Start-up is building the first data centre to use human brain cells

Data centres use huge amounts of energy and chips are in high demand – could brain cells be the answer? Australia-based start-up Cortical Labs has announced it is building two “biological” data centres in Melbourne and Singapore, stacked with the same neuron-filled chips that it has demonstrated can play Pong or Doom.

Cortical Labs is one of a few companies developing biological computers, consisting of neuronal cells wired up to microelectrode arrays that can stimulate and measure the response of cells when fed data. Earlier this month, the firm demonstrated that its flagship computer, the CL1, could learn to play the game Doom in a week.

Now, Cortical Labs has revealed two data centres that it plans to build. The first, in Melbourne, will contain around 120 CL1 units. The second, which is being built in collaboration with the National University of Singapore, will house 20 CL1s initially, but the company hopes it will eventually contain 1000 units in a larger data centre, after regulatory approval. Cortical Labs says this will allow it to expand its cloud-based brain-computing service.

Biological computers like the CL1 are being built and tested by research groups around the world, but they are often hard to build and not easy for others to use, says Michael Barros at the University of Essex, UK. “We spend a lot of money and sweat to build these [systems].”

“What [Cortical Labs] is doing is essentially allowing its biocomputer to be accessible at a large scale,” says Barros, who already uses cloud services from Cortical Labs as part of his research. “They’ll be the first ones to do that.”

Though these systems can be trained for relatively simple tasks, like playing Doom, the exact way in which these neurons function and how best to train them to perform tasks like machine learning is still unclear, says Reinhold Scherer, also at the University of Essex. “Having access to this allows you to explore learning, training and programming,” he says. “You don’t program neurons like standard computers.”

Cortical Labs argues that its data centres will also require far less power than typical computing systems, claiming that each CL1 needs around 30 watts, rather than the thousands of watts that a state-of-the-art conventional AI chip requires.

“When we scale up and have these as whole rooms, just like you have now with data servers, then there could be huge power savings,” says Paul Roach at Loughborough University, UK. There are other resources that biological data centres might need, such as nutrients to feed and keep alive the neuronal chips, but it should require far less cooling than conventional computing, he says. “The amount of energy that’s saved with [Cortical Labs’s] figures is fairly conservative.”

However, the technology is still at an early stage, says Tjeerd olde Scheper at Oxford Brookes University, UK, who has worked with a competing biological computing company, FinalSpark. “Is it going to work as people might think? No, we’re still in the early days of this development.”

It is hard to do a direct size comparison, as CL1 chips can’t do conventional calculations like a regular silicon-based AI chip can, but the proposed biological data centre will have hundreds of biological chips, compared with hundreds of thousands of graphics processing units (GPUs) seen in the largest AI data centres.

“I think it is a very long way from production-ready. It’s a very big step from a small network playing a computer game to a large language model, ” says Steve Furber at the University of Manchester, UK.

One of the remaining issues is that it is still unclear how to store the results of training these neurons in a form of memory, or how to run actual computational algorithms on them, rather than training them for specific uses like playing video games.

Another challenge is how to retrain the neurons once they have completed a particular task. “Whatever they are trained on is lost when the culture ends its life, so there needs to be a proper retraining,” says Scherer. “Then it’s not an optimal solution to keep a technology going if you need to retrain every 30 days.”