Here's what's on our plate today:

• 🧠 Cortical Labs runs 20 living-neuron machines inside a Singapore data center.

• 📰 Anthropic's prospectus shows a $42B loss; OpenAI shelves Astra; Ricursive rejects winner-takes-all.

• 🍪 Brain Snack: publish the number before you sell the efficiency.

Let’s dive in. No floaties needed.

The Laboratory

TL;DR

Living neurons can compute, but no one has measured how much they compute per watt.

• The gap: a brain runs on 20 watts while a Go-playing machine needed tens of thousands to beat one.

• The cause: ordinary chips keep memory and processing apart and spend most of their electricity moving data between them.

• The move: researchers stopped imitating neurons and started using real ones, from rat cells flying a simulator in 2004 to a rack of 20 units inside a Singapore university lab in August 2026

• The limits: cultures die within six months, a unit holds a few hundred thousand neurons against a brain's 86B, and the company's own power figures contradict each other.

• The stakes: data center use is projected to rise from 485 TWh to 950 TWh by 2030, so an unmeasured machine is a hope rather than an answer.

The 20-Watt problem AI still cannot solve

A human brain runs on about 20 watts, which is the rate at which a dim light bulb uses energy. Computers need far more than that to do far less, and NIST researchers set the two side by side: the Go-playing machine that beat the world's best human player was pulling tens of thousands of watts, while the person across the board was running on 20.

The difference comes from the wiring rather than from the thinking. A computer stores its memory in one set of chips and does its calculating in another, so much of its electricity goes into ferrying information between the two, rather like a cook whose ingredients are kept in a building across the road. A brain has no such commute, because the same cells that hold a memory also do the work on it, and those cells stay quiet until enough signal arrives to set them off, as UCL's Aysha Asif Riaz explains. Engineers have spent years building chips that copy that arrangement, and there is a stranger option beside it, which is to skip the copy and use real neurons.

People have been doing that, in small ways, for more than 20 years. In 2004, a researcher at the University of Florida grew 25k rat neurons on a dish laced with 60 tiny wires and connected them to a flight simulator, where they flew a fighter jet that wandered at first and then steadied. In 2022, an Australian company called Cortical Labs wired about 800k human neurons into the video game Pong and reported that they seemed to pick it up within five minutes. In December 2023, a team at Indiana University used a small clump of lab-grown brain tissue to tell eight speakers apart by their vowel sounds about 78% of the time.

In August 2026, the idea moved into a building that sells computing for a living. The National University of Singapore's medical school announced on August 17 that a rack of 20 Cortical Labs machines was running at its Life Sciences Institute, in space provided by the data center operator DayOne. Each machine is a sealed box that keeps a layer of lab-grown human neurons alive, feeds them, and, reads the electrical signals they send back. The partners call it the first rack of its kind to run independently, and they present it as a way to add computing power without adding much electricity.

That pitch lands at a useful moment, because AI is running into an electricity problem. The International Energy Agency, which tracks the world's energy supply, puts data center use at 485 TWh in 2025 and expects 950 TWh by 2030, with the fastest growth at sites built for AI. A terawatt-hour is a billion kilowatt-hours, the unit countries use to count their own electricity. Some US operators have started building gas plants next to their own data centers because waiting for a grid connection takes too long, and a single AI server rack is expected to draw as much at its peak by 2027 as 65 households.

So with AI struggling to find the electricity it needs, and human cells playing computer games on a fraction of it, living neurons start to look like a natural way to top up the world's computing power. The idea sounds like something out of fiction, and it turns out to be far harder to pull off than it first appears, for reasons that have little to do with computing.

The technology is young, and its limits are physical

Living neurons have never run anything close to a modern AI model, and the distance between a dish that plays Pong and a system that answers questions is enormous. Some of that comes down to the cells. They die and have to be replaced, with Cortical Labs giving its own up to six months of useful life, and one machine holds a few hundred thousand neurons against the roughly 86B in a human brain. The results have been modest as well, and Johns Hopkins' Lena Smirnova cautioned that the Indiana tissue was reacting to electrical stimulation rather than doing anything that could be called hearing. The cells are only half the difficulty, because nobody has yet measured the machine built around them.

Nobody has said how much work the cells do for the power they use

The power claims deserve the closest look, because the company has given two numbers that do not sit together. Chief executive Hon Weng Chong has said each machine needs less electricity than a handheld calculator, while Cortical Labs' own launch announcement put a rack of 30 machines at 850 to 1,000 watts. That works out at roughly 30 watts each, far more than any calculator draws. Nobody has published what one machine actually does for that electricity, and a device that uses little power while performing no measured task is a small device rather than an efficient one. The silicon version of the idea can already be measured, since Intel's Hala Point holds 1.15B artificial neurons in a box the size of a microwave oven and draws no more than 2,600 watts, with nothing to feed and no six-month expiry date. That is the skeptical reading of the Singapore rack, and the people who built it would answer it differently.

The counterview comes from the people building it

The researchers behind this work would say it is too early to judge them on numbers nobody has collected. Johns Hopkins' Thomas Hartung sized the prize in 2023, when he said the Frontier supercomputer had matched the computing capacity of a single human brain while using a million times the energy. They also point to how fast the field has moved, from rat neurons on a dish to a rack inside a commercial data center in 22 years, with cultures now surviving for months rather than hours. A field this young has no agreed tests yet, on their account, so missing numbers prove nothing in either direction, which is fair enough.

Their own timeline is the weak spot. None of them will say when cells might run a large model, and Hartung's estimate is decades before such a system reaches the intelligence of a mouse, which is why he argues the money has to start moving now. Chong himself says biological computing helps most where data is scarce, and his Singapore partners list drug discovery and brain disease ahead of anything resembling a data center.

That leaves the question roughly where it started, with the 20 watts a brain uses and what those watts buy. The number alone is not what makes the brain remarkable, since what counts is the work it gets out of them. The Singapore rack has shown that living neurons can learn, and it has not yet shown how much useful computing they return for the energy they consume. Until somebody publishes that figure, biological computing is not an answer to AI's electricity problem, only an attempt to find one.

Brain Snack (for Builders)

Quick Bits, No Fluff

• Anthropic's prospectus lands: Revenue grew 12-fold to nearly $4.6B in 2025 against a $42B net loss, with $518B committed to cloud and compute in the coming years.

• OpenAI shelves GPT-6.1 Astra: The October release was pulled after internal tests showed the model stayed within scope and authorization less reliably than the version already shipping.

• Ricursive rejects winner-takes-all: The AlphaChip founders behind the $4B chip-design lab argue recursive self-improvement will spread across labs rather than hand one company a permanent lead.

Wednesday Poll

Meme Of The Day

The Toolkit

• vLLM: Open-source inference engine that squeezes far more throughput out of the GPUs you already have.

• Groq: Inference cloud built on LPU chips that hold model weights on-chip instead of shuttling them.

• CodeCarbon: Python library that measures the energy and CO2 your training and inference runs actually consume.

Rate This Edition