volume rather than only at the surface. Promising direction. But even 3D-MIND's embedded electrodes serve thousands of neurons, not millions, and no commercial product yet delivers that improvement at anything approaching CL1 scale.

Where Biological Computing Might Actually Win

If silicon beats biology on every efficiency metric, why is a $20 billion data center company investing?

Because the GPU comparison is, in one critical sense, a category error. Biological computing is not trying to multiply matrices faster. It is creating a platform where living human tissue responds to drugs, develops pathological firing patterns, and models neurological disease in a closed-loop experimental environment that has no silicon equivalent whatsoever. Professor Rickie Patani of NUS Medicine described the dual purpose: "We're creating a platform that can help us understand learning and adaptation at their biological source. It gives us a real route to accelerate drug discovery and neurological disease research."

Early work on the CL1 successfully simulated epileptic-like activity in a neural culture and then demonstrated that antiepileptic compounds could restore normal learning behavior. Running pharmacological interventions on a living neural network that can be electrically monitored and stimulated in real time? No GPU replicates that regardless of its TFLOPS rating. A pharmaceutical company spending $50 million to $300 million on a Phase III clinical trial might find genuine, immediate value in a $20 million biological computing platform that narrows the drug candidate pipeline before human trials begin.

Limitations

This analysis compares CL1 neurons and H100 transistors on metrics like watts and cost per unit, but these are fundamentally different computational elements performing fundamentally different operations, and no benchmark exists to compare their useful work output because the useful work barely overlaps. Cortical Labs has not published detailed performance benchmarks for the CL1 at data center scale. The iPSC differentiation cost estimate of $1,000 per batch at scale is an extrapolation from academic literature, not a disclosed figure from the company. NUS's deployment is a research prototype, not a commercial operation, and no peer-reviewed paper has yet been published on its data center results. Interface bandwidth is improving: Princeton's 3D-MIND and similar three-dimensional mesh technologies could significantly change the electrode-to-neuron ratio over the next several years, though no commercial product delivers that improvement at CL1 scale today.

The Bottom Line

Singapore has plugged in the world's first rack of living brain cells and called it a data center. The power numbers are real: 16 million neurons running on less energy than a single GPU. But the bottleneck in biological computing was never watts. It was always the interface, 59 electrodes trying to read the electrical chatter of 800,000 neurons simultaneously, an observation problem masquerading as an efficiency breakthrough. Until electrode density catches up to neural density by several orders of magnitude, biological data centers will remain extraordinary research platforms for drug discovery and neuroscience and thoroughly impractical alternatives to silicon for any task a GPU already handles. If you are evaluating biocomputing for your organization, the honest question is not "Is it more efficient?" but "Do I need to run experiments on living neurons?" If yes, the CL1 at $35,000 standalone or the Cortical Cloud at $300 per week is the only game in town. If no, your H100 budget is safe.