Cortical Labs became widely known in 2022 for DishBrain, described in a paper in the journal Neuron, in which a culture of about 800,000 human and mouse neurons grown on a CMOS electrode chip learned to play the video game Pong inside a closed feedback loop. The neurons were not programmed to play. They were placed in an electrical environment that gave predictable feedback for one behaviour and unpredictable feedback for another, and the culture reorganised itself toward the predictable outcome. The CL1, launched commercially in March 2025 at roughly US$35,000 a unit, packaged that arrangement as a product: living neurons on an electrode array inside a sealed housing that supplies its own life support, with a software interface so that code can be deployed to it without an attached conventional computer. What Singapore has added is scale and independent operation. Twenty CL1 units racked together and run by a commercial data centre operator rather than by the manufacturer's own laboratory is the step the partners are calling a world first - not the biology, which already existed, but its deployment as data centre infrastructure.

Conventional computing separates memory from processing and moves data between them, which is where much of the energy goes. A neuronal culture does not make that separation: the same connections that hold the state also do the computing, and they adapt as signals pass through, which is the property that makes such systems interesting for learning tasks. The term used for hardware that fuses living material with electronics is wetware, distinguishing it from both hardware and software. The costs are equally distinctive. Silicon needs power and cooling; neurons need to be kept alive. Each CL1 runs pumps, a gas-mixing system, temperature control and filtration continuously, and the culture has to be replaced when its roughly six-month viability window closes - a maintenance obligation no silicon rack carries. The efficiency claim should be read carefully. The full rack's draw of roughly 800 to 1,000 watts is modest against an AI server rack, but the partners have released no benchmark showing what work the rack completes per watt, so there is as yet no like-for-like comparison with a conventional processor.

The interest here is less in the neurons than in where they have been put. Singapore is land- and power-constrained and has run a cautious data centre policy for years; its Green Data Centre Roadmap of May 2024 set a power-usage-effectiveness target of 1.3 or lower at full IT load and offered new capacity - at least 300 megawatts in the near term, with a further 200 megawatts available to operators deploying green energy technology - on condition of efficiency. A jurisdiction that rations capacity by efficiency is exactly where a low-power computing substrate gets a hearing, which is why the choice of Singapore is part of the story. The wider context is the energy bill of artificial intelligence. Training and serving large models has made data centre electricity demand a policy question in several countries, and biological computing is one of a set of alternatives - alongside neuromorphic and optical computing - being explored on the argument that the brain performs remarkable computation on about twenty watts. Whether neurons can be made to do commercially useful work at that efficiency is unproven, and the absence of any published workload benchmark is the honest measure of how early this is. For an examination, the reliable content is the terminology, the partners and the numbers rather than the promise.