The Biological Data Center: A Rack of Living Human Neurons in Singapore

Twenty units, each holding roughly 200,000 lab-grown human neurons on a silicon chip, are now processing data inside the National University of Singapore's medical school. The rack went live in August. Cortical Labs, the Melbourne company that built it, calls the single rack the first independently operated biologically integrated server rack in the world.

Nobody involved claims the neurons compute faster than transistors. The bet is narrower and stranger: that biology can carry part of AI's workload for a fraction of the power.

Cortical Labs, founded in 2019, supplied the hardware. DayOne, a Singapore-headquartered data-center developer, supplied the facility and the capital. NUS Medicine grows the cells under neuroscientist Rickie Patani. The prototype is the first of its kind outside Australia, where a Melbourne facility already runs 120 of the same units.

What is actually in the rack

The CL1 is a self-contained biological computer. Stem-cell-derived human neurons are cultured onto a custom electrode array, then wrapped in pumps, gas mixing, temperature control and filtration. That life-support system keeps the culture alive for up to six months, which is roughly how long a research programme needs before it iterates.

Software talks to the cells through the company's Biological Intelligence Operating System, or biOS. It renders a simulated world, translates it into electrical signals, and reads the neurons' responses back into the loop. The cells are not programmed with instructions. They are placed in an environment and nudged with feedback until their firing patterns produce the behaviour a researcher wants.

The lineage is public. The CL1 descends from DishBrain, the experiment in which a culture of roughly 800,000 human and mouse neurons learned to play Pong. The same team later taught an array of CL1 chips to play Doom, a harder test of pattern recognition under a stream of pixels. Pong is why Cortical could raise money at all. The rack is what the money built.

The establishment of this prototype shifts the conversation from research to commercial application. Biological computing supplements AI in areas where data is sparse, learning from far less and adapting as conditions change.— Hon Weng Chong, Founder and CEO, Cortical Labs

The number Cortical does not publish

Everything here rests on power. The company says the CL1 uses "a fraction of energy other technologies use" and describes it as low power "to the extreme". The rack sits inside a country that has tied further data-center growth to efficiency targets under its Green Data Center Roadmap, and the operator frames the prototype as a way to scale AI capacity at "significantly lower power intensity".

What no one has published is a watts-per-inference figure. Efficiency in this field is measured against a moving target — a modern GPU running a quantised model — and no side-by-side benchmark of neurons against silicon on the same task has been released. That gap is the honest centre of the story. Without a number, "a fraction of the power" is a design intent, not a result.

The comparison problem is real on the silicon side too. Data-center operators quote power in megawatts and efficiency in power-usage effectiveness, both of which assume predictable, always-on workloads. A rack whose compute substrate has to be fed, kept warm and replaced every six months does not fit that accounting model yet. Part of the operator's interest is learning how to write the meter for a substrate that decays.

A second constraint is subtler. Biological computing is not a general substitute for silicon. On arithmetic, memory and orchestration, transistors win by orders of magnitude and will keep winning. The interesting cases are the ones AI handles badly — sparse data, shifting conditions, problems where there is no large training set to learn from. Cortical names drug discovery, humanoid robotics, cybersecurity and fraud detection as the places to look.

From Pong to a server rack

The arc matters, because it explains why a research curiosity now has a procurement category. Cortical spent three years proving that neurons could be steered. The last year has been about packaging them into something a facility manager can install and power.

BIOLOGICAL COMPUTING — FROM DISH TO RACK
─────────────────────────────────────────────────────────────
 2019 ────── 2022 ────── 2025 ────── ◉ 2026 ────── 2027
 founded    DishBrain    CL1 ships    NUS rack    DayOne SG1
 Melbourne  800k neurons + 120 units  live        operational
            learn Pong   in Melbourne

The wiring-diagram question is older than the rack. As we wrote in October, a fruit fly's full connectome split the machine-learning field over whether biology's structure is a blueprint worth copying. This rack runs the same question from the other side: instead of abstracting a brain into a diagram, it keeps the brain alive and watches what it does.

What is new is the surround. A neuron culture by itself is a laboratory artefact. Put it in a rack with defined electrical envelopes, cooling and monitoring, inside a commercial facility, and it becomes an asset with a power budget and an uptime expectation. DayOne's plan moves the prototype from the NUS lab into one of its own commercial Singapore data centers for a live deployment phase, where it will be tested under real load.

Who pays, and why Singapore

DayOne is the reason this happened in Singapore rather than Melbourne. The company operates digital infrastructure across nine markets in Asia Pacific and Europe and says it has secured about 2.1 GW of bookings since inception. It has committed capital and engineering, not just floor space.

Cortical's investor list explains the strategic pull. Backers include Horizons Ventures, Blackbird Ventures, LifeX Ventures, Radar Ventures and In-Q-Tel, the venture arm of the CIA. The company has raised in the low tens of millions of dollars, a rounding error next to a single hyperscale build. National-security interest in a Pong-playing neuron culture is its own signal: the same sparse-data problem appears in signals intelligence and threat detection.

Singapore supplies the third ingredient. The city-state is expanding AI capacity under tightening energy and environmental guardrails, which makes a lower-power compute substrate worth a prototype even if it never scales. For the operator, the biological rack is a hedge against the risk that power, not silicon, becomes the binding constraint on the data-center business.

A field that is already commercial — barely

Cortical Labs is not alone, and that changes how the rack should be read. Switzerland's FinalSpark has run a remote platform for three years, renting access to brain organoids on electrodes through a Python API; its organoids carry around 10,000 neurons across eight electrodes. Both companies have moved from demonstrations to paid access, and both sell the same abstraction: a biological substrate you reach over a network.

The difference is scale and packaging. FinalSpark's customers are mostly researchers probing what the system does. Cortical's CL1 wraps a full life-support and compute unit behind a purchase order, and Singapore is the first time those units are operated by someone other than the vendor, inside infrastructure built for commercial workloads. Twenty units is small. The point of the rack is not throughput; it is proving the unit of deployment.

That is why the missing benchmark matters more than it looks. A research platform can run on curiosity. A commercial one runs on economics, and economics wants a denominator — joules per task, cents per result, a figure that can sit next to a GPU cluster on a spreadsheet. Two companies building the same category, both selling access rather than silicon-class performance, is a sign the field understands what it has and what it does not.

There is also a naming problem that marks how early this is. Cortical and its peers cannot agree on what to call the thing they are building. Papers in the space offer competing terms — synthetic biological intelligence, organoid intelligence, wetware computing — and Cortical has published a call for a shared nomenclature. A field that cannot name itself is not yet a market.

One more caveat belongs in the record. Whether these cultures "learn" in any sense a neuroscientist would defend is still contested. The published work shows cultures adapting their firing to reduce a feedback signal, a form of conditioning rather than reasoning. Cortical's own researchers have written about the moral questions raised by embodied neural systems, and the same work that makes the platform useful is the work that makes its edges hard to draw.

The part that is actually being tested

Read the partnership language carefully and the real experiment is operational. The prototype answers whether living neurons can sit inside a data center's power, cooling and monitoring envelope without becoming a full-time biology project. NUS Medicine supplies the neurobiology; DayOne supplies the infrastructure discipline; Cortical supplies the hardware.

If it works, the model replicates. Chong has said Singapore is the template the company intends to scale globally. If it does not, the result will look less like a failed computer and more like a failed facility: biology that could not meet an uptime contract.

The near-term signal to watch is boring: whether the next deployment happens inside a commercial data center, on commercial power terms, with a benchmark attached. That would turn a press release into a category. Until then, a rack of living neurons is a serious experiment wearing a data-center badge.