Somewhere in San Francisco, a cluster of approximately 200,000 human neurons is trying—really trying—to navigate the pixelated corridors of Doom. The performance, by all accounts, is terrible. The cells fumble through hallways, fire wildly at nothing, and generally play like someone who's never held a controller. Which, to be fair, they haven't.
But here's the thing: they're playing at all.
This is Parasma's pitch. Not the gaming prowess of disembodied brain matter, but the energy efficiency. The Y Combinator-backed startup, which emerged in the summer of 2026, is making a bet that sounds either visionary or preposterous depending on who you ask: that living neurons, trained to compute, might solve artificial intelligence's escalating power problem better than the next generation of silicon ever could.
Sean Cole, the founder, holds a master's from the University of Sussex's AI program and apparently made neurons play Doom before launching a company around the concept. His background is academic, but his timing is commercial. Because while most of the tech industry has been racing to build bigger data centers and more efficient chips, a smaller cohort of researchers has been asking a different question entirely: what if the substrate itself is wrong?
On the same day Parasma announced its YC backing, the company published an ethics page. Not the usual platitudes about responsible AI, but a genuine philosophical inquiry: "Does training cultured neurons to play Doom cause suffering?" Cole walks through his reasoning methodically—200,000 neurons lack the recurrent connectivity, the hierarchical structure, the integrated information processing that might add up to something like consciousness. Therefore, probably not.
It's an odd document for a seed-stage startup, the kind of thing you'd expect from a university review board. Then again, most startups aren't growing their compute in petri dishes.
The Math Nobody Wants to Do
The backdrop here matters. Data center electricity demand grew approximately 17% in 2025, with projections for continued growth through 2030 according to International Energy Agency analysis. Data centers now eat roughly 2.6% of global electricity, with the United States accounting for 45% of that consumption as of 2024 figures. The IEA expects annual growth around 15% through the end of the decade.
Ireland offers a particularly stark case study. In 2025, data centers consumed 7,663 GWh—23% of the entire country's electricity supply, per data from Ireland's Central Statistics Office. That's nearly equivalent to all residential power use nationwide.
Big Tech's answer has been to spend harder. Cumulative capital expenditure topped $400 billion in 2025. The IEA forecasts a 75% increase for 2026. Industry watchers like TrendForce and Bloomberg NEF peg 2026 guidance somewhere between $710 billion and $830 billion, with installed data center capacity hitting around 155 gigawatts, up 29% year-over-year.
Microsoft's 2025 sustainability report—summarized by TechRadar last July—showed emissions up 25.1%, driven by AI and cloud infrastructure buildout. Google's 2025 environmental disclosure, published the following year, documented substantial electricity increases since 2021, even as the company highlighted TPU efficiency improvements and water replenishment initiatives.
The message is clear enough, even if nobody says it outright. Silicon scaling has hit an infrastructure wall, and that wall is starting to cost real money.
When Lab Experiments Become Computing Paradigms
"Organoid intelligence" entered formal scientific vocabulary in February 2023, when researchers published a roadmap in Frontiers in Science. The proposal: use living human brain cells—either 2D neuronal cultures or 3D brain organoids—as a new computing substrate, distinct from neuromorphic chips that merely mimic biological processes.
By 2022, this had moved beyond theoretical papers. Cortical Labs and Monash University published results in Neuron demonstrating that neurons grown in a dish could learn to play Pong through a closed-loop system. They called it "DishBrain." In December 2023, Indiana University researchers showed "Brainoware" in Nature Electronics, using brain organoids for reservoir computing. The work drew National Science Foundation backing and follow-on grants through 2025.
Then in 2024, a Swiss outfit called FinalSpark launched something commercial: the "Neuroplatform," offering remote access to biocomputing hardware for $500 monthly, according to Scientific American's August 2024 coverage. The platform featured modules of four organoids each, about half a millimeter in size, containing roughly 10,000 neurons, with eight electrodes per organoid. FinalSpark reported organoid survival averaging around 100 days and claimed interest from 34 universities, including Michigan and Berlin's Free University.
Cortical Labs went a different direction, building what it calls the CL1 platform—a "code-deployable biological computer" with a Python SDK and API, detailed in a February 2026 preprint. The company raised $10 million in April 2023 from investors including Horizons Ventures and In-Q-Tel, positioning itself as infrastructure play for an industry that barely exists yet.
The Doom demonstrations in February and March 2026 marked a milestone, though not for the reasons you might expect. The gameplay was uniformly bad—The Guardian, PC Gamer, and Tom's Hardware all covered the demos, with PC Gamer noting the cells "play a lot like a beginner who's never seen a computer—and in fairness, they haven't." What mattered was the tooling: developers could now write Python scripts to translate game state into electrical stimulation patterns, read neuronal responses, and decode those spikes into in-game actions.
The novelty wasn't the biology. It was the API.
Cole's Gambit
Cole's involvement predates Parasma. His YC profile notes he "made human brain cells play Doom" as prior work. His approach, laid out on the company's research page as of June 18, 2026, uses what he describes as "surprise-scaled reinforcement via a silicon critic." The system delivers low-amplitude, biphasic electrical pulses—microampere range, charge-balanced to avoid harm.
The stated mission reads like moonshot brochure copy: build "algorithms and infrastructure that turn living neurons into programmable compute" to "massively reduce the energy strain of AI." FinalSpark's co-CEO Fred Jordan told Scientific American in August 2024 their goal was AI with "100,000 times less energy" consumption. National Geographic, in a July 2025 overview, cited expert projections of potential energy gains ranging from 10^6 to 10^10.
These are aspirations, not benchmarks. As of 2024 reporting, published demonstrations remained at modest neuronal counts—10,000 to 800,000 cells—performing relatively simple tasks. The million-fold energy advantages remain theoretical for complex AI workloads.
Parasma is hiring, and the job posting offers a window into the company's ambitions. The YC listings page advertises a "Founding Scientist" role with compensation that raises eyebrows: $1 million salary plus 1-3% equity, requiring 11-plus years of experience. For a solo-founder, seed-stage startup, it signals either serious funding or serious desperation. Possibly both.
The Reality Behind the Hype
Biological computing runs headlong into some stubborn problems. Organoid manufacturing lacks anything resembling standardization. Cultures are non-stationary, noisy, prone to drift. FinalSpark's 100-day average lifespan was considered state-of-the-art as of 2024 reporting, but that's three months of viability before the substrate simply degrades.
Reproducibility is worse. The NIH announced a dedicated organoid development center in September 2025, specifically to standardize organoid models using AI and robotics and reduce reliance on animal testing. That this infrastructure is only now being built tells you where the field actually stands: early. Very early.
Scale poses another constraint. The largest published demonstrations involve hundreds of thousands of neurons, not millions or billions. Multi-electrode arrays have density limits. Achieving bidirectional input-output at the resolution needed for genuinely complex tasks requires advances that sit at the intersection of neuroscience, materials science, and electrical engineering.
Then there's the question of what "learning" actually means in these systems. TechRadar's March 2026 coverage of the Doom demonstrations noted skepticism about whether neurons were genuinely learning or simply responding to carefully engineered input-output mappings. No independent, peer-reviewed paper had formally quantified learning performance as of mid-2026; most technical details came from the CL API preprint, company documentation, and media coverage.
In other words: demonstrations, not proofs.
Ethics When Your Computer Is Alive
Cole's consciousness reasoning is thorough, perhaps necessarily so when you're zapping brain cells for a living. He argues that a 200,000-neuron system lacks the structural complexity for consciousness—no recurrent connectivity at scale, no hierarchical organization, no integrated information processing. His stimulation parameters are carefully bounded: charge-balanced biphasic pulses in the microampere range, designed to minimize any conceivable harm.
But as capabilities advance, the questions get considerably harder. A May 2024 review in Nature Reviews Bioengineering argued ethics frameworks must evolve alongside technical capability. An August 2025 survey found many tissue donors oppose broad consent for brain-organoid research. The field lacks clear regulatory guardrails beyond baseline ISSCR guidelines from 2021, which essentially suggest existing oversight committees can handle review.
China issued Human Organoid Research Ethical Guidelines in 2025. The U.S. response has been fragmented at best. NIH BRAIN Initiative funding emphasizes NeuroAI integration and includes neuroethics working groups, but there's no unified framework. The FDA's January 6, 2025 draft guidance on AI/ML device software doesn't directly address organoid computing, though it would presumably become relevant if biological systems were deployed in regulated medical applications.
Georgetown Law's Tech Review published an explainer in 2025 outlining the legal landscape. What happens when organoid systems become meaningfully more capable? What are the thresholds for consciousness or suffering? Who bears liability if something goes wrong? These aren't abstract philosophy problems—they're questions any company building toward commercial products will eventually confront.
Other Bets on Post-Silicon Computing
Parasma isn't alone in betting against GPU dominance. Normal Computing raised $50 million in March 2026, led by Samsung Catalyst, for thermodynamic computing chips. Neuromorphic, analog, and photonic approaches are all positioned as energy-efficient alternatives. The broader biocomputing landscape includes work using bacteria and fungi as unconventional substrates—explored in a 2024 Scientific American feature as potentially less ethically fraught than human-derived neurons.
Johns Hopkins maintains an organoid intelligence initiative, with projects at its Applied Physics Laboratory and active neuroethics collaboration. Indiana University continues Brainoware research. In October 2025, UC Santa Cruz received a $1.9 million NSF grant to explore learning and reasoning in organoids as part of the agency's EFRI "BEGIN OI" program.
The infrastructure and funding exist. What remains unclear is whether living substrates will carve out a genuine commercial niche or remain a research curiosity that produces interesting papers but limited products.
The Long Bet
Reviews published in 2024 and 2025 suggest hybrid bio-silicon systems and improved interfaces could open niches where biological substrates offer advantages—energy efficiency, sample efficiency, specific pattern recognition tasks. Nobody credible is predicting wholesale GPU replacement. The question is whether there's a wedge: certain optimization problems, tasks that benefit from biological priors, applications where wetware makes economic sense.
The technical obstacles remain formidable. Standardization, longevity, scalability, reproducibility, and the fundamental difficulty of quantifying what constitutes genuine learning versus sophisticated input-output mapping—all open problems. Scientific American's August 2024 reporting identified lack of standardized manufacturing and high variability as recognized bottlenecks across the field.
Then there are constraints unique to living compute: maintenance requirements, ethical oversight, donor consent protocols, and the basic strangeness of managing biology as infrastructure. These aren't software bugs to be patched. They're inherent to the substrate.
Yet the energy crisis is demonstrably real, the capital expenditure is astronomical, and policy pressure continues mounting. The IEA's 2026 updates reinforce that AI compute growth will remain constrained by power availability, water resources, and infrastructure spending. Alternative paradigms will keep attracting research funding and entrepreneurial attention, whether or not they ultimately deliver.
Parasma sits at an unusual intersection: frontier neuroscience, desperate infrastructure economics, and genuinely unsettled ethical terrain. Cole's neurons playing Doom represent a proof of concept for a control system. Whether that control system can scale from laboratory demonstration to commercial infrastructure is an entirely separate question.
For now, the neurons keep firing. The game keeps running. And the energy meters keep spinning—just fewer of them, if the hypothesis proves out. The bet is that biology, given appropriate algorithmic scaffolding, can outperform silicon where it arguably matters most: efficiency per joule.
Whether that bet pays off may well determine if we're witnessing the emergence of a new computing paradigm or simply an expensive experiment that taught us a bit more about neurons and perhaps a bit more about our own limitations. Either way, someone's going to learn something. Even if it's just that some problems can't be solved by throwing more compute at them—silicon or otherwise.
