CoreWeave Says Caterpillar Is Deploying Vera Rubin for Physical AI Training

CoreWeave says Caterpillar is deploying Nvidia Vera Rubin infrastructure on its cloud platform as a data factory for training specialized physical-AI models at industrial scale. The claim appears in a CoreWeave investor presentation filed with the U.S. Securities and Exchange Commission.

The infrastructure is intended to support Caterpillar’s model training. CoreWeave places the deployment among physical-AI, robotics and simulation workloads used to train and simulate world models, foundation models and generative-design systems. The presentation cites autonomous construction equipment and vehicles, humanoid robots and engineering simulation as examples for the broader workload category, but does not say Caterpillar will pursue each application.

The presentation was furnished as Exhibit 99.2 to a CoreWeave Form 8-K filing dated Sept. 17, 2026. It describes Caterpillar as a Vera Rubin customer use case but does not state how many systems the company plans to use, where they will run or what the commercial arrangement is worth.

Nvidia describes a Vera Rubin NVL72 rack as integrating 72 Rubin GPUs and 36 Vera CPUs connected by NVLink 6, along with ConnectX-9 SuperNICs and BlueField-4 data-processing units. Nvidia says the platform supports AI pretraining, post-training, test-time scaling and inference. For additional platform context, DataPhoenix has covered CoreWeave’s Vera Rubin NVL72 cloud access. CoreWeave did not specify Caterpillar’s rack count or say which workload stages it will use beyond training.

The presentation and other opened official sources do not provide a Caterpillar-specific installation or training-start date. They also do not identify the models, training data, evaluation criteria, target machines or measured production outcomes for the deployment.