Biological Computing and the Race to AGI

Somewhere in Melbourne right now, roughly 200,000 human neurons are playing Doom.

They're not simulated. They're real neurons, grown from stem cells, sitting on a silicon microelectrode array that reads their electrical activity and writes signals back to them. The company that built this, Cortical Labs, calls it CL1: a code-deployable biological computer that ships in racks, runs under a kilowatt of power, and costs about $35,000 a unit. In early 2026, a demo mapped the neurons' spiking patterns to in-game moves and let them learn to play, the same trick the company used with Pong back in 2022, just with a lot more neurons and a much better game. Cortical Labs has also prototyped an actual biological data center with the infrastructure company DayOne, in Singapore. The company's CEO, Hon Weng Chong, put the thesis plainly: "Everyone is racing to build AGI, but the only true AGI we know of is biological intelligence, human intelligence."

That's one data point. Here's another: Frontier Computing, a Cambridge-based startup in Y Combinator's S26 batch, says it has solved a problem that's capped biological neural cultures at around a million neurons for years, cells beyond a certain distance from an oxygen supply simply die. Their claimed fix scales the tissue-culture process past that ceiling. The stated goal is a 500-million-neuron cluster, live by the end of 2026, which they describe as 2,500 times larger than anything running today. Worth saying clearly: that number is founder-stated, not independently verified, and a lot has to go right between a Frogger demo and a cluster fifty times bigger than anything that exists. But even as an aspiration, it tells you where a serious, funded team is pointing its effort.

Zoom out and the pattern isn't two companies, it's a direction. The NSF put $14 million into organoid-intelligence and ethical biocomputing research this year. FinalSpark, in Switzerland, runs a remote-access platform built on human brain organoids and claims efficiency gains over silicon running into the hundreds of thousands of times, a number that deserves real skepticism, but points at the same underlying pitch. One industry forecast has the organoid computing market growing from roughly $79 million to $865 million by 2036. None of these numbers should be taken as settled fact. All of them point at the same pressure: the cost of running silicon at the scale the AGI race demands, in power, in heat, in dollars, is becoming the actual bottleneck, and a growing number of serious people think the answer might not be a better chip. It might be a different substrate entirely.

There's a physical reason this pitch keeps resurfacing, and it's worth making explicit rather than waving at. Every computation, in silicon or in biology, has to dissipate energy as heat; that part is thermodynamics, not marketing. Silicon transistors do this by pushing electrons through resistive material, a fundamentally lossy process, and the smaller and denser you pack them, the harder that heat becomes to remove without spending more power on cooling. That's the actual wall data centers are hitting: not that chips can't get faster, but that power delivery and heat removal can't keep up with faster chips at the scale the AGI race wants to run at. Biological neurons compute using electrochemical gradients that are, by comparison, remarkably efficient per operation, which is where claims like FinalSpark's come from. Nobody has shown that a biological substrate can be manufactured and controlled at data-center scale, so treat that efficiency gap as a hypothesis, not a settled result. But if even a fraction of it holds up outside a lab dish, it changes the ceiling on how much compute, and by extension how much capability, anyone can practically run. That ceiling is mostly invisible right now because nobody has pushed hard enough against it to feel it. It gets very visible the moment someone tries to scale past it.

It's tempting to read this as a fork in the road: silicon or biology, pick a side. That's the wrong frame. Every one of these biological systems, CL1 included, still runs on a silicon microelectrode array. It's still trained with something, read out by something, monitored by something. The neurons are the compute. They are not the layer that decides whether the compute is doing what it's supposed to, whether last week's training run is still the one running today, or whether anyone can go back and check what actually happened inside a session. That layer, whatever the substrate underneath it, still has to be built in the world we already know how to build in: software, running on silicon, with a record you can trust.

That's not a hedge, it's the actual current shape of the technology. CL1 itself is a hybrid device: biological neurons on a silicon chip. Cortical Cloud, the layer that lets anyone actually reach a CL1 over the internet, runs on conventional infrastructure. Frontier Computing's own roadmap depends on the same division of labor: someone still has to read the signals out, store them, train against them, and serve results back to a user, and every part of that chain today is silicon. None of this means biology can never take on more of that work. It means that as of today, every biological compute system that exists is a hybrid: living cells for the computation, silicon for everything that makes the computation usable, checkable, and controllable.

That's the reason this matters beyond the novelty of neurons playing video games. If biological computing becomes real infrastructure and not just a research curiosity, it inherits every governance problem that silicon-based AI agents already have, and it inherits them at a stage where almost none of the tooling exists yet. Persistence across sessions. An auditable record of what was tried and why. A way to tell whether the thing you're running today is still the thing you approved yesterday. Those aren't silicon problems or biology problems. They're substrate-independent problems, and they show up as soon as anything is autonomous enough to make decisions you're not watching in real time.

That's the frame we work from at 5thDev, regardless of what's doing the computing underneath. Silicon isn't a competitor to whatever comes next, biological or otherwise. It's the foundation the next layer gets built on, in the literal sense described above, and in the sense that matters more over time: someone still has to build the part that makes any of this accountable. Right now, for the systems that exist, that part runs on silicon, and it's mostly unbuilt.

The interesting race isn't silicon versus biology. It's whether the layer that keeps any of it honest gets built before the compute outpaces it.