Startup Engineers Living Human Brain Cells to Play Video Games

In February 2026, a clump of roughly 200,000 living human brain cells grown on a silicon chip learned to play Doom — the 1993 first-person shooter that has become an unofficial benchmark for running on anything with a processor. The neurons, cultured by Australian biotech startup Cortical Labs, could locate enemies, move through corridors, and fire weapons, though they died frequently [1]. The demonstration was the latest and most visible milestone in an emerging field that its proponents call "organoid intelligence" and its critics call premature hype built on ambiguous evidence.

The question is no longer whether human neurons can be wired into a computer. They can. The question is whether they should be — and whether doing so produces anything that silicon cannot already do better.

The Players and Their Money

Three companies sit at the center of biocomputing's commercial push. Cortical Labs, founded in Melbourne, has raised approximately $15 million across multiple rounds, with investors including Horizons Ventures, Blackbird Ventures, and notably In-Q-Tel, the CIA's venture capital arm [2][3]. Koniku, a California-based company founded by Osh Agabi in 2015, has raised roughly $37 million over seven rounds, including backing from SoftBank, and has pivoted toward chemical-sensing applications rather than general-purpose computation [4]. FinalSpark, a Swiss startup with just six employees, is seeking CHF 50 million ($62.7 million) to fund its next stage of research, though its current funding comes primarily from its parent company AlpVision [5].

Source: Crunchbase / PitchBook

Data as of Mar 1, 2026CSV

The broader market these companies are chasing is small but growing fast. Industry analysts at Intent Market Research project the wetware computing market will reach $3 billion by 2030, up from $260 million in 2023 — a compound annual growth rate of 41.5% [6]. The U.S. Department of Energy announced $15 million in funding for 22 wetware computing research projects in 2022 [6].

Source: Intent Market Research

Data as of Jan 1, 2025CSV

These numbers remain a rounding error compared to the $500+ billion global semiconductor industry, but they reflect growing institutional interest in whether biological substrates might address AI's mounting energy problem.

What These Systems Actually Do

Each company takes a different approach. Cortical Labs' CL1 — marketed as the "world's first commercial biological computer" — consists of human neurons grown on a planar microelectrode array with 59 electrodes [7]. Gameplay input is converted into electrical stimulation patterns: when an enemy appears on the left side of the Doom screen, electrodes zap the corresponding region of the neural culture. The neurons fire back spikes that the system interprets as actions — move, turn, shoot [1].

The original DishBrain system, published in Neuron in December 2022, used approximately 800,000 human and mouse cortical neurons to play Pong and showed measurable performance improvement within five minutes [8]. A standard deep reinforcement learning algorithm required about 90 minutes to reach comparable Pong performance [9]. The Doom demonstration in 2026 used a smaller, 200,000-neuron array on the CL1 platform.

FinalSpark takes a different path. Its Neuroplatform hosts 16 human brain organoids — small, three-dimensional clusters of neurons derived from induced pluripotent stem cells — connected to electrodes for remote access by researchers worldwide [5]. The company claims these bioprocessors consume "a million times less power" than digital processors for equivalent computations [10], though independent verification of this specific claim has not been published in peer-reviewed literature.

Koniku's approach is arguably the most commercially grounded. Rather than pursuing general-purpose computation, its Konikore device programs neurons with specific chemical receptors to detect target substances. In 2017, Koniku secured an $8 million contract with Airbus to develop biological threat-detection systems for aviation security, and the technology was reportedly deployed at San Francisco International Airport [4].

The Scale Problem: 200,000 vs. 86 Billion

The gap between current biocomputing systems and the human brain is staggering.

Source: Various academic sources

Data as of Mar 1, 2026CSV

Cortical Labs' CL1 operates with roughly 200,000 neurons. FinalSpark's organoids contain tens of thousands. A fruit fly brain has about 100,000 neurons. A mouse brain contains 71 million. The human brain has approximately 86 billion neurons connected by roughly 100 trillion synapses [11].

Current brain organoids in research settings typically measure under 500 micrometers in diameter and contain fewer than 100,000 cells [11]. At what neuron count does sentience or pain perception become plausible? Neuroscientists do not agree on a threshold. The concept itself may be misframed — many researchers argue that sentience depends not on raw neuron count but on functional architecture, connectivity patterns, and specific structures. One influential framework, proposed by philosopher Jonathan Birch, suggests a "brainstem rule": if a neural organoid develops or connects to a functioning brainstem capable of regulating arousal, registering needs, and producing sleep-wake cycles, it becomes a candidate for sentience [11]. No current biocomputing system comes close to this.

Performance Claims Under Scrutiny

The central marketing claim of biocomputing — that neurons compute more efficiently than silicon — rests on a real observation wrapped in significant extrapolation. The human brain operates on about 20 watts of power, roughly equivalent to a dim lightbulb, while performing computations that would require orders of magnitude more energy on conventional hardware [12]. Digital computers are estimated to be 10,000 to 100 million times less energy-efficient than biological neurons for certain parallel processing tasks [12].

But translating that theoretical advantage into working technology is a different matter. The CL1 unit draws between 850 and 1,000 watts — mostly for the life-support systems that keep neurons alive, including temperature regulation, nutrient delivery, and waste removal [7]. At $35,000 per unit, the cost-per-computation is not competitive with conventional hardware for any current application.

Cortical Labs plans to build data centers housing CL1 racks — 120 units in Melbourne and 20 units at the National University of Singapore, in partnership with data center company DayOne [7]. The company's roadmap targets "Wetware as a Service" cloud access in late 2025, with ambitions to scale to 1,000 units [7].

As for the Doom demonstration itself: Cortical Labs acknowledges the gameplay "resembles a complete beginner who has never seen a keyboard, mouse, or indeed a computer before." The neurons perform better than random firing but lose frequently [1]. No independent replication of the Doom results has been published. The earlier Pong results in Neuron were peer-reviewed but drew criticism — a response paper by Fuat Balci and colleagues in Neuron in March 2023 argued that the original study's use of terms like "sentience" and "intelligence" was "not based on any established or robust consensus" and relied on "relatively weak evidence" [13].

The Skeptics' Case

Tony Zador, a computational neuroscientist at Cold Spring Harbor Laboratory, has described efforts to produce organoid intelligence on par with silicon-based AI as "a scientific dead-end" [14]. Several lines of criticism are worth taking seriously.

Scalability. Biological neurons are fragile. They require precise temperature, pH, nutrient supply, and waste removal. Scaling from 200,000 neurons to millions — let alone billions — while maintaining viability and functional connectivity is an unsolved engineering problem fundamentally different from scaling transistor counts on silicon.

Reproducibility. Each organoid develops differently. Unlike manufactured chips, biological cultures exhibit high variability, making standardized manufacturing and quality control extraordinarily difficult [14].

Misattributed learning. The "learning" observed in DishBrain may not represent goal-directed adaptation. Neurons naturally form connections and produce activity patterns in culture — a phenomenon known as spontaneous network activity. Distinguishing genuine learning from noise artifacts or inherent self-organization requires careful controls that critics argue have not been adequately demonstrated [13].

Competition from neuromorphic chips. Intel's Loihi 2 neuromorphic processor, which mimics the spiking behavior of biological neurons in silicon, can process neural networks up to 5,000 times faster than biological neurons [15]. The Hala Point system, built from Loihi 2 chips, can simulate 1.15 billion neurons at 20 times the speed of a biological brain [15]. A single Loihi 2 chip supports up to 1 million neurons and 120 million synapses [15]. These systems offer the brain-inspired advantages — energy efficiency, event-driven processing — without the complications of keeping cells alive.

Stanford neural organoid researcher Sergiu Pasca has cautioned that "using accurate terms that neither hype nor misrepresent the work really does matter," reflecting broader concern among organoid scientists that inflated commercial claims could trigger public backlash that spills onto basic research [14].

The Optimists' Case

Proponents counter with properties of biological neurons that silicon has not successfully replicated at scale.

Spike-timing-dependent plasticity (STDP). Biological synapses strengthen or weaken based on the precise timing of pre- and postsynaptic spikes — a form of learning that occurs continuously, locally, and without external training signals [16]. This stands in contrast to the backpropagation algorithm used in artificial neural networks, which requires a separate, computationally expensive training phase.

Catastrophic forgetting resistance. Large language models and other deep learning systems notoriously suffer from catastrophic forgetting — when learning new information overwrites previously learned knowledge. The brain handles this through multiscale plasticity mechanisms, where neuromodulators regulate learning across brain regions [16]. A 2023 paper in Science Advances demonstrated that brain-inspired algorithms using STDP principles reduced learning energy consumption by approximately 98% while achieving higher accuracy than conventional methods [16].

Analog signal processing. Biological neurons process information using continuous, graded signals rather than discrete digital bits. This analog processing enables certain types of computation — pattern recognition in noisy environments, for example — that require extensive resources on digital hardware.

Sub-milliwatt operation. Individual biological neurons consume energy on the order of femtojoules per synaptic event. If the life-support overhead problem can be solved, proponents argue, biological substrates could offer computation at energy levels that silicon will never reach [10].

The energy argument carries particular weight given the trajectory of AI infrastructure. Training a single large language model can consume gigawatt-hours of electricity. If biological systems could perform inference at a fraction of that energy cost, the economic incentive would be substantial — even if the systems are slower in absolute terms.

The Regulatory Vacuum

No regulatory framework specifically governs the use of human-derived neural tissue in computing applications. The gap is wide and growing.

In the United States, NIH guidelines restrict federal funding for certain human-animal chimeric research but contain no provisions addressing organoid computing [17]. The FDA Modernization Act 2.0, signed in December 2022, permits alternatives to animal testing — including organoids — but addresses them as tools for drug development, not as computational substrates [18]. The International Society for Stem Cell Research (ISSCR) guidelines do not specifically address biocomputing applications [11].

In the European Union, the European Group on Ethics in Science and New Technologies (EGE) provides ethical guidance, but specific regulatory classifications for wetware computing systems do not exist [17]. Are these systems research subjects? Medical devices? Neither category fits cleanly.

The NIH announced $87 million in contracts in 2025 to establish a Standardized Organoid Modeling (SOM) Center, but its mandate focuses on standardizing organoid methods for drug testing, not computing [18]. STAT News reported in November 2025 that ethics concerns had prompted calls for global oversight of neural organoid research more broadly, with several scientists arguing existing frameworks are inadequate [17].

Who Donated the Cells?

Cortical Labs' neurons are derived from human blood cells converted into induced pluripotent stem cells (iPSCs), which are then differentiated into neurons [7]. iPSC technology, developed by Shinya Yamanaka (who won the Nobel Prize for the work in 2012), allows researchers to reprogram ordinary adult cells — typically blood or skin cells — into a stem-cell-like state capable of becoming any cell type.

The informed consent landscape for iPSC-derived research remains shaped by the 1990 California Supreme Court decision in Moore v. Regents of the University of California. In that case, the court ruled 4-3 that John Moore did not have property rights over cells taken from his body — specifically, his spleen cells, which his physician used to develop a lucrative cell line without Moore's knowledge [19]. The court held that while physicians must disclose research interests to patients, patients do not have property claims over excised biological material [19].

This precedent means that donors of blood or skin cells used to create iPSCs for biocomputing generally have no legal claim to profits derived from those cells — provided informed consent was obtained for the initial cell collection. However, the consent forms used in most iPSC banking operations were written before biocomputing was a commercial possibility, raising questions about whether donors understood the full range of potential applications [20]. The legal landscape varies by jurisdiction, and no court has specifically addressed the question of commercial biocomputing applications derived from donated tissue.

The Academic Surge

Research interest in brain organoids has exploded over the past decade, with more than 43,000 papers published since 2011.

Source: OpenAlex

Data as of Jan 1, 2026CSV

Biocomputing as a field has seen a parallel, if smaller, surge in academic publications, peaking at 1,560 papers in 2023 [21].

Source: OpenAlex

Data as of Jan 1, 2026CSV

This academic activity suggests the field has momentum, but the gap between laboratory demonstrations and commercial viability remains large.

Timelines and Economic Implications

Cortical Labs' public roadmap targets several milestones: shipping CL1 racks in late 2025, achieving 95% uptime for cloud access, validating its "Minimal Viable Brain" through peer review, and securing at least ten non-academic commercial clients [7]. The company has discussed applications in drug discovery, pattern recognition, and as co-processors alongside conventional AI hardware.

If these systems advance from playing rudimentary video games to performing economically useful tasks, the realistic timeline — based on the companies' own projections — spans years, not months. Drug discovery, where biological systems might model cellular responses more accurately than simulations, is frequently cited as a near-term application. Financial modeling and autonomous systems are mentioned as longer-term possibilities, but no company has published benchmarks demonstrating competitive performance on any of these tasks.

The question of job displacement is, for now, speculative. Biocomputing systems in their current form are not replacing any existing technology. If they eventually prove useful, the most likely early applications — pharmaceutical screening, specialized sensor systems, and low-power inference — would augment rather than replace existing computational infrastructure. Koniku's chemical-detection work is arguably the only biocomputing application currently generating revenue from non-research customers [4].

What Comes Next

The field of biological computing sits at an uncomfortable intersection of genuine scientific potential, commercial incentive, ethical uncertainty, and regulatory absence. The neurons on Cortical Labs' chip are, in a narrow technical sense, playing Doom. Whether that represents the first step toward a new computing paradigm or an expensive parlor trick depends on answers to questions that have not yet been asked in the right forums — regulatory bodies, ethics boards, and peer-reviewed journals — with the rigor the subject demands.

The brain organoid researchers who built the scientific foundation for this work are watching the commercial applications with a mix of interest and anxiety. As one unnamed organoid scientist told STAT News: the worry is not that the science is bad, but that the marketing outpaces it [14]. In a field where the basic science is still being established, the distance between a promising result and a proven technology is measured not in neurons, but in years of careful, transparent work.