A single rack of advanced AI servers will soon draw as much electricity as 65 homes. According to forecasts from Gartner, AI servers alone could be consuming more power than all conventional data center hardware combined within the next couple of years. The International Energy Agency's projections are equally stark: global data center consumption could reach 1,000 terawatt-hours by 2026, roughly doubling what it was just a few years ago.
Meanwhile, cloud providers are charging around five dollars an hour for access to a single H100 GPU, and power availability—not chip supply, not talent, not capital—has become the binding constraint on AI's expansion.
The math is brutal. Perhaps more brutal than the industry wants to admit.
Which is why a small group of startups has started doing something that sounds lifted from the pages of science fiction: training living human neurons to perform computation. Not as a metaphor. Literally. Growing brain cells in laboratories, wiring them to electrodes, and teaching them to process information as an alternative to silicon.
From Academic Curiosity to Commercial Bet
The field calls itself "organoid intelligence," though some prefer the more visceral term "wetware computing." At its core, it involves computation performed by living biological substrates—typically neurons grown from stem cells and interfaced through multi-electrode arrays. The scientific groundwork was laid gradually, then suddenly.
In December 2022, researchers published results in Neuron showing that lab-grown neurons could play Pong through a system they called DishBrain. A year later, a Nature Electronics paper demonstrated what the authors termed "Brainoware," using brain organoids as reservoir computers capable of speech recognition and time-series prediction. By early this year, UC Santa Cruz researchers had trained brain organoids to solve the cart-pole inverted pendulum control problem—a classic robotics challenge—improving their success rate from 4.5 percent (essentially random chance) to 46 percent through adaptive reinforcement learning.
The neurons learned. They also forgot after rest periods, a reminder that biological systems trade plasticity for stability in ways silicon does not. But the principle was proven.
Now the technology is migrating from academic labs into commercial development, carried forward by a handful of startups with varying approaches to the same ambitious goal: harnessing the extreme energy efficiency of biological neurons, which outperform digital transistors by orders of magnitude per operation. Whether that efficiency advantage can survive the journey from petri dish to production environment remains an open question.
The market timing appears anything but accidental. Data from the IEA shows data center electricity demand grew roughly 17 percent in 2025. Gartner's analysts forecast global consumption climbing from 447 TWh in 2025 to 565 TWh in 2026—a pace that makes utility executives nervous and CFOs even more so. The grid constraints are real. The cost pressures are mounting. And silicon-based solutions, whether traditional or neuromorphic, aren't closing the gap fast enough to matter.
Why Now?
Three forces have converged to make biological computing viable in 2026 rather than 2036.
First, the technical infrastructure has quietly matured. Multi-electrode arrays capable of sub-millisecond closed-loop stimulation and recording exist on the market. Python APIs allow developers to interface with living neurons as if they were just another cloud compute resource. FinalSpark launched its Neuroplatform in 2024, offering remote access to living neuronal cultures. Cortical Labs commercialized its CL1 system last year with a Python SDK and what it calls a "Cortical Cloud" for remote experiments. The plumbing is in place.
Second, stem cell science has advanced to the point where human neurons can be derived reliably from induced pluripotent stem cells. Cortical Labs CEO Hon Weng Chong derived neurons from his own blood for the company's demonstrations—a detail that somehow makes the whole enterprise feel both more legitimate and slightly unsettling. These aren't mouse neurons or cell lines of uncertain provenance. They're human brain cells, grown to specification.
Third, demonstration projects have provided tangible proof points that capture public imagination. In March, Sean Cole, an independent developer fresh off an MSc in AI from the University of Sussex, spent about a week getting Cortical Labs' neurons to "play" Doom. The system used roughly 200,000 neurons on a multi-electrode array. Game states were encoded as stimulation patterns; neural responses were decoded back into game actions.
The Guardian covered it. The technical community paid attention. The performance was rudimentary—think early arcade games, not esports—but the fact that it worked at all was the point. Living neurons, derived from human cells, were controlling the movements of a digital avatar through a pixelated hellscape. It's the sort of milestone that makes you pause.
Around the same time, Cortical Labs announced a partnership with data center operator DayOne to prototype "biological data centers" in Melbourne and Singapore. Princeton researchers unveiled a 3D device integrating living brain cells and electronics, emphasizing the million-fold power efficiency gap between brains and AI systems. Northwestern demonstrated what it called "printed neurons" that communicate with living brain cells, aimed at better bio-electronics interfaces.
The technical scaffolding, in other words, is in place. What's missing is the next layer: algorithms that can train and control these biological substrates reliably.
The Startups Making the Bet
Parasma represents the newest entrant and perhaps the most explicit bet on this algorithmic gap. The company joined Y Combinator's Summer 2026 batch with a stated mission to "train human brain cells for AI compute." Its founder is Sean Cole—the same developer behind the Doom demonstration—and he describes the company as building "bio-ML algorithms for training brain cells" to achieve orders-of-magnitude more energy-efficient computation. In a LinkedIn post announcing the YC acceptance, Cole thanked Cortical Labs and University of Sussex researchers Adam Barrett and James Knight, suggesting the venture has academic roots.
No public funding amounts or customer pilots have been disclosed. The company is pre-seed, pre-product, perhaps pre-clarity on what the business model even looks like. But the thesis is straightforward: if wetware is going to compete with silicon, someone needs to write the software layer that makes biological neural networks trainable at scale.
Cortical Labs remains the most visible commercial player. Founded in 2019, the Australian startup raised $10 million in April 2023 from investors including Horizons Ventures, Blackbird, In-Q-Tel, Radar, and LifeX. Its CL1 hardware platform, launched commercially last year, provides closed-loop stimulation and recording with declarative Python APIs—the kind of developer-friendly tooling that suggests the company is serious about building a platform, not just publishing papers.
The Doom demo in early 2026 generated widespread press, though both Chong and Cole were careful to note that the neurons weren't "conscious" and their learning was limited. Medical applications remain a long-term goal, they said, but one gets the sense they're being diplomatic. For now, the focus is compute.
The DayOne partnership, announced in March, represents the first serious attempt to explore how biological computing might fit into actual data center infrastructure. Details on performance, scalability, and economics remain sparse—understandably, given the experimental nature—but the fact that a data center operator is prototyping with living neurons suggests the technology has crossed some threshold of credibility. Or perhaps it's a sign of how desperate the industry is becoming for alternatives.
FinalSpark, a Swiss company founded in 2014, offers a different model: low-cost research access. Its Neuroplatform, launched publicly in 2024, allows remote experiments on living neuronal cultures. A report from LiveScience cited rental access at $500 per month, positioning it as infrastructure for academic and early-stage commercial research. The company published a peer-reviewed platform paper in Frontiers in AI detailing its experimental architecture. It's the picks-and-shovels play in a market that doesn't quite exist yet.
The Biological Computing Co. made the biggest funding splash: $25 million in seed funding announced on February 12, 2026. Co-founded by neurosurgeon-scientists Alex Ksendzovsky and Jon Pomeraniec—both with MD-PhDs, both with the kind of credentials that make investors lean forward—TBC is based in San Francisco's Mission Bay and positions itself as building a "neuron-based alternative to silicon" for computer vision, generative video, and AI infrastructure. The company describes an encode-decode pipeline that maps neuronal responses into AI models. Strategic advisors from big tech and academic networks were announced shortly after.
Intactis Bio, a Utah-based startup, demonstrated "Hello, World!" and matrix math operations on an automated biohybrid platform in March. The company raised $250,000 from Nucleus Fund and describes its goal as colocating "biocomputers" with energy-constrained data centers. It's the earliest-stage of the cohort, but the technical milestone of basic computation on living neurons represents another proof point, however modest.
Across these companies, a pattern emerges: iPSC-derived human neurons, multi-electrode arrays for interfacing, Python SDKs or APIs for developer access, and positioning around energy efficiency as the killer application. The specific use cases vary—generic compute, computer vision, reservoir computing—but the underlying bet is identical.
What Could Go Wrong
The challenges are substantial. And non-negotiable.
Stability and reproducibility remain unsolved. The UCSC cart-pole study showed that organoids learn but then forget after rest periods. Researchers concluded that larger, more complex organoids and improved protocols would be necessary for retention and scalability. If a wetware processor loses its training weights every time you power it down, it's not yet competitive with silicon, no matter how energy-efficient it is during operation. That's not a software problem. That's a fundamental biological limitation that may or may not be surmountable.
Interface and measurement bottlenecks persist. Higher-density electrode arrays, more robust stimulation protocols, and standardized encode-decode toolchains are all works in progress. Cortical Labs' CL API and the Brainoware reservoir frameworks represent early attempts, but the field lacks the equivalent of CUDA or PyTorch for biological neural networks. Building that standard will take time, coordination, and likely a few false starts.
Ethics and regulation loom large, though perhaps less ominously than one might expect. The International Society for Stem Cell Research's updated guidelines, posted in August 2025, caution against overclaiming cognitive attributes of organoids and recommend appropriate institutional oversight and informed consent. The NIH BRAIN Initiative's "BRAIN 2.0" documents emphasize emerging ethical questions around engineered neural circuits as organoids gain sophistication through vascularization and multi-region linking.
No wetware-specific statutory framework exists beyond existing human tissue, institutional review board, and ISSCR guidelines. The EU AI Act, finalized in June 2024, regulates AI by use-case and risk level rather than substrate, so it would apply to biological computing systems used for certain applications, but it doesn't address the substrate itself. Which is to say: the regulatory landscape is ambiguous, but not hostile.
Public attitudes appear cautiously supportive. An analysis published this spring notes persistent ethical concerns but finds that support for organoid research often remains robust regardless of whether applications are framed as information technology or medical. People seem able to hold two thoughts simultaneously: this is weird, and this might be necessary.
Performance metrics remain frustratingly vague. The field has no standard benchmarks for wetware systems. Published demonstrations like Doom, cart-pole, and reservoir speech tasks show feasibility, not performance parity with GPUs or even neuromorphic chips. Cortical Labs claims its neurons operate at orders-of-magnitude better energy efficiency than silicon, but without standardized benchmarks, comparisons are difficult. It's a bit like comparing early transistors to vacuum tubes—technically true, contextually incomplete.
The Tailwinds Are Real
Yet the forces pushing this technology forward are undeniable.
IEA scenarios project data center electricity demand could reach 945 TWh by 2030, nearly double recent levels. The base case has data centers consuming roughly 3 percent of global electricity by the end of the decade, up from 1.5 percent just a few years ago. AI-centric facilities are already drawing more than 100 megawatts. Some are being designed for 300 MW or more. The power constraints aren't theoretical. They're forcing the industry to consider alternatives that would have seemed fringe not long ago.
Biological computing won't replace silicon in the near term. The technology is too early, too unstable, too poorly understood. But as a complementary compute layer for specific tasks where energy efficiency matters more than raw throughput? As a research platform for understanding intelligence itself? As a hedge against the physics limits of Moore's Law? The case becomes more plausible.
A Nature Computational Science review published in July traced the trajectory from neuromorphic algorithms to organoid intelligence, describing biofeedback-driven computation as an emerging path. A Nature Communications paper from March 2023 titled "Catalyzing next-gen AI through NeuroAI" identified energy efficiency as a grand challenge and argued that understanding neural circuit computation could inform new AI systems.
The field is no longer purely academic. Venture capital has entered. Y Combinator has backed a startup. Data center operators are prototyping. The question isn't whether biological computing will work in principle—the demonstrations have answered that, at least provisionally. The question is whether it can scale, stabilize, and deliver on the energy efficiency promise at a pace that matches AI's voracious demand for compute.
The startups racing to answer that question are betting that data centers consuming the power of small cities will eventually be forced to consider growing their processors instead of fabricating them. It's a radical proposition. But then again, so was the idea of training trillion-parameter models on datasets scraped from the entire internet.
Radical only sounds impossible until it doesn't.
