The visual quality wasn't great. The gameplay was worse. But in a lab somewhere in Australia, a clump of human brain cells—about 200,000 neurons cultured on a silicon chip—was playing Doom.
"Like a beginner," one description generously offered. The marine lurched through pixelated corridors. Occasionally, it fired at demons. The cells weren't conscious, not in any meaningful sense. They were responding to electrical patterns translated from the game's visual feed, and their firing patterns—along with external algorithms—were being decoded back into movement commands. A closed loop between living tissue and virtual hellscape.
That was in February. The demonstration emerged from Cortical Labs, an Australian startup, working with Sean Cole, an AI researcher who would shortly launch his own company in this nascent space. No peer-reviewed document was available for the Doom demo by July 2026. But it spread. By March, the Doom clip had ricocheted through tech circles, provoking the kind of reactions biological computing tends to trigger: fascination mixed with unease, skepticism laced with curiosity.
Here's what made it more than a stunt: For the first time, living human neural tissue was interacting with a real-time virtual environment in something that looked, however crudely, like purposeful behavior. Not intelligence, exactly. But not randomness either.
Which raises an interesting question, given the trajectory of AI infrastructure. What if the future of computing isn't smaller transistors or clever chip architectures? What if it's wet?
The Grid Can't Keep Up
The timing of all this isn't coincidental, though perhaps it's more opportunistic than the founders might admit.
Data centers are developing an appetite problem. According to International Energy Agency projections from their reporting cycle covering this period, data center electricity consumption could hit 950 terawatt-hours by 2030—roughly double recent levels, representing something close to 3% of global electricity demand. AI is the culprit. Industry forecasts cited in mid-year trade coverage estimated consumption around 565 TWh for a recent annual period, up more than a quarter year-over-year. Grid constraints, multiple analysts warned, were coming.
Every ChatGPT query burns watts. Every image generation. Every training run adding up to an infrastructure headache that's starting to look intractable through conventional means.
Enter biological computing, where the pitch is almost too neat: neurons run on electrochemical gradients at power levels that make silicon transistors look wasteful by comparison. FinalSpark, a Swiss outfit that launched a remote "wetware" platform back in spring 2024, claimed its organoids consumed "a million times less power" than digital chips.
A company claim, not an independently verified benchmark. Take it with appropriate skepticism. But the underlying physics isn't wrong—biological systems evolved to be thermodynamically efficient in ways that semiconductor fabrication plants never optimized for.
The harder question: can neurons actually compute at scale, or is this just highly efficient noise?
From Pong to Cart-Pole to Doom
The academic lineage here is surprisingly short, which tells you how new this all is.
December 2022: A team publishes results in Neuron demonstrating that cultured neurons in a dish—they called it "DishBrain"—could learn to play Pong within minutes using closed-loop electrical feedback. Electrode arrays stimulated the cells and recorded their responses. The neurons learned to keep the ball in play. Simple, but deliberate.
December 2023: Another group shows "Brainoware" reservoir computing in Nature Electronics, using brain organoids for speech recognition and solving nonlinear equations. The organoids functioned as adaptive computational reservoirs—processing input through their internal dynamics, requiring only lightweight training of output mappings.
February 2024: UC Santa Cruz researchers publish in Cell Reports showing cortical organoids performing goal-directed learning in a cart-pole balancing task. Through reinforcement-learning-style "coaching," success rates climbed from 4.5% (random baseline) to 46%. Not reliable, but definitely not random.
Three papers in just over a year. Not a mature field. But enough to suggest the technology might be crossing a threshold from pure research into something that could conceivably be productized.
Which, naturally, is where the startups come in.
The Commercial Gambit
Cortical Labs moved first. By early 2025, they were selling the CL1 "biological computer"—a self-contained unit housing roughly 200,000 human neurons on a multi-electrode array, complete with life-support systems rated for six months. Trade press reported pricing around $35,000 per unit, with weekly rentals available for approximately $300. They also launched Cortical Cloud, an API for remote experimentation.
The Doom demo ran on this infrastructure.
On March 11, 2026, Cortical Labs announced a partnership with DayOne to deploy CL1 units in data center configurations. An initial Australian facility would house 120 units. Plans called for scaling to 1,000 units in Singapore. The pitch: ultra-low-power biological co-processors handling niche workloads alongside conventional servers.
FinalSpark, meanwhile, took a different tack. Rather than selling hardware, the Swiss startup offered remote access to living organoids through a Python API. Their Neuroplatform—documented in a Frontiers in Artificial Intelligence paper published in May 2024—featured 16 organoids available for 24/7 experiments. Multi-electrode array control, extensive logging. They'd accumulated over 18 terabytes of data across three years of operations (later claiming north of 30 TB on their website). By May 2024, interest from three dozen universities.
Then in mid-February came The Biological Computing Co., founded by two neurosurgeons who raised $25 million in a seed round. Their vision, laid out in a Fortune interview: integrate living neuron networks with foundation models to slash cost and energy consumption in AI inference. Timeline: 5-10 years to deploy into live compute infrastructure.
Ambitious doesn't quite capture it.
And in June, Parasma emerged from Y Combinator's summer batch. Founder: Sean Cole, the same researcher behind the Doom collaboration. Mission, per the YC directory: "algorithms and infrastructure that turn living neurons into programmable compute."
Team size: one.
What Can They Actually Do?
Let's be clear about the gap between these demonstrations and replacing a GPU rack. It's enormous.
The UC Santa Cruz cart-pole experiment—arguably the most rigorous to date—required careful tuning of stimulation protocols and feedback timing. Performance was probabilistic. The organoids showed learning, yes. Reliability? Not so much.
The Doom demo raises even more questions. Coverage at the time noted uncertainty about how much "intelligence" resided in the neurons versus the encoding and decoding algorithms running on silicon. Military.com, in a March write-up, pointed out the absence of peer-reviewed documentation months after the video circulated. The neurons might genuinely be playing Doom. Or they might just be reacting to patterns while software handles the heavy conceptual lifting.
What seems clearer is that biological neural networks excel at specific types of computation—pattern recognition, adaptation, parallel processing that doesn't map cleanly onto von Neumann architectures. A mid-year review in Nature Computational Science positioned organoid intelligence as "biohybrid computation," a middle ground leveraging strengths of both wetware and silicon.
The technical challenges? Considerable. Organoids need nutrients, oxygen, stable temperature. They drift. They're contamination-sensitive. Finite lifespans measured in months, not years. Standardization barely exists—each organoid is slightly different, which is a nightmare for reproducibility. Long-term memory retention remains an open question.
You can see why data center operators haven't exactly rushed to embrace this.
The Data Center Bet
Cortical Labs' partnership with DayOne represents the most aggressive commercial deployment yet—assuming it proceeds as planned.
The company claims its CL1 units consume less power than a handheld calculator. That figure appears in press coverage, though independent validation is absent. Even if accurate, the computational workload that power supports remains unclear. Useful work, or just baseline metabolic activity keeping the cells alive?
The initial Australian facility with 120 units is a pilot. The planned Singapore expansion to 1,000 units would constitute a meaningful scale test, though still minuscule compared to conventional data centers.
The question isn't whether neurons consume less power per neuron—they almost certainly do. It's whether they can deliver enough computational throughput to justify the complexity of keeping living cells alive in production environments.
One possibility: niche applications where biological neural networks offer advantages in power efficiency or adaptability even if raw throughput lags silicon. Sensory processing. Anomaly detection. Tasks that play to biological strengths.
The Biological Computing Co. founders, in their February Fortune interview, sketched a 5-10 year timeline for integrating neuron networks into "real-time compute circuits assisting data centers," with 10-20 years for deeper integration.
That's a very long runway. Whether investors have that kind of patience remains to be seen.
Government Money Enters the Picture
Institutional support is materializing, which lends the field a certain legitimacy.
In 2024, the National Science Foundation launched its EFRI "BEGIN OI" program—Bioengineered Systems for Ethical Biocomputing and Organoid Intelligence—awarding $14 million to seven projects in October 2024. The solicitation explicitly called for research bridging organoid neurobiology and computational applications.
More telling: DARPA announced its "O-Circuit" program in May, calling for proposals to develop "a new class of biologically inspired computers centered on organoid cytomorphic intelligence." Proposers' Day was held in April.
DARPA programs often signal where defense and intelligence agencies see future strategic advantage. Their interest suggests this isn't entirely academic fantasy.
Academic consortia are coalescing too. Johns Hopkins hosts an organoid intelligence hub that published the "Baltimore Declaration" roadmap in a February 2023 Frontiers in Science paper, formalizing the term "organoid intelligence" and outlining technical and ethical tracks. UC Santa Cruz's Braingeneers group produced the cart-pole work. Princeton researchers published in April Nature Electronics on 3D micro-instrumented neural network devices for computing applications.
Research is accelerating. Though multiple reviewers caution that many claims remain "demo-grade"—not directly comparable to silicon benchmarks. Which is a polite way of saying: impressive in principle, unproven at scale.
The Consciousness Problem
Here's where the conversation gets uncomfortable. Brain organoids are clusters of human neurons. They're not brains—they lack the scale, organization, connectivity of even primitive nervous systems.
But as they grow more complex, ethical questions intensify.
Parasma published a detailed research note in mid-June—the same day it announced Y Combinator backing—addressing "consciousness and suffering." The note argued that the ~200,000-neuron cultures used in the Doom demo lacked "global integration, nociception, embodiment, or self-model" necessary for meaningful consciousness. Stimulation protocols were described as "conservative, charge-balanced," designed to avoid damage. The feedback loop ran through a silicon "critic" scaling reinforcement by prediction error, not raw punishment or reward.
It reads like preemptive ethics documentation. The kind of thing you publish when you know the questions are coming and want to get ahead of them.
The International Society for Stem Cell Research updated guidelines in August 2025 (version 1.2), treating brain organoids under general organoid research guidance with specialized review as needed. A 2021 National Academies report highlighted governance gaps for neural tissues, chimeras, and organoids.
The academic debate through the recent period emphasizes donor consent, moral status thresholds, and a concern that commercial hype could damage the broader biomedical research ecosystem. Organoid models are critical for disease study. If public backlash forces restrictions, legitimate neuroscience suffers.
STAT reported in November 2025 that some organoid pioneers feared exactly that—that "organoid intelligence" marketing could jeopardize funding and public support for their work. Overpromising on biocomputing triggers restrictions that constrain basic research.
It's a valid concern, and not one these startups seem particularly worried about navigating.
What Happens Next
The trajectory from here is genuinely uncertain.
Parasma, at team size one, is at the earliest possible stage—an idea, a demonstration, Y Combinator validation. The Biological Computing Co. has $25 million and a decade-long vision. Cortical Labs has actual deployment plans in motion, though at pilot scale. FinalSpark has operational experience accumulated over two-plus years.
Technical gaps remain substantial: stability, scalability, standardization, memory retention, reproducibility. A mid-year Nature Computational Science review catalogued these constraints alongside the field's promise, positioning organoid intelligence as "the new frontier beyond neuromorphic chips" while acknowledging most work remains exploratory.
What makes the space compelling—perhaps more than founders might admit—is the confluence of pressures. AI energy consumption is a genuine infrastructure problem. Neuromorphic chips and other silicon-based alternatives haven't delivered promised efficiency gains at scale. And neurons demonstrably compute, adapt, learn while consuming negligible power.
Whether that translates into infrastructure? Speculative.
But DARPA is funding proposals. Data centers are being planned, if not yet built. Y Combinator is backing founders. The speculation is being taken seriously, at least by some institutional actors with resources to deploy.
The race isn't to replace silicon. That's not realistic, at least not on any timeline that matters for current investment horizons. The race is to identify workloads where biology offers enough of an edge—in power, adaptability, or some capability silicon can't easily replicate—to justify the considerable complexity of keeping neurons alive in production.
The cells are playing Doom, stumbling through corridors, occasionally shooting demons. Whether they can do anything more useful than that—and whether anyone can build a sustainable business around it—remains very much an open question.
But someone's going to try. In fact, several someones already are.
