Human Brain Cells Run New Data Centers in Singapore, Melbourne

Human Brain Cells in Data Centers: A Technical Deep Dive into Cortical Labs’ Biological Computing

Executive Summary
Cortical Labs is deploying the first small-scale biological data centers in Singapore and Melbourne that use living human neurons cultured on multi-electrode arrays (MEAs) as the core computational substrate. The CL1 “body in a box” system integrates lab-grown human brain cells with conventional silicon hardware to create a hybrid biological-silicon computing platform. Early prototypes demonstrate real-time spike-based processing with extremely low power consumption compared to GPUs, while a 30-unit biological neural network server stack is scheduled to come online in 2025. This represents an early but radical departure from transistor-based architectures toward wetware computing that may eventually challenge Nvidia-style silicon accelerators in specific low-power, adaptive workloads.

Technical Architecture
At the heart of Cortical Labs’ system is the DishBrain / CL1 architecture: human pluripotent stem-cell-derived neurons are seeded onto a high-density multi-electrode array (typically 4,096–16,384 electrodes) fabricated on a CMOS-compatible substrate. The neurons self-organize into a living 2D/3D culture that forms synaptic connections over days to weeks.

Input and output occur entirely through electrical stimulation and recording:

• Sensory or data inputs are encoded as precisely timed electrical pulses delivered through the MEA electrodes.
• The biological network processes information via natural spike-timing-dependent plasticity (STDP) and other Hebbian learning mechanisms.
• Output spikes are decoded back into digital signals for downstream silicon logic.

The CL1 hardware package includes:

• A sealed, environmentally controlled “body in a box” that maintains precise temperature, pH, nutrient flow, and waste removal for long-term cell viability (months).
• On-board FPGA or low-power microcontroller for real-time spike sorting, stimulation control, and interfacing with the host computer.
• Closed-loop feedback that allows the biological network to be trained in situ using reinforcement or supervised paradigms.

The company has reported significant improvements in “wetware” optimization: better media formulations, glial support cells, and genetic or pharmacological tuning to increase network stability and reduce spontaneous background activity. A single CL1 unit is described as containing hundreds of thousands of live human neurons.

The upcoming biological neural network server stack consists of 30 individual CL1-style units mounted in a rack, each communicating with a conventional orchestration layer. This stack is expected to be offered as “Brain-as-a-Service” via cloud APIs by late 2025, with four stacks targeted for commercial availability.

Performance Analysis
Cortical Labs has not yet published formal peer-reviewed benchmark numbers against modern GPUs or TPUs, which is expected given the pre-commercial stage. However, public demonstrations and technical briefings provide early indicators:

• Power consumption: A single CL1 unit operates in the milliwatt range for the biological component (the neurons themselves consume ~10–20 μW per mm² of culture). Including support electronics, total system power is still orders of magnitude lower than an equivalent GPU workload for certain adaptive tasks.
• Energy efficiency: Early DishBrain experiments showed the biological network could learn to play Pong in minutes using far less energy than reinforcement-learning agents on silicon. Claims suggest 1–3 orders of magnitude better energy efficiency per synaptic operation compared to current neuromorphic chips.
• Latency: Spike propagation and network dynamics operate on biological timescales (milliseconds), which is slower than digital clock cycles but acceptable for many edge or control applications.
• Scalability: A 30-unit stack represents roughly 10–30 million live neurons. While impressive for wetware, this is still tiny compared to the ~100 billion neurons in a human brain or the trillions of parameters in frontier LLMs.

Comparison with Nvidia and Neuromorphic Silicon

The biological system’s advantage lies in its native analog, low-precision, highly adaptive computation and extreme energy efficiency. Its disadvantages are obvious: limited lifespan of the culture, ethical and regulatory complexity, slower clock rates, and lack of mature software ecosystem.

Technical Implications
If successfully scaled, Cortical Labs’ approach could open an entirely new branch of computing architecture—“Synthetic Biological Intelligence” (SBI). Potential applications include:

• Ultra-low-power edge AI for sensors and robotics where continuous training in dynamic environments is required.
• Research platforms for studying brain-like computation at scale.
• Hybrid systems that use biological networks as co-processors for specific tasks (anomaly detection, temporal pattern recognition, adaptive control) while silicon handles deterministic high-speed math.

For the broader AI ecosystem, this introduces a biological substrate into the current silicon-dominated stack. Cloud providers may eventually offer “biological accelerators” alongside GPUs, with developers needing new abstractions for programming living neural networks (spike encoding schemes, training protocols that respect biological constraints).

It also accelerates convergence between neuroscience and machine learning. Techniques developed to train these wetware networks could feed back into more efficient spiking neural network (SNN) algorithms on digital neuromorphic hardware.

Limitations and Trade-offs
The technology faces severe practical constraints:

• Biological variability: Every culture is slightly different; batch-to-batch performance can vary significantly.
• Maintenance: Cells require continuous nutrient supply, temperature control, and sterile conditions. A power or fluidics failure can kill the “computer.”
• Speed: Millisecond-scale dynamics limit use in high-throughput training or inference compared to GHz silicon.
• Ethical & regulatory: Using human-derived neurons raises consent, moral, and biosafety questions. Data-center deployment will require new compliance frameworks.
• Scalability ceiling: Growing and sustaining billions of neurons in 3D cultures with adequate oxygen and nutrient delivery remains an unsolved engineering problem.

Expert Perspective
Cortical Labs’ work is one of the most credible demonstrations yet that living neural tissue can be turned into a stable, programmable computational substrate outside a research lab. While it is still many years from competing with Nvidia on raw FLOPS or serving frontier foundation models, it opens a genuinely new computational paradigm. The real significance may not be in replacing GPUs but in creating hybrid systems that exploit the complementary strengths of biological plasticity and silicon precision. The 2025–2026 timeline for cloud-accessible biological server stacks will be a critical test: if they can deliver reliable, remotely accessible wetware computing, it will mark the beginning of a new chapter in post-Moore computing.

Technical FAQ

### How does programming a biological neural network differ from training an SNN on Loihi or SpiNNaker?
Instead of setting weights in simulation, developers send electrical stimulation patterns and use reinforcement signals (e.g., dopamine-like reward pulses) to shape the living network’s synaptic strengths via biological plasticity. The “model” is the physical connectivity of the culture itself, requiring new toolchains for spike encoding, closed-loop training, and readout.

### What are the projected performance metrics for the 30-unit stack?
Specific FLOPS or synaptic-ops/second figures have not been disclosed. The primary metric emphasized is energy per operation and adaptability rather than peak throughput. Expect early cloud offerings to target niche low-power, always-on learning workloads rather than large-scale matrix multiplication.

### Is the CL1 API expected to be compatible with existing neuromorphic frameworks?
No public SDK details have been released. Integration will likely require new abstractions that treat the biological network as a black-box co-processor with spike I/O, rather than a drop-in replacement for PyTorch or Lava.

### How long can a deployed CL1 unit remain operational?
Current systems have demonstrated stable operation for several months. Long-term viability beyond 6–12 months in a commercial data-center environment has not yet been publicly validated.

References

• Cortical Labs technical briefings and CL1 launch materials (Mobile World Congress 2025)
• DishBrain foundational papers on STDP-based learning in vitro
• New Atlas, TechRadar, and ABC News technical coverage of the CL1 and server stack plans

Sources

• Bloomberg - Human Brain Cells Run New Data Centers in Singapore, Melbourne
• ABC News - Melbourne start-up launches 'biological computer' made of human brain cells
• New Atlas - World's first "Synthetic Biological Intelligence" runs on living human cells
• TechRadar - Brain-as-a-service set to start before 2026