TBC says its neuron-derived video model runs 5x faster

The Biological Computing Co. says it has a text-to-video model that runs five times faster and costs 80% less to serve than the open-source model it is built on, with the gain coming from a software layer it derived from measurements of lab-grown nerve cells. The company announced the work with AWS on September 22, 2026.

The layer is the whole pitch. According to AWS's announcement, it adds less than 0.1% to the model, needs no biological hardware, and is not a new customer workflow. The plan is to run it on Amazon's Trainium chips, serve it through SageMaker AI and sell it in AWS Marketplace.

Sizes and dates are the parts worth reading twice. The two headline figures are not independent: The Decoder pointed out that five times faster and 80% cheaper describe the same effect, because inference cost scales with compute time. The company gave no standard quality metric to support its claim of improved quality, and it has not named the open-source base model.

Some adaptation numbers exist, and they come from TBC's own write-up rather than from the announcement. The Decoder records a 19% improvement in differential entropy against the unmodified base model on the Oasis world-model benchmark, a 15% improvement over conventional fine-tuning at the same parameter budget and a 5% improvement over LoRA adaptation. In that experiment the adapter carried about 156,000 parameters on top of Oasis's roughly 600 million and left the base model's architecture unchanged, and TBC says the adapted model held sharp object edges and stable spatial relationships far longer than the base model did.

What a creator can do with it

Nothing yet. The only public surface is an early-access form, which offers access "at no cost to you". There is no price for the commercial product, no partner list, and no model ID to call. AWS is not just hosting a demonstration: putting the component in Marketplace alongside Trainium and SageMaker is how an enterprise buys compute it can invoice.

Two things follow for anyone budgeting AI video this quarter. First, per-second API prices carry the inference cost of the model behind them, so a working 80% cut is the kind of thing that shows up later as a lower rate card, not as a tool you can use. Second, "five times faster" is the marketing form of a compute saving. Ask for the quality number on the same prompts before you treat a speed figure as a capability figure.

Watch for a named base model and a price when the early access list opens; until then the claims sit in a press release and a signup page.