Caterpillar, CoreWeave cut physical AI data loop to hours

The push into physical AI lands in a construction market where demand for new data centers, power plants and highways is fueling a boom even as the industry deals with declining productivity and a shortage of skilled machine operators. Caterpillar's answer draws on its years of autonomous equipment experience in mining, but as Brandon Hootman describes it, mining and construction are polar opposites: a mine site, once instantiated, does not change frequently, while construction requires matching a structured system to an unstructured environment.

What makes the pairing with CoreWeave notable is the data problem underneath. Caterpillar's digital ecosystem already holds about 18 petabytes of federated data from machines, dealers and customers, yet Hootman calls that just a drop in the bucket compared with what is needed. A single machine can produce terabytes of Lidar, camera, multi-second control and performance data in a given day. CoreWeave, which named Caterpillar among its enterprise customers in its second-quarter results announcement, brings graphics processing unit capacity and applied expertise, plus a recently launched Physical AI Field Engineering service that embeds its engineers with customers' domain experts. Ahlfeld frames physical AI as an entirely different beast that requires a lot of storage and different infrastructure.

The payoff the two describe is speed: annotation and labeling that once took months or weeks now happens in hours, with the loop into simulation or training closing inside the workday. Whether that holds up likely hinges on how well the approach copes with the far more variable construction setting, and on whether the data pipeline keeps pace as more machines come online. Equipment makers, cloud providers and the operators supplying that field data are the parties whose work this reshapes.