CoreWeave launches Physical AI Field Engineering service

In an interview with Superintelligence, Dr. Richard Ahlfeld, SVP of Physical AI at CoreWeave, explained that the most common failure in physical AI is not a bad model, but training data that lacks critical real-world moments. He cited NEURA Robotics, which uses a facility with roughly 100 cells to collect specific edge-case data, arguing that finding the right data matters more than volume. Ahlfeld noted that while synthetic data can fill gaps in areas like lighting, physical tests remain unavoidable for tasks involving granular, liquid, or soft materials.

The service uses CoreWeave’s internal stack, including Weights & Biases and the research agent ARIA. Customers named in launch materials include Nissan and the Aston Martin Aramco Formula One Team. For the F1 team, the system processes 40 radio channels, transcribing and categorizing messages within five seconds to support time-critical race decisions. Ahlfeld emphasized that customers retain ownership of the data and models, though performance depends on the underlying GPU hardware.