You Can Now Buy a Computer Made of Human Brain Cells for $35,000. Here's What It Actually Does.

Two hundred thousand neurons, grown from human stem cells, sitting on a silicon chip inside a climate-controlled box with pumps, gas lines, and nutrient feeds. Price: $35,000. Lifespan before the cells need replacing: six months. Best demonstrated capability: playing the 1993 video game Doom at a level the creators themselves describe as "a beginner who's never seen a computer."

That is the Cortical Labs CL1, announced at Mobile World Congress in Barcelona in March 2025 and entering its first commercial shipments. It is the world's first commercially available biological computer, and it represents a genuine inflection point for a field that has spent decades as academic curiosity. You can now buy computing power that runs on living human tissue.

Whether the thing you can buy is a computing revolution or a very expensive science fair project is the question worth interrogating.

A Scorecard Nobody Has Built

Cortical Labs is not alone. At least four platforms in three countries now offer some form of biological computing, either for sale or for rent. Nobody has compared them side by side. So I did.

Platform	Origin	Neurons/Unit	Price	$/Neuron	Power/Unit	Best Demo
Cortical Labs CL1	Australia	~200,000	$35,000	$0.175	~33W	Doom (beginner)
FinalSpark Neuroplatform	Switzerland	~100K (4 organoids)	$1,000/mo	Rental	Not disclosed	Data processing
Brainoware (Indiana U.)	USA	Organoid scale	Research only	N/A	N/A	78% speaker ID
MetaBOC (Tianjin U.)	China	Not disclosed	Research only	N/A	N/A	Robot control

For context, here are the systems these biological computers are supposed to eventually compete with:

System	Processing Units	Price	Power	Best Capability
Human brain	86 billion neurons	N/A	20W	General intelligence
NVIDIA H100 GPU	80 billion transistors	~$30,000	700W	GPT-4 class inference

A CL1 costs $5,000 more than an H100 and contains 400,000 times fewer processing elements. Its best trick is navigating a 33-year-old video game poorly. Not a kind comparison, but an honest one.

What Efficiency Claims Actually Mean

One number circulates in every biocomputing press hit: "1 million times more energy-efficient than silicon." FinalSpark's strategic advisor Ewelina Kurtys told Interesting Engineering that "living neurons are 1 million times more energy efficient compared to silicon-based hardware." That claim traces back to a real biological fact: a single neuron consumes approximately 10 attowatts per synapse operation, while a transistor switching in a modern GPU consumes roughly 10 picojoules. A factor of about a million.

But that comparison strips away everything that keeps neurons alive.

A CL1 rack of 30 units consumes 850 to 1,000 watts, according to Cortical Labs' own disclosures. That powers 30 units containing a combined 6 million neurons. Do the math: 1,000 watts divided by 6 million neurons equals 0.000167 watts per neuron.

A human brain runs 86 billion neurons on 20 watts total: 0.00000000023 watts per neuron.

That makes the CL1 approximately 725,000 times less energy-efficient per neuron than the biological system it was engineered to replicate. Life support equipment, temperature regulation, pumps cycling nutrients, sterile gas systems: they erase virtually all of the bare-neuron efficiency advantage. When FinalSpark says "1 million times more efficient," they are describing the neurons in isolation. Plug in the machine that keeps those neurons alive, and the advantage vanishes.

What Living Cells Can Actually Do

Cortical Labs first demonstrated biological computing in a 2022 paper published in Neuron. Their experiment, called DishBrain, placed approximately 800,000 human and mouse neurons on a multi-electrode array and trained them to play Pong. Cells received electrical stimulation corresponding to the ball's position and returned signals that moved the paddle. Performance improved over time, and neurons adapted their firing patterns based on feedback. Researchers described the behavior as "sentient" in the paper's language, a word choice that drew immediate scrutiny from the neuroscience community.

In February 2026, the team escalated to Doom. Using commercially available CL1 hardware, they fed the game's visual environment to 200,000 neurons via electrode stimulation and translated their electrical responses into movement and shooting commands. Cells moved through corridors and engaged enemies. "Is this an esports champion? Absolutely not," the team noted in their demonstration video. "Right now, the cells play like a beginner who's never seen a computer."

Separately, at Indiana University, a team led by Feng Guo published results in Nature Electronics in December 2023 showing that a brain organoid connected to a chip (dubbed "Brainoware") could distinguish between eight different speakers from vowel sounds with 78% accuracy after just two days of training using 240 audio clips. A conventional artificial neural network trained on the same task required substantially more data and compute to reach similar accuracy, though it eventually outperformed the organoid on longer training runs.

In China, researchers at Tianjin University and Southern University of Science and Technology built MetaBOC, a brain-on-chip system that autonomously controlled a simple robot. Their organoid processed sensor data and directed the robot through obstacle avoidance and object grasping tasks. Published in 2024, the work demonstrated that biological tissue could close a perception-action loop in real time, not merely respond to stimuli in a game environment.

Wetware as a Service

Cortical Labs is pursuing two revenue streams. CL1 hardware sells at $35,000 per unit or $20,000 per unit when purchased in 30-unit server racks ($600,000 per rack). A Wetware-as-a-Service (WaaS) option runs $300 per week per unit, providing remote access to neurons maintained in Cortical Labs' own facilities. Initial shipment covers 115 units, with four server racks planned to be online by end of 2025.

FinalSpark, based in Switzerland, has scaled to over 1,000 brain organoids and collected more than 18 terabytes of neural activity data. Pricing starts at $1,000 per month for access to four organoids, with universities receiving some projects for free. Co-founder Fred Jordan frames the value proposition bluntly: "A human brain consumes roughly 20 watts to power 100 billion neurons. To simulate a human brain, you would need the energy output of a nuclear power plant. Bioprocessors could be up to a million times more energy-efficient."

Target customers are not companies looking to replace their GPU clusters. Pharmaceutical companies testing drug compounds on living neural tissue, neuroscience labs studying network dynamics, and AI researchers exploring what biological architectures can teach about learning efficiency form the early market. Drug toxicity screening alone represents a use case where a $35,000 device that keeps human neurons alive and measurable could save millions in failed animal trials.

An Ethics Question That Scales Poorly

A 2026 paper in Scientific Reports surveyed attitudes toward embodied brain organoids and found that perceptions of consciousness significantly shaped ethical judgments. A 2024 review in Nature Reviews Neuroscience by Bayne, Seth, and Massimini asked directly: "Are brain organoids conscious?" Their answer: current organoids almost certainly lack the structural complexity for consciousness, but the question becomes less theoretical as organoids grow larger and more interconnected.

Cortical Labs' CL1 contains roughly 200,000 neurons. A fruit fly brain contains about 100,000. A honeybee brain contains about 960,000. A mouse brain contains 71 million. No scientific consensus exists on what threshold of neural complexity, if any, produces subjective experience. Cortical Labs maintains an ethics board and states that their systems are tools for research, not sentient beings.

But commercial logic points toward scale. More neurons means more compute. More interconnection means more capability. This industry's own trajectory is a collision course with its own ethical boundary, and no regulatory framework exists for adjudicating "is this cluster of neurons a someone?" Europe's AI Act, which took effect in 2025, does not address biological computing substrates. No national regulatory body has published guidelines specific to organoid computation.

Why This Might Be a Dead End

Here is the strongest argument that biological computing is a dead end rather than a beginning: the lifespan problem. CL1 neurons survive about six months before the culture degrades and needs replacement. That is not a battery running down. It is a fundamental biological constraint: cells outside a body lack the vascular systems, immune protection, and repair mechanisms that maintain neural tissue for decades. Every six months, you lose your "trained" network and start over.

Silicon does not forget. An H100 trained on a model retains those weights indefinitely, consumes zero maintenance power when idle, and can be replicated exactly across thousands of identical chips. A biological computer's state dies with its cells. Until someone solves the persistence problem, biological computing functions more like a research instrument that generates insights about neural learning than a computing platform that accumulates capability over time.

This is not a minor engineering detail. It is the central question: can you build a computing industry on hardware that expires?

What Living Data Is Worth

FinalSpark's 18 terabytes of neural activity data may end up being more valuable than the organoids that generated it. If biological computing's real contribution is not direct computation but rather datasets that teach silicon AI how neurons learn, then a $1,000-per-month rental fee is not buying compute. It is buying training data for the next generation of neuromorphic algorithms. Researchers studying spike-timing-dependent plasticity, dendritic computation, and metabolic constraints on learning have never had access to this volume of real-time, living neural network data.

Brainoware's result at Indiana University points in the same direction. Its organoid reached 78% accuracy on speaker identification after two days. A conventional neural network trained on the same dataset eventually surpassed it. But the organoid learned from far fewer examples and used orders of magnitude less energy during training. If that learning efficiency can be reverse-engineered into silicon algorithms, biological substrates become research tools, not production platforms.

Limitations of This Analysis

Several important caveats apply. No independent benchmarks exist for any biocomputing platform. All performance claims come from the companies and labs themselves. Cortical Labs has not disclosed customer outcomes or third-party validation of its learning capabilities. Brainoware's 78% accuracy was measured on a narrow task (speaker identification from vowel sounds with eight speakers) and does not indicate generalizable intelligence. FinalSpark's energy efficiency claims compare bare neuronal metabolism to GPU power draw without accounting for life support overhead. MetaBOC's robot control demonstrations have not been independently replicated. Pricing for FinalSpark's industrial tier and MetaBOC is not publicly available.

The Bottom Line

Biological computing crossed from hypothesis to product in 2025. You can order a computer made of human neurons, plug it in, and run code on living tissue. Four platforms across Australia, Switzerland, the United States, and China are operating simultaneously. Combined investment in the sector exceeds $50 million. But the gap between marketing ("a million times more efficient than silicon") and engineering reality (725,000 times less efficient per neuron than a brain, with cells that die every six months) is enormous. What the first generation of biological computers likely delivers best is not computation itself. It is the data they generate about how biological networks learn, data that could make silicon AI better. Whether that makes a $35,000 box of neurons a bargain or a curiosity depends entirely on what the next five years of neuroscience can extract from 18 terabytes of living network activity.