How Singapore’s 20-Unit Biological Data Centre Uses Living Neurons

How the CL1 computers process data

Each CL1 contains at least 200,000 lab-grown human neurons. The cells are derived from blood cells that have been reprogrammed into stem cells, then cultivated across a silicon chip fitted with electrodes.

Those electrodes create a two-way interface:

1. Conventional computers translate information into electrical stimulation for the neural network.
2. The neurons respond through patterns of electrical activity.
3. The electrode array measures that activity and sends it back to software as an output.

This closed feedback loop lets the living network interact with a simulated environment and adjust its responses. The neurons are therefore not merely being used as biological decoration; they form the adaptive part of the computational system, while silicon provides the interface and control layer.

Why the life-support system matters

A CL1 is not a standalone silicon server. The neurons must remain alive and in controlled conditions.

Technicians replenish a nutrient mixture containing sugar, micronutrients and pH buffers every three days. A gas-mixing system supplies carbon dioxide, oxygen and nitrogen, while the wider system manages the conditions needed for the cell culture.

That makes the technology a different kind of infrastructure rather than infrastructure-free computing. It may reduce some electrical and cooling demands, but it introduces laboratory consumables, biological maintenance, controlled environments and the need for skilled operators.

Where biological computing could be useful

Cortical Labs argues that biological neural networks may learn and adapt from relatively few examples and respond more flexibly when circumstances change. That could make them relevant to tasks where collecting large, carefully labelled datasets is difficult.

Potential applications discussed by the company and its partners include:

• Humanoid robotics: Robots working in homes, offices and public spaces encounter unpredictable situations that are difficult to represent in advance.
• Cybersecurity: Biological systems could be tested for detecting unusual activity when there are not enough labelled examples of every possible attack.
• Adaptive research computing: Researchers can study how living neural networks respond to controlled inputs and feedback.
• Robotics and gaming experiments: Cortical Labs’ Melbourne facility reportedly has corporate-research and university customers exploring these kinds of uses.

These remain proposed or experimental applications. The available reporting does not establish that biological computers have already outperformed silicon systems in production robotics or cybersecurity workloads. Independent, workload-specific benchmarks will be needed.

The energy case—and its limits

Cortical Labs says an individual CL1, including its life-support equipment, uses about 30 watts and does not require conventional server cooling. The company has contrasted that figure with an Nvidia H100 SXM rated for up to 700 watts and an eight-H100 server reported to consume about 10,200 watts with supporting hardware.

Those figures suggest why the idea is attracting attention in Singapore. Data centres accounted for 7% of the country’s national electricity use in 2020, and electricity and water constraints have been part of the country’s debate over expanding data-centre capacity.

But a low device power rating is not the same as a proven lower total cost or environmental footprint. The public information available here does not provide a definitive water footprint, lifecycle assessment or workload-for-workload comparison with equivalent GPU infrastructure. Nutrient media, gas control, cell production, laboratory equipment and human maintenance all belong in a full accounting.

Biological computers will not replace GPUs

The strongest case for CL1 is specialised adaptive processing, not general-purpose high-performance computing. Cortical Labs’ chief executive has acknowledged that silicon remains substantially better for the fast, precise and repeatable calculations used by large language models and similar workloads.

The practical model is therefore complementary:

• silicon processors handle conventional high-throughput computation;
• biological neural networks are tested on tasks that may benefit from adaptation and sparse learning data;
• software and electronics connect the two systems.

That distinction matters. Calling the installation a biological data centre does not mean that mainstream cloud servers are about to be replaced by racks of living cells.

Capacity, commercial use and pricing

The Singapore project begins with 20 CL1 units for joint internal research and development. The partners are using the prototype to build operational expertise and examine the manufacturing and support processes required for a larger deployment.

Subject to technical validation, biosafety requirements and regulatory approvals, DayOne and Cortical Labs are exploring a phased expansion to as many as 1,000 units in a Singapore commercial facility.

Cortical Labs’ Melbourne site operates 120 CL1s and reportedly serves about 20 corporate-research and university customers. Remote access to one CL1 has been reported at US$2,200 per month, compared with a reported US$4,300 monthly rental for a high-end AI chip on major cloud services.

The remote-access price is a customer fee, not the operating cost of the biological computer or the NUS facility. The partners have not publicly disclosed the full cost of maintaining the Singapore installation.

The main limitations

Biological computing faces several unresolved challenges:

• Maintenance: Neurons require continuous nutrient, gas and environmental support.
• Consistency: Living systems may be less deterministic and repeatable than digital hardware.
• Scaling: Expanding from a 20-unit research rack to 1,000 units would require reproducible cell manufacturing and reliable operations.
• Regulation and biosafety: Larger commercial deployment depends on approvals as well as technical performance.
• Benchmarking: Claims about learning efficiency, energy use and adaptability still need independent, real-world testing.
• Environmental accounting: Lower electrical demand does not by itself establish a lower total water or lifecycle footprint.

These constraints explain why the Singapore installation is being treated as a staged validation project rather than a finished replacement for a conventional server farm.

What the project means for Singapore

Singapore’s biological data centre is best understood as a strategic experiment. If the claimed ability to learn from limited data and adapt to changing conditions survives independent testing, wetware computing could give the country a niche way to add selected AI capacity without scaling cooling and electricity demand in the same way as GPU-heavy infrastructure.

For now, the evidence supports a narrower conclusion: Singapore has created a small, living-neuron computing prototype that combines biological processing with silicon control systems. Its importance lies in testing whether that unusual architecture can become useful, reliable and commercially scalable—not in demonstrating that biological servers can displace mainstream data centres.