Startup Claims Neuron-Derived Chip Speeds Up AI Video

The Biological Computing Co. says a tiny software add-on built from lab-grown neuron activity could make text-to-video generation five times faster and far cheaper, though the claims remain unverified and the product isn't public yet.

A startup called The Biological Computing Co. is pitching a novel shortcut for AI video generation: a lightweight adapter trained on patterns recorded from real, lab-grown neurons. The company says this so-called Neural Dynamics Adapter can be bolted onto an existing text-to-video model to make it run about five times faster while cutting compute costs by roughly 80 percent. What makes the pitch unusual is the adapter's size. According to the company, the layer contains around 156,000 parameters and adds less than 0.1 percent to the footprint of whatever base model it's attached to — a rounding error compared to the billions of parameters typical large video models carry. TBC has not named which underlying model the adapter works with, making it hard to independently gauge how the efficiency gains would translate to real-world use. To bring the technology to market, TBC says it's pursuing a partnership with Amazon Web Services, positioning the adapter as a way for cloud customers to generate video content without the heavy infrastructure costs usually associated with the task. As The Decoder reports, these performance figures come entirely from the company's own announcements rather than from outside testing or peer review. The product itself isn't available to try yet — TBC is currently only collecting sign-ups for early access, meaning the promised speed and cost improvements can't yet be checked against actual output. Until independent benchmarks or hands-on demos surface, the claims sit in the realm of a bold pitch rather than a proven breakthrough in how video models are trained or run.