Amazon Web Services and The Biological Computing Co are selling a text-to-video model that claims to run five times faster and cost 80 percent less than its open-source predecessor. The performance boost comes from a proprietary software layer derived from measurements of real nerve cells.
The San Francisco startup, TBC, announced the partnership in a press release. The new tool builds on an unspecified open-source video model but adds a layer that increases the model size by less than 0.1 percent. The company states the adaptation offers better quality while reducing inference costs. TBC plans to host the model on AWS Trainium chips via Amazon SageMaker AI, but the product is currently available only as a signup for early access.
The neurons stay in the lab
Despite the name, biological hardware is not involved in the final product. Customers do not need special equipment or new workflows. The biology exists only during the development phase.
Researchers grow cortical nerve cells on chips with 4,096 electrodes. They stimulate the cells electrically and measure how activity spreads across the culture and fades over time. These metrics feed into small software modules called adapters that plug into existing diffusion models. The cells themselves remain in the laboratory.
Neurosurgeons Alex Ksendzovsky and Jon Pomeraniec founded the company, originally named Biological Black Box. They raised $25 million in February 2026. The team chose generative video as its entry point because the task exposes the hardest problems with AI inference regarding quality, speed, and cost. Future plans include world models and letting neurons compute alongside silicon.
How TBC turns a cell culture into a software module
The company explained its process using the Minecraft world model Oasis as an example. Oasis is a diffusion model with roughly 600 million parameters that builds game scenes frame by frame. Small errors pile up as each new frame depends on the last, causing the scene to blur into a shapeless mush. TBC calls this effect “washout.”
In the lab, TBC grows cortical nerve cells on a chip with 4,096 electrodes. The cells wire themselves into an active network. The team stimulates individual spots electrically and records how excitation spreads and fades.
For the experiment, the team translated image data into patterns of electrical pulses fed to the culture. The neurons responded by spreading excitation to neighbouring cells before the activity faded. The researchers recorded where the network became active, how far the activity reached, and how long it lasted.
These measurements were translated into numbers to build the “Neural Dynamics Adapter.” This small software module holds a pattern that mimics how a signal spreads within limited space. A second component translates 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.
The adapter builds one basic rule into the model: activity acts mainly in the neighbourhood and then fades. It works like a damper that holds neighbouring image regions together and prevents disturbances from spreading unchecked.
The module contains about 156,000 parameters. The base model’s architecture remains unchanged. The adapter was trained on Minecraft video sequences while the base model received a light nudge with a low learning rate. No cells are involved during operation because the adapter is now a fixed mathematical transformation running on any GPU.
TBC claims Oasis with the adapter held sharp object edges and stable spatial relationships far longer than the original. The company measured this using the differential entropy of the images, a gauge of how much visual information a frame holds before it blurs. Across ten test videos, the adapter variant scored about 19 percent better than the original. It performed 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. This pushed the frame rate from about two to just under ten frames per second. The speed gain did not come from the adapter itself but from cutting the number of compute steps per frame without quality collapsing. 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 is 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. This is probably one effect stated in two units, not two breakthroughs.
Whether the biological detour beats purely digital tricks like distillation or caching remains an open question. TBC has not published a controlled comparison on the same base model. The company’s real claim is not that neurons compute better than chips, but that they point to algorithms nobody would have found at a desk. Whether that holds up cannot be judged from the data so far.
What it means
For people making video with AI, the change is practical rather than theoretical. The new tool promises to keep images sharp for longer without needing more powerful hardware. The claim is that you can generate more coherent footage in less time. However, the lack of specific benchmarks for the commercial version means users must trust the initial test results before committing to the service.
