Silicon neurons and living ones
From the papers and the company page you'll find in the sources below.
Side by side
| Neuromorphic chip (IBM TrueNorth, 2014) | Wetware (Cortical Labs CL1) | Human brain | |
|---|---|---|---|
| What does the computing | Silicon circuits that imitate spiking neurons | Living human neurons on a chip | Living neurons |
| Neurons | 1 million, programmable | About 800,000 | About 86 billion, by the usual estimate |
| Connections | 256 million configurable synapses | Whatever the cells grow | Whatever the brain grows |
| Power | 63 milliwatts on live video | No published figure I could find | About 20 watts |
| How long it lasts | As long as the chip works | Up to six months, says the company | A lifetime |
I'd read the table as two different answers to the same question, which is how you get some of the brain's efficiency into a computer. Neuromorphic engineers copy the brain's layout in silicon. Wetware people skip the copying and use the cells themselves, and you'll see each choice brings its own problems.
What a neuromorphic chip actually is
If you want one example, I'd pick IBM's TrueNorth, which was described in Science in 2014. I was surprised how big it is. There are 5.4 billion transistors on the chip, split into 4,096 cores, and between them they hold 1 million spiking neurons you can program (plus 256 million synapses). You can also tile the chips together if you need more. The team fed it video at 400 by 240 pixels and 30 frames a second, it picked out and classified several objects as they went past, and I think the striking number is that it only drew 63 milliwatts doing that.
What I'd stress is that every one of those neurons is a circuit you've programmed. You're able to copy the chip, reset it and run the same network again tomorrow, and you can't do that with cells.
Where wetware is different
With living neurons you don't get to program the connections at all, since the cells grow their own. I'd say you shape what they do by stimulating them through an electrode array and giving them feedback, which is what the DishBrain team did with Pong. The best outside result I've seen on a CL1 is a July 2026 paper where one culture sorted images into two classes with accuracy up to 95.6%, and I think you'd agree that's a much smaller task than the video demo.
Another way round the programming problem is to leave the cells untrained and only train a readout on top, which is what Brainoware did with a brain organoid in 2023 (the method is called reservoir computing).
The cells also die. FinalSpark's organoids last up to about 100 days and Cortical Labs quotes six months for its neurons, so anything a culture learns is lost when it's replaced (the guide on how long brain organoids live has the figures).
Which one is further along
As far as I can tell neuromorphic chips are, by a long way, because they already run real tasks at low power and they're made in a chip factory like any other silicon. I think wetware's case rests on the brain itself, which runs on about 20 watts (the brain versus GPU guide sets that against the 21 megawatts of the Frontier supercomputer, with the caveats). Whether living cells can deliver that efficiency in a machine is still an open question, and I wouldn't claim either side has won it.
Questions
What is the difference between neuromorphic computing and wetware?
I'd put it simply. A neuromorphic chip is silicon that imitates spiking neurons and you program it, whereas wetware uses living neurons that grow their own connections, so you're training them with stimulation and feedback.
How much power does a neuromorphic chip use?
It depends on the chip, but IBM's TrueNorth drew 63 milliwatts while it handled 400 by 240 pixel video at 30 frames a second (that's from its 2014 paper in Science).
Is wetware better than neuromorphic chips?
I don't think so, not yet. You can copy and reset the chips and they're already running real tasks, whereas living cultures die within months and so far they've only handled much smaller jobs.
Sources
- Merolla et al., A million spiking-neuron integrated circuit with a scalable communication network and interface, Science, 2014
- Cortical Labs, CL1 product page
- Schottlender et al., Evaluating Encoding Strategies for Closed-Loop Classification in Biological Neural Networks, arXiv, 2026
- Kagan et al., In vitro neurons learn and exhibit sentience when embodied in a simulated game-world, Neuron, 2022
- Jordan et al., Open and remotely accessible Neuroplatform for research in wetware computing, Frontiers in Artificial Intelligence, 2024
- Eighty-six billion and counting: do we know the number of neurons in the human brain?
- Balasubramanian, Brain power, PNAS, 2021