Last year, data centers worldwide burned through 415 terawatt-hours of electricity. The International Energy Agency warns that by decade's end, that figure could more than double to 950 TWh, claiming about 3% of all electricity generated globally. In Ireland, the numbers have already turned surreal: data centers were projected to consume 23% of the national grid by 2025, nearly matching the power draw of every household in the country.
The culprit, as anyone tracking tech infrastructure knows, is AI. And the trajectory is unsustainable.
Which helps explain why a small cadre of startups and academic labs are now pursuing what sounds less like engineering than biological alchemy: training living human neurons—actual brain cells grown in lab dishes—to perform computational work. The pitch is seductive in its simplicity. Your brain runs on about 20 watts while executing pattern recognition, contextual reasoning, and adaptive learning tasks that would demand kilowatts from even the most advanced silicon. Harness a fraction of that efficiency, the thinking goes, and you might rewrite the energy economics of computing itself.
Of course, the distance between a compelling pitch and a functional product tends to be measured in years of unglamorous troubleshooting. And in this case, the hurdles are technical, ethical, and scientific all at once.
Where the Science Stands Now
Biological computing occupies an unusual corner of the research landscape—part neuroscience, part bioengineering, part machine learning. The field has moved through distinct phases: symbolic AI in the early decades, neuromorphic silicon chips designed to mimic neural architectures more recently, and now what researchers call "biohybrid" systems that incorporate actual living neural tissue. A July 2026 perspective published in Nature Computational Science traced this evolution and identified biohybrid computers as the current frontier, albeit one with significant unresolved challenges around interfacing, stability, and reproducibility.
Some of the recent demonstrations have been genuinely striking. In February 2026, a team at UC Santa Cruz published results in Cell Reports showing that brain organoids—millimeter-scale clusters of lab-grown neurons—could learn to balance an inverted pendulum in a simulated cart-pole task. With adaptive training protocols, these organoids achieved success rates of 46%, up from 4.5% in control conditions. A month earlier, Melbourne-based Cortical Labs showed off its CL1 system running a stripped-down version of the video game Doom, using cultures of 200,000 to 800,000 living human neurons interfaced through high-density microelectrode arrays.
These aren't parlor tricks, exactly. They're proofs of concept that living neural tissue can be trained to perform goal-directed tasks through closed-loop feedback—essentially the same reinforcement learning principles underlying modern AI systems. But they're also very far from replacing a GPU cluster. The organoids live for months, not years. Performance varies wildly from one batch to the next. And no one has yet demonstrated that these systems can scale beyond highly controlled lab environments.
Still, the fact that neurons can learn to play simple games at all suggests something worth investigating.
Why This, Why Now
The energy problem is accelerating, not stabilizing. In June 2026, Gartner projected that data center electricity demand would climb 26% year-over-year to reach 565 TWh. AI-optimized servers were forecast to account for 31% of data center power consumption in 2026, surpassing conventional servers by 2027. Meanwhile, major technology companies poured more than $400 billion into AI infrastructure in 2025, with spending expected to jump another 75% this year. Some firms have gone so far as to sign offtake agreements for 45 gigawatts of power from small modular nuclear reactors—a hedge against future grid constraints.
Against that backdrop, the human brain's 20-watt operating budget looks almost taunting. Pattern recognition, contextual reasoning, adaptive learning—all for less power than a compact fluorescent bulb. If computation could run on biological substrates at even remotely comparable efficiency, the cost savings would be measured in orders of magnitude.
Government agencies have begun placing serious bets. In August 2024, the National Science Foundation launched its EFRI BEGIN OI program, committing $14 million across seven projects aimed at advancing "organoid intelligence" with embedded ethical oversight. In May 2026, DARPA announced its O-Circuit program—Organoid Cytomorphic Intelligence—focused on developing ultra-efficient biological processing units. These aren't basic neuroscience grants. They're strategic investments in a computing paradigm that doesn't yet exist at commercial scale.
Perhaps the most telling signal: the money is real, the timelines are vague, and nobody seems entirely sure what success looks like.
The Players
Cortical Labs emerged from foundational academic work published in the journal Neuron back in December 2022, when researchers demonstrated that neurons cultured on microelectrode arrays could learn to play Pong through closed-loop feedback. The company has since commercialized that discovery into the CL1 platform, which networks racks of what it calls "biological computing units" in a pilot facility in Melbourne. In March 2026, Cortical raised funding led by Horizons Ventures and 3C, with participation from Tom Oxley, the founder of brain-computer interface company Synchron. The company is building a manufacturing hub in Malaysia and plans to open a Singapore facility with around 1,000 CL1 units by September 2026, according to earlier company statements.
The CL1 system works by reading neural activity through high-density microelectrode arrays, translating game states into electrical stimulation patterns, and delivering reward or penalty signals based on performance. The Doom demonstration in early 2026 showed neurons learning to navigate a first-person shooter environment—though calling it gameplay might be generous. Complexity was limited. Still, in January 2026, Reply, an Italian IT consultancy, partnered with the University of Milan to study CL1 learning dynamics, energy efficiency, and robustness. Early signs of enterprise interest, however modest.
FinalSpark, a Swiss startup, has taken a different tack: cloud-based access to living neural tissue. In May 2024, the company launched its Neuroplatform, a remote research infrastructure that provides institutions with API and Jupyter notebook access to brain organoids cultured on microelectrode arrays. A paper published in Frontiers in Artificial Intelligence detailed the architecture and positioned the platform as shared infrastructure for wetware computing experiments. Nine institutions received free access at launch, with 16 organoids initially available. The platform reportedly scaled to 1,000 organoids by February 2026, though that figure is based on media reports and lacks peer-reviewed verification. FinalSpark's pitch is democratization: researchers can run experiments on living neurons without the overhead of maintaining their own wetlab.
On the academic side, UC Santa Cruz's Braingeneers program represents the research vanguard. The cart-pole study published in Cell Reports in February 2026 showed that brain organoids could learn a control task through high-frequency training and adaptive coaching signals. The team received a $1.9 million NSF EFRI grant in October 2025 to scale parallel organoid experiments. In June 2026, genomics pioneer David Haussler, speaking to UCSC News, described the goal as using AI and organoids together "to understand how our brains actually work"—positioning biological computing less as a product opportunity and more as a research tool for decoding neural function.
Then there's Parasma, a San Francisco-based startup that joined Y Combinator's Summer 2026 cohort. Founded by Sean Cole, Parasma isn't building hardware at all. Its focus is the software layer: algorithms that train living neurons to perform useful computation. The company's website is admirably blunt: "We write the algorithms to turn brain cells into compute." Parasma cites internal experiments involving Doom-playing neural cultures and notes collaborations with Cortical Labs and the University of Sussex. The team consists of one person at present. Parasma is actively recruiting a founding scientist to design closed-loop training experiments, reward structures, and evaluation metrics. Funding beyond YC participation hasn't been disclosed. Neither has product pricing or customer pilots. It's very, very early.
The Hard Parts
For all the recent momentum, biological computing is awash in unresolved problems. Reproducibility may be the most vexing. Brain organoids exhibit donor-to-donor variability, non-stationarity under sustained stimulation, and limited lifespans—typically measured in months. Standardization across labs remains a work in progress, though the Organoid Standards Initiative published manufacturing and quality control guidance in May 2024.
Interfacing presents another bottleneck. High-density microelectrode arrays offer read/write access to neural tissue, but closed-loop latency, signal-to-noise ratios, and limited observability constrain what can be achieved. These systems lack the long-range connectivity and architectural complexity of even small animal brains, let alone human ones. Life support and bioreactor engineering introduce layers of operational complexity that silicon never has to contend with. You can't just reboot a neuron culture.
And then there are the ethical questions, which are harder to dismiss than you might expect. In May 2026, the UK's Nuffield Council on Bioethics published a report calling for stronger governance, improved consent procedures, updated Home Office guidance, and centralized monitoring of neural organoid research. The report flagged gaps in existing human and animal research frameworks, particularly around consciousness and suffering thresholds. Parasma responded with its own ethics statement in June 2026, arguing that the ~200,000-neuron cultures used in Doom experiments lack the architecture and scale necessary for consciousness, with conservative stimulation protocols and no nociceptors present.
But uncertainty around markers of sentience remains. The scientific community is divided on where to draw the line, and the public conversation tends toward alarm. A November 2025 investigation by STAT quoted leading organoid scientists cautioning against hype and warning of backlash if overclaims proliferate. Public trust in bioengineering is fragile. The specter of "conscious computers" could easily trigger regulatory overreach that stifles legitimate research for years.
What Happens Next
Biological computing is not a near-term replacement for silicon—let's be clear about that. The July 2026 Nature Computational Science perspective emphasized that interface challenges, stability issues, and ethical considerations will likely remain gating factors for years. But the trajectory is real enough to warrant attention. Government funding through NSF and DARPA signals scientific legitimacy. Academic labs are publishing reproducible benchmarks. Startups are raising capital and building platforms, however tentative.
And the energy crisis isn't going away. Gartner expects AI servers to dominate data center power consumption by 2027. The IEA projects a doubling of data center electricity demand by 2030. Neuromorphic silicon, analog photonic chips, and other efficiency initiatives are already in-market, but they offer incremental gains—perhaps 20× to 100× improvements in specific workloads. Biological substrates, if they ever scale, represent a different kind of inflection point entirely.
What remains unclear is the timeline and the application. Will wetware computing first find traction in niche use cases like adaptive control systems, pattern recognition, or low-power edge inference? Or will it remain primarily a research tool for understanding neural function, as UCSC's Haussler suggested? Parasma's bet on the algorithm layer is intriguing. If platforms like Cortical Labs' CL1 or FinalSpark's Neuroplatform become standardized substrates, there could be a market for training protocols, reward architectures, and closed-loop optimization frameworks that make biological compute genuinely useful.
The field is also wrestling with its own hype cycle. FinalSpark's early marketing materials claimed "million times less power" than digital systems—a figure that lacks peer-reviewed support and has drawn skepticism from researchers. Cortical Labs has been more measured in its public statements, but the Doom demo still invites comparisons to gaming PCs that aren't remotely fair. The challenge for founders in this space will be managing expectations while delivering incremental, verifiable progress.
For investors and technologists tracking this space, the near-term opportunity is infrastructure: platforms, interfaces, training algorithms, and ethics frameworks that turn today's lab experiments into tomorrow's reproducible systems. The long-term opportunity is harder to quantify but potentially transformative—a computing paradigm that runs on watts instead of kilowatts, scales through biology instead of fabrication capacity, and learns through plasticity instead of backpropagation.
It may take a decade. It may not work at all. But the energy math is forcing the conversation, and the science—strange as it is—is no longer purely theoretical.
