A nineteen-year-old in London has raised $10 million in pre-seed money to grow brains, and the launch video is a dish of neurons playing Frogger.

The company is Frontier Computing, out of YC's Summer 2026 batch, round led by General Catalyst. The founder, Michael Domarkas, has apparently been culturing neuronal tissue since he was seventeen, which is a sentence that makes me feel a certain way about my own teenage years.

The pitch is one line: a GPU burns most of its energy shuttling data between memory and compute, and a neuron does both in the same place. So stop simulating neural networks on hardware that's bad at it, and grow the network instead.

And the demo numbers are specific enough to take seriously. One hour of real-time learning, 92% peak crossing rate over a 25-game sample. That's a real result written down like a real result, not a vibe.

It's also the least interesting thing in the announcement.

Frogger is a reproduction, not a breakthrough

We've been here. In 2022, Cortical Labs' DishBrain — 800,000 human and mouse neurons on a CMOS chip — taught itself Pong. That became the CL1, which is a shipping product you can buy, and which has since been wired into a prototype biological data centre. FinalSpark has been renting time on 160,000 neurons over the internet for a couple of years now.

Frogger is Pong with more lanes and a worse failure mode. As a demonstration that cultured neurons will close a loop with an environment and adapt, it's a replication of a four-year-old result on a bigger substrate.

Which is fine! Replication is good. But if you read the headline and thought the dish learned a video game, therefore the models are next, that's the wrong takeaway, and it's the takeaway the video is engineered to produce.

The actual claim is plumbing

Here's the bit worth caring about.

Biological neural tissue in a dish has been stuck at roughly a million neurons, and not because nobody wanted more. It's the vascularization problem: neurons in the middle of a growing lump can't get oxygen, because there's no circulatory system delivering it, and diffusion falls off with geometry. Grow past a certain radius and the core starves and dies. Every biocompute outfit on the planet has been pressed against that ceiling.

Frontier's claim is that they've built a tissue-culture approach that breaks it, and they're using that to build 100M and 500M neuron systems — the larger one supposedly live by the end of this year, though some write-ups put it in January 2027, which is the kind of wobble that tells you it's a target and not a delivery date. At 500 million they'd be about 2,500x the largest publicly available system in the field.

That's the news. Not Frogger. If someone has genuinely solved oxygen delivery in thick cultured tissue, that's a result with implications well outside computing, and the appropriate response is to want to see the method rather than the gameplay footage.

And right now, the 500M cluster is a plan. It's a build-out that has begun. There is a large and well-documented gap between we have a scaling primitive and we have a rack.

The energy argument is the honest part

The von Neumann bottleneck is real, and I don't want to be glib about it. Moving a weight from HBM into a multiply-accumulate unit costs vastly more energy than the multiply does. That's not marketing, that's the actual reason your training bill looks like that.

And the biology numbers are genuinely absurd. A CL1 unit draws around 30 watts; a full rack sits under a kilowatt, against roughly 6kW for a single data-centre GPU. A human brain does its thing on about 20 watts, which remains the most embarrassing benchmark in computing.

So the direction of the argument is sound. Compute that doesn't move data is better than compute that does.

The thing is, that's also the pitch for SRAM-heavy accelerators, for compute-in-memory silicon, for every wafer-scale part shipping today. Those approaches chase the same physics without needing an incubator. Biology's edge has to be more than "co-location is good" — it has to be that co-location at neuron density is unreachable in silicon, and that's a much harder claim to stand behind.

The part that makes me itch

Cultured neurons die.

Not as a failure mode — as a schedule. Neuron cultures run something like six months; organoids stretch from about a hundred days to fifteen depending on how heroic the maintenance is. Without circulation the tissue eventually necroses. That's the same constraint the vascularization work is aimed at, and even the optimistic version doesn't end with and then it lives forever.

Now put that next to the product. They're selling a training substrate. Training implies runs you can repeat, compare and roll back. And the literature on this stuff flags reproducibility as the central concern: developmental stochasticity means two cultures grown identically aren't identical, neurons fire spontaneously as normal behaviour, and that spontaneous firing is noise sitting underneath every measurement you take.

So the substrate is non-deterministic, individually unreliable, and has a shelf life. You can't diff two dishes. You can't git checkout last Tuesday's culture. When a run goes sideways you cannot tell, without a lot of work, whether your protocol was wrong or those particular neurons were just having a day.

Every one of those is a known research problem with people working on it. None of them is a rounding error, and all of them get harder at 500x the cell count, not easier.

"Training" is doing an enormous amount of work in that sentence

The part I'd want pinned down before anything else: what is the interface?

You do not backpropagate through tissue. There's no gradient to chase, no optimizer, no checkpoint. What these systems actually do is closed-loop stimulation — you encode state as electrical stimulus, read the population response, and shape behaviour with structured feedback. It works, demonstrably, for a paddle and now for a frog.

But the pitch says cost-effective training of large parameter count models, and those two things are not on the same continuum. A dish reaching 92% on Frogger isn't a small language model. There's no established path from one to the other, no equivalent of "just add layers", and nobody has shown what a loss function even means here.

That's not fatal. It's just the entire hard problem, and it's downstream of the biology they're currently raising against.

So

I want this to work, which is exactly why the framing bugs me.

There's a genuinely important claim buried in this launch — that the oxygen ceiling which has capped the whole field at a million neurons might be breakable. If that holds up under scrutiny it's a serious piece of science, and $10M is cheap for finding out.

Wrapped around it is a demo replicating a 2022 result, a cluster that doesn't exist yet, and the word "training" borrowed from a discipline that assumes determinism the substrate can't currently provide.

Both things are true at once. The way you tell which one you're being sold is to watch what they publish next. If it's the culture method and survival curves at 100M, they're doing the science. If it's another game, they're doing the round.

Anyway. Somewhere in London there's a lump of neurons whose entire experience of existence is a frog crossing a road, and I've been thinking about that all morning.