US and Chinese capital is piling into humanoid and foundation-model robotics. Several companies are currently hard at work building general AI models that you can plug into any robot so it performs whatever task is put in front of it. The promise is a ChatGPT moment for robotics: physical manipulation solved the way language was solved, by throwing enough data at a Transformer that it learns the structure of our physical world. The assumption is that whatever makes robotics different from language is a set of bottlenecks that you can trick your way past, and is not a reason to change what you build.

But this transition may not be as swift as they imagine. The buyers of these robotic models are concrete companies, each with concrete problems. They buy a solved operation, they don't buy "pluggable brains."

The product that is currently being funded assumes a general model that can adapt to any running factory: one master model, any task in front of it. However, in a running factory, whatever robot you add has to be compatible with machines already in place. Not only that, but the factory has to remain fully operational after the addition. While in an ideal world everything could be rebuilt from scratch, in reality capex limits how fast you can replace the old system. The sequence of what you change is what guarantees factory uptime. The machinery, the replacement sequence, and the business constraints in factories will always be plant-specific, and so the model fit has to be too.

General-purpose robotics is trying to cover the whole of the manufacturing sector at once. This sector contains companies of all sizes, with completely different machinery and business requirements.

Venture capital is funding adaptable models. Factory owners, however, are paying for immediate integration with their existing lines. Those are not the same product.

Some plants will buy the general-purpose pitch anyway, and pay the capex and the service fees. Most companies cannot make that change in one step. The ones that wait for a fully generalized model will lose the years in between. Upgrading the line you already have, station by station, beats waiting for one model that claims to do every job.