A two-person startup backed by Y Combinator announced Tuesday that it has trained human brain cells to predict the next token in a sequence with 90% accuracy. That's the same fundamental task large language models like GPT perform millions of times per second. More striking, Parasma says the neurons hit 78.1% accuracy on a harder, non-linear version of the problem—beyond what the team described as a 75% ceiling for a conventional linear silicon decoder.
The claim, if it holds under scrutiny, positions lab-grown neural tissue as something more than a research curiosity. It suggests living cells might one day offload certain kinds of computation from the GPU farms that now consume as much electricity as small countries. Whether that day arrives in five years or fifty is the subject of considerable debate among neuroscientists, investors, and the small cohort of biocomputing entrepreneurs trying to make brain cells compute for a living.
Parasma emerged from Y Combinator's Summer 2026 batch with a stated mission to replace GPUs with human neurons, according to its profile on the accelerator's website. Sean Cole, the founder, previously worked with Cortical Labs in Melbourne, where he helped build the first implementation of Doom running on roughly 200,000 living human neurons inside the CL1 biological computer. Tom's Hardware covered that milestone in March 2026. Cole framed Tuesday's token-prediction result as validation of a broader thesis.
"Crossing this boundary that a linear silicon decoder cannot represent is exactly the kind of result we built Parasma to pursue," he said in the August 13 launch announcement. The company did not disclose the size of the dataset, the duration of training, or how many attempts preceded the reported 78.1% run. As of the latest available information, no independent replication has been published.
Power, Memory, and the Case for Neurons
Data-center operators are in a bind. The International Energy Agency projected in its 2024 special report that electricity consumption from data centers, AI workloads, and cryptocurrency mining could roughly double by 2026 compared to 2022 levels. At the same time, high-bandwidth memory supply remains tight. TrendForce warned in June that HBM supply is tight and contract prices are expected to surge through 2027. HBM's share of DRAM wafer input is forecast to hit 30% in 2027, up from 22% this year, as AI server shipments continue growing more than 20% annually.
An eight-GPU Nvidia HGX node can draw upward of 5.6 kilowatts at the system level, according to recent technical documentation. Uptime Institute tallied more than 350 announced data-center campuses rated at 100 megawatts or higher as of 2026, and staffing shortages combined with high-density rack adoption are now among the top operational headaches.
Enter the biological alternative. Neurons are absurdly energy-efficient by comparison—brains run on about 20 watts total. They learn from sparse examples, adapt continuously, and don't need to be retrained from scratch every time you add new information. Those traits appeal to anyone trying to build intelligence without melting the grid.
Parasma automates stem-cell differentiation and tunes electrode spacing to interface with cultured neurons, Cole explained in an August 13 transcript. The company uses microelectrode arrays, grids of tiny sensors and stimulators, to send input signals into the neural culture and read out electrical activity in real time. The technique itself is not new. It was pioneered in academic labs decades ago and commercialized more recently by companies including Cortical Labs, FinalSpark, and The Biological Computing Company.
The innovation, such as it is, lies in applying closed-loop feedback to train the neurons on a task borrowed directly from transformer architectures. Next-token prediction is the core operation of models like GPT: given a string of words, predict what comes next. Parasma's claim is that living neurons can learn this task through reward signals delivered via the electrode array, adjusting synaptic weights and firing patterns until performance improves. The company declined to specify how many neurons it used or the architecture of its culture system.
A Field Taking Shape
Parasma is wading into what researchers call "organoid intelligence," or OI for short. The term describes efforts to use two-dimensional neural cultures or three-dimensional brain organoids, interfaced with microelectrode arrays, to perform computation. A roadmap published in Frontiers in Science in February 2023 laid out the ambition: to make organoids more computer-like, not computers more brain-like. The distinction matters. Brains excel at sample efficiency, energy efficiency, and continual learning. Silicon does not, or at least not yet, and that gap is driving interest in biological substrates.
Academic labs have shown that closed-loop learning in living tissue is possible. Researchers at UC Santa Cruz published results in Cell Reports this past February demonstrating that mouse cortical organoids could learn a cart-pole balancing task through feedback. A team at Indiana University Bloomington reported in Nature Electronics in late 2023 that brain organoids paired with high-density microelectrode arrays could recognize spoken digits and predict nonlinear equations in a reservoir-computing framework, though that work is now more than a year old.
Thomas Hartung, a professor at Johns Hopkins who co-authored the 2023 OI roadmap, told Frontiers in an interview that his dream is "to form a channel of communication between an artificial intelligence program and an OI system that would allow the two to explore each other's capabilities." It is an expansive vision, and one that remains speculative.
The Money and the Government
Venture capital is betting, carefully, that neurons can scale. The Biological Computing Company raised a $25 million seed round in February, led by Primary Ventures, to deploy neuron-based co-processors for computer vision and generative video workloads. Intactis Bio launched public demos in July claiming up to three million times greater energy efficiency per decision than silicon, though that figure is a vendor assertion that lacks independent verification. Several analysts cautioned this year that efficiency claims from biocomputing startups often lack public audits.
The U.S. government is also paying attention. The National Science Foundation announced $14 million across seven projects in 2024 under its EFRI "BEGIN OI" program. Notably, the program required or encouraged embedded ethicists as co-principal investigators, a signal that ethical concerns are not being treated as an afterthought. DARPA issued a solicitation on May 1 for its O-CIRCUIT program, seeking biological processing units for edge learning and inference with milliwatt-hours-per-day power draw.
Cortical Labs shipped its CL1 biological computer in 2025 and began installing racks of the devices in Melbourne in September of that year, according to a March profile in Information Age. On August 17, DayOne Data Centers and the National University of Singapore Medicine announced a "Biological Data Center Prototype" in Singapore. Jamie Khoo, CEO of DayOne, said the market is "responding with new approaches, beyond just bigger builds." DayOne had raised more than $2 billion in a Series C announced in January.
FinalSpark, a Swiss company, launched its Neuroplatform in May 2024, offering remote subscriptions for experiments on human brain organoids. The company cited more than 30 terabytes of datasets in analysis published this year. Intactis Bio said in July that early customers spend more than $20,000 per month on compute and that the company is planning data-center pilots. Cortical Labs raised $10 million in seed funding in 2023 led by Horizons Ventures, with additional investors Gobi Partners and 3C joining in 2026. Total funding stood at around $11 million as of March, according to Data Center Dynamics.
The Skeptics Speak
Not everyone is convinced. Tony Zador, a neuroscientist at Cold Spring Harbor Laboratory, told STAT in a November 2025 report that getting circuits to "wire up to do what we want them to do is completely beyond what we could even conceive of right now." The NSF program's requirement for ethicist co-principal investigators reflects broader unease. Journals have tightened reporting standards for biocomputing research.
Reproducibility is a persistent problem. Brain organoids show high batch-to-batch variability, a challenge documented in Nature papers from December 2024 and August this year. The latter showed that organoids can record the passage of time over multiple years but require extensive quality control. Typical useful culture lifetimes run months. Replacing biological units at scale is an operational cost not yet benchmarked against GPU fleets.
Energy-efficiency claims from vendors, often framed as improvements of several orders of magnitude, require neutral, task-normalized audits that include full life-support overheads. Those audits are not yet common. Ethics questions are surfacing as well. A Nature commentary published in August noted that tissue donors may not realize their cells are being used for biocomputing and called for updated consent frameworks.
What Parasma Does Now
Parasma says it is pursuing revenue through near-term applications while building toward brain-cell compute as a general substrate, according to the August 13 transcript. The company has not disclosed funding beyond its Y Combinator batch, headcount beyond the two founders listed on its profile, or named customers. YC partner Tyler Bosmeny is listed as the primary partner on the batch page.
The field is moving from lab demonstrations to pilot deployments. API access, small racks, public gameplay. But standardized benchmarks comparing biological and silicon systems on narrow tasks like token prediction, control, and noisy-sensor classification are still taking shape. Early customers are likely research labs and deep-tech firms with specific workloads, according to sector reporting from mid-2026.
Hon Weng Chong, founder and CEO of Cortical Labs, told Data Center Dynamics in March that "AI capacity is accelerating faster than most people realize, and everyone is talking about chips, models, and megawatts."
The question, perhaps, is whether anyone will soon be talking about neurons.
