The Pick-and-Shovels Thesis for Physical AI

The venture-capital cliché is that in a gold rush you should sell picks and shovels. It is a cliché because it is usually right. The people building the models are the ones taking the biggest technical and business risk; the people supplying them are taking a smaller, more diversified one.

In physical AI, the model builders are the frontier labs, OpenAI, Genesis AI, RLWRLD, and the humanoid pure-plays. The picks and shovels are capture hardware, operator networks, and the data-pipeline infrastructure that sits between them.

Not every AI cycle has a clean pick-and-shovel layer. Language-model training data was largely scraped from the public web, there was no meaningful supplier tier between the scrape and the model. Physical AI is different. Manipulation data does not exist on the web in the shape a model needs. It has to be captured, deliberately, with instrumented hardware, by paid operators. That entire capture stack is a supplier tier that has to exist and be paid for.

This is why data-infrastructure companies in robotics have been able to raise real capital on real revenue in a way that data-infrastructure companies in language never could. The customers are the frontier labs, and they are writing checks because there is no free alternative.

Two things. The one whose cost per usable operator-hour is lowest at scale wins the volume tier of the market. The one whose data quality and delivery integration is best wins the frontier-lab tier. The two are not necessarily the same company, and the market is likely large enough for several to coexist for years.

The bet BLO LAB is making is that wearable capture, an operator network built on Talika, and a delivery pipeline designed for foundation-model training will let us play in both tiers over time. It is not the only strategy in the market. It is the one that fits the shape of demand we see.