TBC plans to run the model on AWS Trainium chips, serve it through Amazon SageMaker AI, and sell it in the AWS Marketplace. All of that is still on paper. For now, the only thing available is a signup for early access. It’s the first commercial product to come out of TBC’s development process.

The neurons stay in the lab

The company name suggests brain cells are doing the computing here, but they aren’t. What TBC sells is ordinary software: a proprietary layer that adds less than 0.1 percent to the size of the base model and runs entirely on standard AI infrastructure. Customers need no biological hardware and no new workflows.

The biology sits in the development process instead. TBC grows cortical nerve cells on chips with 4,096 electrodes, stimulates them electrically, and measures how the activity spreads across space and fades over time. From that, the company derives metrics like how far a signal reaches and how long it lingers. Those principles feed into small software modules, called adapters, that plug into existing diffusion models. The neurons themselves never leave the lab.

TBC, originally called Biological Black Box, was founded by neurosurgeons Alex Ksendzovsky and Jon Pomeraniec, and raised $25 million in February 2026. The company picked generative video as its entry point because that’s where the hardest parts of AI inference show up: quality, speed, and cost. After that, TBC wants to tackle world models and eventually let neurons compute alongside silicon.

How TBC turns a cell culture into a software module

TBC laid out the path from cell chip to AI model in a blog post, using the open Minecraft world model Oasis as an example. Oasis is a diffusion model with roughly 600 million parameters that builds game scenes frame by frame from a player’s inputs. Its typical problem is that each new frame builds on the last, so small errors pile up. After a few seconds, the scene blurs into a shapeless mush that TBC calls “washout.”

The biological part starts in the lab. TBC grows cortical nerve cells on a chip with 4,096 tiny electrodes, and the cells wire themselves into an active network on their own. Through the electrodes, the team can stimulate individual spots electrically while recording how the excitation spreads across the culture and fades again.

For the Oasis experiment, TBC essentially “showed” the nerve cells images. The image data was translated into patterns of electrical pulses fed to the culture through specific electrodes. The neurons responded, and not just at the stimulated spot: the excitation spread to neighboring cells and faded after a short time. That’s exactly the behavior TBC recorded. Where does the network become active, how far does the activity reach, and how long does it last?

The company translated these measurements into numbers and built a small software module from them, the “Neural Dynamics Adapter.” Put simply, it holds a pattern that mimics how a signal spreads across the cell culture within limited space. Because the video model can’t do anything with that raw data on its own, a second component comes in: a kind of translation layer that converts the biological pattern into the language of the AI model’s intermediate results. Only then can the model use the structure drawn from the cells when it computes images.

The adapter builds one basic rule into the model, a rule TBC read off the cells: activity acts mainly in the neighborhood and then fades. It works like a damper that holds neighboring image regions together and keeps disturbances from spreading unchecked.

At about 156,000 parameters, it’s small, and the base model’s architecture stays unchanged. It was then trained on Minecraft video sequences, while the base model got a light nudge at the same time with a low learning rate. No cells are involved during operation, since the adapter is by then a fixed mathematical transformation that runs on any GPU.

According to TBC, Oasis with the adapter held sharp object edges and stable spatial relationships far longer than without it. As its measure, the company used the differential entropy of the images, roughly a gauge of how much visual information a frame still holds before it blurs into a flat surface. Across ten test videos, the adapter variant scored about 19 percent better than the original, about 15 percent better than a same-sized conventional fine-tuning, and about 5 percent better than a LoRA adaptation.

In a follow-up project, TBC built a larger adapter with roughly three percent extra parameters and pushed the frame rate from about two to just under ten frames per second. The speed gain didn’t come from the adapter itself, but from cutting the number of compute steps per frame without the quality collapsing the way it usually does. An outside test by infrastructure provider Bluesky Compute in July found about 4.4 times lower inference costs and more than three times as much coherent video as the base model.

Plenty of numbers, little proof

Comparable details are missing for the commercial model. TBC names neither the base model nor the hardware, resolution, clip length, or quality metric behind its claims. It’s also striking that five times faster and 80 percent cheaper are mathematically the same thing: if you need a fifth of the GPU time, you save 80 percent. So this is probably one effect stated in two units, not two breakthroughs.

Whether the biological detour beats purely digital tricks like distillation or caching also remains an open question, and TBC hasn’t published a controlled comparison on the same base model. The company’s real claim isn’t that neurons compute better than chips anyway, but that they point to algorithms nobody would have found at a desk. Whether that holds up can’t be judged from the data so far.