
led by Sean Cole
YC S26 startup training human brain cells for compute, as an alternative to GPUs.
Parasma trains human neurons to do compute and says its cells beat a linear silicon decoder on one hard task. The team is two people and the startup sits in Y Combinator's summer 2026 batch. It has no public product, API or price.
| Product | What it is |
|---|---|
| Next-token prediction demo | Language task run on a cultured neuron sample |
| Reinforcement learning demo | Control task inside a changing environment |
Part of Y Combinator's summer 2026 batch, Parasma is a biocomputing startup that trains human neurons to perform computation. Sentences become electrical stimulation, the cells' responses are decoded into tokens, and the system feeds back whether each guess was right. It has also shown a reinforcement-learning demo. Parasma sells no product yet, and there is no public API or price.
In late February 2026 founder Sean Cole worked with Cortical Labs to train human neurons to play Doom. Parasma's August launch release then claimed 90% accuracy on defined next-token tasks, with randomized controls consistent with chance. One isolated run on a harder non-linear task reached 78.1%, above the 75% ceiling of a linear silicon decoder. Nobody outside the company has replicated these figures.
Sean Cole founded Parasma and is its CEO. Y Combinator's profile notes his MS in Artificial Intelligence and lists a team of two. He describes earlier work on neuromorphic computing and reinforcement learning, and the profile shows an open Founding Scientist role, based in San Francisco or remote. Parasma has published no leadership page, board or advisory list.
Parasma's confirmed backing is Y Combinator, announced on the company's site on 18 June 2026. The company has not disclosed an amount, a valuation or any other investor. Some third-party databases show a small seed figure, but Hokai could not confirm it with the company or Y Combinator, so it is not stated here. No revenue or customer contracts have been reported.
Parasma has published no donor-consent, ethics-review or safety documentation, which is the main open question for any company working with human neural tissue. It has no compliance certifications or trust page. Its closest competitors are Cortical Labs, maker of the CL1 biological computer, and FinalSpark, which offers remote access to living neurons for researchers. Both already give some outside access to their platforms; Parasma does not yet.
Founded by Sean Cole, Parasma reports that cultured human neurons predicted most next tokens correctly on defined tasks, in a company release from August 2026. Randomized controls scored near chance. Parasma has no API, no pricing and no disclosed funding amount beyond Y Combinator backing.
Parasma is a biocomputing startup in San Francisco that trains human brain cells to do computational work, and it is part of Y Combinator's Summer 2026 batch. The founder and CEO is Sean Cole. The work first became public in late February 2026, when Cole collaborated with Cortical Labs to train a culture of human neurons to play Doom, a result Parasma's press release describes as the start of the project. Parasma announced its public launch on 13 August 2026, and its site now describes the pitch in one line: it trains brain cells for compute, as an energy-efficient alternative to GPUs.
The company is not building the biological hardware. Its stated focus is the algorithms and infrastructure that stimulate a culture of neurons, read the responses back and feed back whether each answer was right. In the launch release, a sentence is converted into electrical stimulation, the cells' responses are decoded directly into tokens, and no GPU sits in that loop. Parasma reports that cells correctly predicted 90% of next-token decisions on precisely defined tasks while taking in the full sentence context, and that randomized control runs scored significantly worse, consistent with chance. On a harder, non-linear task where no single context token determines the answer, one isolated run reached 78.1% accuracy against the 75% ceiling that a linear silicon decoder could achieve. These are the company's own figures from one announcement, not independent replications.
Parasma's second public demo is reinforcement learning: each frame of an environment becomes a stimulation pattern, the neurons select an action, and the result is measured and fed back. The near-term plan in the release is repeatable benchmarks across cultures, devices and time, plus richer contextual prediction and adaptive-control tasks. The long-term claim is that programmable biological compute could replace silicon where continual learning or extreme energy and sample efficiency matter. That is a goal, not a shipped product: there is no public API, hosted platform, pricing or customer announced.
Funding is thin on the record. Y Combinator's profile lists Parasma as active with a team of two, and the company posted that it is now backed by Y Combinator on 18 June 2026. No round size, valuation or other investor has been confirmed by the company, so treat any third-party dollar figure as unverified. The YC listing also shows an open Founding Scientist role in San Francisco or remote, which signals where the team is heading: more wet-lab and neuroscience capacity. For a sense of the silicon this approach wants to displace, compare the inference-chip makers Groq and SambaNova, the GPU incumbent NVIDIA and the GPU cloud CoreWeave. Another YC-backed company on Hokai is Leaping AI, and LiteLLM is a YC tool in the gateway category. The robotics lab Physical Intelligence is the closest Hokai profile to Parasma's stated robotics ambition.
Where Parasma is weaker than the alternatives: it has no peer-reviewed paper of its own, no independent benchmark, and no product anyone outside the company can use, while Cortical Labs and FinalSpark already offer some form of research access to biological compute. It also depends on neuron cultures that drift over time, which is exactly why the company says it is working on repeatability. Parasma has published no donor-consent, IRB or bioethics documentation that Hokai could find, which matters when the substrate is human tissue. Anyone who needs working compute today should use silicon; the useful way to follow Parasma is as a research signal. To compare AI companies by what they ship, browse the directory or use Smart Match.
Everything on this page (funding, leadership, compliance posture, the product list) is read from Parasma's own pages and filings, or from the Hokai records for the products it ships. A field the company does not publish is left off the page rather than guessed at.