The experiment sounded like science fiction, or maybe just science gone slightly mad. In early 2026, an independent researcher managed to train roughly 200,000 living human neurons—cultured in a dish, kept alive with carefully calibrated electrical pulses—to play Doom. Not simulate playing Doom. Actually play it.
Sean Cole, the researcher behind that viral demonstration, didn't stop at the proof of concept. By June, he'd turned the work into Parasma, a one-person startup backed by Y Combinator with a pitch that sounds almost too simple: "We write the algorithms to turn brain cells into compute."
The timing, perhaps more than the founders expected, turned out to matter. As 2026 rolled on, data center operators found themselves scrambling for power contracts, AI labs waited months for GPU allocations, and the industry faced an uncomfortable truth—electricity, not silicon, had become the binding constraint on artificial intelligence development. Against that backdrop, Cole's proposal to sidestep the bottleneck entirely by training neurons instead of transistors began to look less like a curiosity and more like a serious, if still speculative, alternative.
Whether it will work at scale remains an open question. But in an industry increasingly defined by energy scarcity, the question itself carries weight.
Power Hunger Meets Biological Efficiency
The numbers tell a stark story. Gartner projected that data centers would consume 565 terawatt-hours of electricity in 2026, up significantly from the year before. The International Energy Agency's base case anticipated that figure nearly doubling to around 945 TWh by 2030, driven largely by AI workloads.
One blunt July 2026 analysis framed it this way: access to power will determine winners and losers in the AI race.
Enter biological computing. Human brains run on about 20 watts while performing tasks that require hundreds of kilowatts from GPU clusters. If even a fraction of that efficiency could be replicated in engineered systems, the economic calculus would shift dramatically. It's that promise—energy savings measured in orders of magnitude—that's drawing venture capital into what was, until recently, a purely academic curiosity.
Cole's path into this frontier began with hardware from Cortical Labs, an Australian company that pioneered what it calls the DishBrain platform. That system had already made headlines in October 2022 when peer-reviewed research published in Neuron showed neurons learning to play Pong. Cole's Doom demo took the concept further, demonstrating a closed-loop training paradigm: neurons receive electrical stimulation encoding the game state, respond with activity patterns decoded into control commands, and get feedback signals that guide learning over time.
His research note from June 2026 describes the stimulation as "biphasic, charge-balanced... on the order of one to a few microamperes," with training protocols optimized for task performance. Multiple outlets identified Cole as the researcher behind the work, which drew on support from Cortical Labs scientists Brett Kagan, Hon Weng Chong, and Alon Loeffler, as well as the University of Sussex.
By June 18, Cole announced Parasma's acceptance into Y Combinator's Summer 2026 batch. His profile notes a nontraditional trajectory for someone now tackling one of computing's most speculative frontiers.
From Lab Bench to Commercial Ambition
Parasma is entering a field that has moved, with startling speed, from academic theory to tentative commercialization. The foundational framework—organoid intelligence, or OI—took shape around a February 2023 roadmap published in Frontiers in Science by Johns Hopkins researchers and collaborators. The so-called Baltimore Declaration laid out what would be needed to scale brain organoids as compute substrates: larger and longer-lived cultures, high-density electrode arrays, closed-loop algorithms, robust data infrastructure.
Three years later, parts of that vision are materializing. Cortical Labs began shipping its CL1 device in 2025—a benchtop biocomputer widely reported at around $35,000 per unit. The device supports neuron cultures for up to six months and provides a Python API for control, effectively commoditizing what had been cutting-edge neuroscience equipment not long before.
Switzerland's FinalSpark launched a remote Neuroplatform in May 2024, granting nine institutions cloud access to living neuron processors for experiments running several months. The Biological Computing Co., a U.S.-based startup, raised $25 million in seed funding in February 2026 with the stated goal of integrating living neurons with foundation models to reduce compute costs on vision and generative tasks.
Academic labs, meanwhile, continue pushing complexity. Indiana University's "Brainoware" organoid reservoir computer, described in Nature Electronics in 2023, demonstrated speech recognition and nonlinear prediction tasks. Johns Hopkins' SURPASS program, funded through late 2024 and into 2025, advanced work on larger organoids and new testing modalities.
Federal interest is growing, too. The National Science Foundation's EFRI program solicited proposals for biocomputing through engineered organoid intelligence in 2024–2025. DARPA launched an "O-CIRCUIT" organoid cytomorphic intelligence program, with notices appearing in 2026.
Cortical Labs itself secured investment from Horizons Ventures and In-Q-Tel in April 2023 and expanded with a Gobi Partners-backed Malaysian manufacturing hub in March 2026. The company has positioned its technology around the "free energy principle"—a neuroscience framework suggesting that biological systems are inherently optimized for efficient inference under uncertainty. An April 2026 LiveScience feature reported that Cortical Labs is piloting a Melbourne facility and planning a larger Singapore site it describes, perhaps optimistically, as a "biological data center."
Technical Realities and Ethical Uncertainties
The challenge, of course, is translating lab demonstrations into reliable, scalable systems—something silicon solved decades ago but biology has yet to match.
Current neuron-based platforms face limitations that underscore just how early-stage this technology remains. Biological variability. Limited lifespan, measured in months rather than years. Reproducibility difficulties. The CL1's six-month viability window is a benchmark, certainly, but it also highlights the consumable nature of wetware. Neurons must be refreshed. Cultures maintained. Environments controlled with microfluidic precision.
Cole's June 2026 research note on consciousness and suffering offers a window into the constraints—and the ethical considerations. The Doom experiment used a single culture of around 200,000 neurons, a far cry from the roughly 86 billion in a human brain. The note argues methodically that consciousness is negligible at that scale and complexity, a calculation that becomes both scientific and ethical as systems potentially grow larger.
A Nature editorial published in late July 2026 flagged governance gaps that could slow the field's development. Biocomputing projects using human stem-cell-derived neurons may fall outside traditional institutional review board oversight if they're housed in computer science or engineering departments rather than biomedical labs. The editorial urged independent ethics review for all biocomputing work—a procedural hurdle that responsible companies may face even as less scrupulous actors sidestep it.
Donor consent presents another unresolved complication. Neurons are often derived from induced pluripotent stem cells created from donor blood or skin samples, originally collected for medical research. Using those cells for commercial compute applications raises questions that existing consent frameworks don't cleanly address. Chile's 2021 neurorights constitutional reform (Law 21.383) has been analyzed in 2026 legal scholarship as one early attempt to protect brain data, though enforcement mechanisms remain underdeveloped.
The technical substrate also remains fragile and expensive. High-density microelectrode arrays—the interface between neurons and software—are supplied by companies like MaxWell Biosystems, 3Brain, Multi Channel Systems, and Axion BioSystems. These platforms can map tens to hundreds of thousands of electrodes onto organoid surfaces, but the equipment requires specialized facilities. Cortical Labs operates at Australia's Physical Containment Level 2 laboratory standards, a requirement that adds operational complexity and cost.
Uncertain Returns, Undeniable Urgency
Parasma's job posting on Y Combinator's Work at a Startup board, listed as of early August 2026, offers a glimpse of both ambition and challenge. The company is seeking a "Founding Scientist" with 11-plus years of experience. Compensation is listed as "$1M" with equity between 1.00 and 3.00 percent—a figure unusual for an early-stage role and one that should be treated cautiously. But it signals the kind of scarce, specialized talent biocomputing startups are competing for.
The near-term trajectory for the field appears less about displacing GPUs than carving out niches where biological systems might offer advantages. Adaptive learning under sparse or noisy data, perhaps. Low-power edge inference. Or research platforms for neuroscience itself. The 2023 Organoid Intelligence roadmap positioned hybrid bio-AI systems as a long-term goal, acknowledging that silicon and wetware might complement rather than replace each other for decades.
What remains unclear—genuinely unclear, not just uncertain—is whether venture-scale returns are possible in biocomputing's current form. The Biological Computing Co.'s $25 million seed round and Cortical Labs' expansion into Malaysian manufacturing suggest some investors see a path. But the technology is far from proving it can compete with silicon on cost, reliability, or performance for mainstream AI tasks.
FinalSpark has made vendor claims about energy savings, though independent benchmarking of energy-per-task metrics remains sparse. The competitive case for biological compute hinges almost entirely on energy efficiency, and hard data on real-world performance is still limited.
For now, Parasma represents a bet—shared by a handful of startups, a growing academic community, and a small but notable set of investors—that the AI energy crisis is urgent enough to justify exploring radically unconventional solutions. Whether training neurons to predict tokens or navigate virtual corridors will scale into a new compute paradigm or remain a research curiosity is a question that may take years, possibly decades, to answer.
But in an industry where power has become as scarce as talent, it's a question worth asking. Even if the answers, for now, remain frustratingly incomplete.
