Hey crew,

The Weird Thing I Can't Stop Thinking About,

A dish of neurons just started earning a paycheck.

A startup in Salt Lake City put living human neurons online.

Then they let random strangers play Tetris against them.

Not a simulation. Not "neural network" as a metaphor. Actual brain cells. Grown from stem cells. Wired to electrodes. Deciding, piece by piece, where to drop the next block.

You can try it right now, for free, at play.intactis.bio.

It's called Biostack, built on what they call a Biohybrid Processing Unit — a rack-mountable box that sits in a GPU's slot, except it's alive.

I've spent real hours this year on quantum, running circuits on IBM's Fez processor — thinking hard about qubits as a different computational substrate. Turns out quantum isn't the only exotic-substrate bet on the table right now.

Biocomputing is the other one. And it's moved a lot faster than I expected.

So I did the deep dive. Not "brain cells are the next GPU" — what's actually real, what's marketing, and what's worth doing about it.

Strip away the sci-fi framing and the definition is simple: use actual biological matter — DNA, engineered cells, or living neurons — to store or process information instead of transistors.

That's it. But it's not one field. It's three, chasing three different problems.

Guess which one just raised real venture money?

The third one. That's this issue.

2022. A Melbourne team led by Brett Kagan wired roughly 800,000 human and mouse neurons into a simplified game of Pong. Within five minutes of feedback, the culture started improving. The paper called it "sentience" — a word that made neuroscientists wince. The company has since walked it back: the culture adapts to a goal. It isn't conscious.

2025. Cortical Labs turned the experiment into a product. The CL1 launched in Barcelona — a shoebox-sized "body-in-a-box," living neurons on a silicon chip, kept alive up to six months by onboard pumps, filtration, and temperature control. Price: $35,000. Or rent time on it in the cloud — "wetware-as-a-service." They've since gotten it to run Doom.

2024–26. A Swiss company, FinalSpark, built something different: 16 lab-grown brain organoids, accessed remotely over Python, for $500 a month. Nine-plus universities already run experiments on it. Their claim: roughly a million times less power than silicon. Treat that as a marketing ceiling, not a measured floor, until someone reproduces it independently.

February 2026. This is the one that actually got my attention. The Biological Computing Company — TBC, founded by two neurosurgeons — came out of stealth with a $25 million seed and became the first company ever paid for biological compute. Not a grant. An invoice. Their pitch: encode a hard problem into living tissue, extract the solution, run it as ordinary software on regular GPUs. Claimed gain: 23x, on a real video model.

July 2026. Intactis Bio — the Tetris team — shipped their own BPU, targeting anyone already burning $20K+ a month on GPUs. Claiming up to three million times the efficiency of silicon. Same asterisk applies: unverified, pending independent proof.

Lab curiosity to $35K product to cloud rental to paid enterprise contract. In under four years.

One number shows up in every pitch deck in this space.

The human brain runs on roughly 20 watts. NVIDIA's Blackwell needs around 120,000 watts for comparable throughput, by TBC's own framing.

Whether that exact multiplier survives scrutiny or not, the direction of the gap is real — and it's the same gap driving neuromorphic chips and optical computing right now. AI's energy bill got brutal enough that even strange ideas started looking fundable.

Here's my opinion, flagged clearly as opinion: I don't think "replace the GPU with neurons" is where the value lands.

Keeping living tissue alive, fed, and electrically legible is an enormous tax. Nutrient delivery. Waste removal. Contamination control. Six-month lifespans. Variability transistors simply don't have. That's a lot of fragility to bolt onto a data center.

What I do believe: TBC's actual model. Biology as a discovery engine, not a replacement substrate. Let neurons find a solution the way four billion years of evolution taught them to. Decode the principle. Implement it in silicon, where it's cheap and needs no bioreactor.

That's a different business than "sell you a brain in a box." I think it's the one that survives — because most of these startups won't. Hard tech with living, dying inputs goes through brutal culling before real winners emerge. Quantum's going through the same thing right now, just further along.

Near term (1–2 years), high confidence. More companies copy TBC's "biological adapter" model — cheapest, least fragile version of the idea, already generating revenue. Expect it inside AI infra tools you'd never think to connect to biology.

Medium term (3–5 years), moderate confidence. The hardware plays — CL1, the BPUs — mostly stay in research and pharma niches: drug screening, disease modeling, an ethical stand-in for animal testing. Smaller market than "replace NVIDIA." Still a fine outcome for them.

Long term (5–10 years), genuinely unsure. Watch Cortical Labs' biological data center pilots with DayOne, in Singapore and Australia. If those hit anywhere near claimed efficiency at scale, this stops being a curiosity. If they quietly stall — the likelier outcome — biocomputing stays a fascinating side field. I genuinely don't know which way this breaks. Anyone who claims certainty here is selling something.

Worth sitting with too: as organoids get more sophisticated — Johns Hopkins has reported multi-region organoids with early signs of learning and memory — this field runs into a governance question nobody has a good answer for. Nobody thinks today's systems are conscious. Nobody has a test for when that stops being true.

Not "go get a neuroscience degree." Something better.

Go touch it first, for free. Play Biostack at play.intactis.bio. Five minutes, and "biological compute" moves from your head into your gut.

Want to go deeper? Rent real infrastructure. FinalSpark's Neuroplatform: $500/month, academic-style access, Python API. No biology PhD required — just comfort sending electrical signals in and reading them back out.

Learn the interface, skip the biology. Same bottleneck as always in AI: encode the problem as a signal, decode the response. That's signal processing and ordinary ML — not cell culture.

Watch TBC's playbook, not the hardware. "Biology as algorithm discovery" is the version of this thesis actually making money right now.

Size your attention correctly. Same bucket I put quantum in a year ago — high-risk, high-upside, worth two hours a month, not a roadmap rewrite. Exactly the kind of early, unpriced signal I like feeding into Kubera's research loop.

The real story here isn't "brain cells could replace your GPU."

It's that for the first time in decades, serious money is betting computation doesn't have to be exclusively electronic.

Quantum. Neuromorphic. Biological. Optical. Same question, different substrate: what happens when silicon stops being the only way to compute?

Nobody's won that argument yet. But for the first time, somebody's actually getting paid to test it.