Caterpillar and CoreWeave cut physical-AI learning loop to hours

The conversation at the Fully Connected event framed construction as a harder problem than mining, where Caterpillar has years of experience with autonomous equipment. Brandon Hootman argued that a mine site changes infrequently, while construction is its polar opposite, requiring systems that match a structured setup to an unstructured environment. Richard Ahlfeld added that physical AI is an entirely different beast from the AI clouds built for foundation-model training and agentic inference, requiring a lot of storage and a different infrastructure.

Caterpillar began working with CoreWeave this year, drawn by both graphics processing unit capacity and applied expertise, and CoreWeave later named Caterpillar among its enterprise customers in its second-quarter results announcement. Working with Nvidia, the partners use AI models to annotate and label incoming field data. Training an autonomous excavator means ingesting telemetry and vision data, simulating a digging scenario a million times and adding reinforcement learning. Hootman noted that a single machine can produce terabytes of data within a given day, encompassing Light Detection and Ranging data, camera data, multi-second control data and performance data.

Whether the hours-long loop holds may depend on how those terabytes accumulate, since Hootman described the current 18 petabytes as only a drop in the bucket. For construction and mining firms facing declining productivity and fewer skilled operators, the value of this approach hinges on whether that accelerated loop survives that scale, and on whether CoreWeave's embedded-engineer model proves repeatable beyond this partnership.