The Computer You Grow

#Design. Grow. Compute.

In my last piece I argued that the next decade of chip progress will come from materials and heat as much as from lithography, that some people are wrapping silicon in diamond to push it further. This is a story about the people going the other way entirely. Not making silicon better. Growing the computer out of a mushroom.

On June 10, 2026, Scientific Reports published a paper with a quietly absurd title: "Morphologically Tunable Mycelium Chips for Physical Reservoir Computing." The authors, a team out of Ecovative led by Orkan Telhan and including the company's co-founder Eben Bayer, describe a working analog AI chip that you do not fabricate. You farm it. Their numbers, printed on the paper's own figure: more than 3 million chips per harvest, stable after three months, under a dollar each. Their workflow is three words. Design. Grow. Compute.

It sounds like a stunt. It is not. It is the most visible surfacing of a strange, patient research lineage that has been quietly growing for years, and it is worth understanding properly, because both the promise and the limits are more interesting than the headline.

#What It Actually Is

First, kill the obvious misreading. This is not a living mushroom sitting there thinking. The chip is dead. The team grows fungal mycelium, the root-like network of hyphae, into a target structure, then dries it and wicks a conductive polymer called PEDOT:PSS into the tissue. What you end up with is a non-living, biodegradable slab of fungal material laced with conductivity, shaped into resistors, capacitors, and the nonlinear elements a circuit needs. The mushroom is the scaffold. The computation rides on the electrical behavior of that scaffold.

"Morphologically tunable" is the clever part, and it is not marketing. The way the fungus grows, how densely it branches, how its network tangles, directly changes how charge moves through the dried chip and how much memory it holds. So you do not program this thing by writing code. You program it by changing how the mushroom grows. The structure is the software. That is a genuinely different idea about what a chip is.

The honest framing, which the authors themselves use, is that this is a single prototype, a proof of concept. They benchmarked it on NARMA-10, a standard nonlinear time-series task, and showed it has the right ingredients: nonlinearity, temporal dynamics, the ability to separate inputs. They also concede, in plain language, that they are accepting performance trade-offs in exchange for sustainability. They are not claiming to beat silicon. They are claiming to compute at all, for almost nothing, out of something that composts.

#Why a Mushroom Can Compute at All

The trick that makes this possible is called reservoir computing, and once it clicks, the whole field opens up.

In a normal neural network you train every layer, adjusting millions of weights through backpropagation, which is what makes training so expensive. Reservoir computing cheats. You take any complex, nonlinear physical system with a bit of memory, the reservoir, and you pour your input signal into it. The reservoir's own messy physics smears that input out across a high-dimensional space of states. Then you train exactly one thing: a single linear readout layer that learns to interpret those states. The hard part, the nonlinear mixing, is done for free by the material itself.

This is why the reservoir can be almost anything with the right properties. People have built physical reservoirs out of photonics, spintronics, buckets of water, and now mushrooms. Any medium with nonlinear dynamics and fading memory can do the work in its physics rather than in math. And computing in matter instead of digital logic is the whole energy argument. One in-materio analog reservoir reported energy use around a hundred times lower than conventional CPUs, FPGAs, and GPUs for the same job. When the computation happens in the physics, you stop paying the digital tax.

#This Has Deep Roots

The mushroom chip did not appear from nowhere. It is the commercial, scalable edge of a field that academics have been tending for a decade, and the back story is half the fun.

The patron saint here is Andrew Adamatzky at the Unconventional Computing Laboratory in Bristol. His 2018 paper "Towards a fungal computer" laid out the idea that fungi process information as spikes of electrical activity, the same basic currency as neurons. In 2022 his group published "Language of fungi derived from their electrical spiking activity," which found that the electrical spikes of four mushroom species cluster into patterns resembling words, with vocabularies up to around fifty, and length distributions statistically similar to human language. The EU funded a 2.5 million pound project called FUNGAR to grow buildings out of computing fungus. A separate team at Ohio State built working memristors out of shiitake mycelium that switch up to nearly 5,900 times a second.

And behind all of it sits the oldest unconventional-computing party trick of all: slime mold, Physarum polycephalum, which solves a maze in a single pass and will lay out a near-optimal map of a country's motorways if you bait it with oat flakes where the cities go. Biology has been doing this kind of analog problem-solving for a billion years. We are very late to notice.

#The Part Where I Stay Honest

So is this going to run your models? No. Not close, and anyone selling you that is lying.

The limits are real and the field admits them. Living fungal signals are glacially slow, on the order of half an hour to travel a meter, which is why even Adamatzky says the goal is not to replace silicon but to build large living sensors. The Ecovative chip is dried rather than alive, so it sidesteps the worst of that, but it is still one prototype on one benchmark. The best fungal memristors are still slower than the worst conventional ones. This is not general-purpose computing and it will not be for a long time, if ever.

What it is, is a different answer to a real and worsening problem. Global data-center electricity is on track to pass a thousand terawatt-hours this year, and as I keep writing, power, not chips, is the binding constraint now. The human brain does roughly exaflop-scale work on about 20 watts and is something like a million times more efficient than our AI hardware. That gap is the entire reason serious people, and DARPA, who funded this paper, are pouring money into computing that happens in living or once-living matter.

The mushroom chip is one bet on that frontier. At the other end sits Cortical Labs and its dish of live human neurons learning to play Pong, which I wrote about in the biological computing teardown, a 35,000 dollar machine keeping cells alive in a box. One approach grows cheap dead scaffolds by the millions for pennies. The other keeps expensive living tissue alive for months. Both are reaching for the same prize: computation that does not cost a power station.

#What the Story Actually Is

Here is the thing I cannot stop turning over. For fifty years, progress in computing meant bending nature to our will, forcing ever-smaller patterns into silicon against everything physics wanted to do, and paying for it in heat and power, the wall I wrote about last week. The mushroom chip is the opposite instinct. It lets biology do the messy nonlinear thing it already does effortlessly, and just learns to read the answer out.

Most of these grown chips will never be in a data center. The realistic near future is humble and a little beautiful: ultra-cheap, compostable, mass-grown analog sensors and inference at the edge, a computer you bury in a field to listen to the soil and then let rot. Three million at a time. Under a dollar each. Programmed by how they grew.

We spent half a century learning to build computers. We are now, very tentatively, learning to grow them. I would not bet on the mushroom replacing the GPU. But I would bet that the substrate of computing, the thing we always treated as a fixed given, is about to stop being one. Diamond on one side, mushrooms on the other, and silicon, for the first time in a long time, looking like just one option among several.