“If you want me to predict when your car fails and you give me 100,000 hours of flawless driving data, I am blind.” That is how Richard Ahlfeld, SVP of Physical AI at CoreWeave, put it in a written interview published by Superintelligence on October 4, 2026.
First things first: what physical AI means
Physical AI acts in the real world: self-driving cars, factory robots, humanoid robots or equipment failure prediction. It works with sensor, camera, test-bench and telemetry (measurements an operating machine sends back live) data. A wrong call can mean a crash, downtime or damaged equipment.
Data from normal operation alone cannot teach a model to predict failures
AI cloud company CoreWeave launched its Physical AI Field Engineering service on September 10, 2026. It embeds engineers with automotive, aerospace and mechanical backgrounds in client teams to build models from clients’ test-bench results, simulation outputs, production-line sensors and live telemetry. The team and approach come from Monolith, an engineering AI company CoreWeave acquired in 2025.
Richard Ahlfeld, Monolith's founder and now SVP of Physical AI at CoreWeave, put it bluntly: the most common failure is "a training set that never contained the moments that matter." No amount of flawless driving hours can fill in a failure that never appeared.
NEURA Gym: a data loop that closes every day
He points to customer NEURA Robotics as an example. NEURA Gym runs roughly 100 cells where a robot attempts a customer's real task. The loop closes daily: evaluate what happened, decide what data to collect, collect it, train, validate in simulation, and test on the physical robot again.
The lesson: finding the right data matters more than volume. Field engineers use domain knowledge to decide which edge cases (rare situations that cause real trouble when they occur) deserve synthetic coverage and which anomalies are meaningful signals. Every day the loop checks the same thing: whether the data in hand is missing a critical moment.

75 iterations for decisions made in under 30 seconds
Another example is the radio system for the Aston Martin Aramco Formula One Team: starting from seven hours of annotated audio, it was refined over 75 iterations to cope with engine noise, 300 km/h wind, multilingual drivers and team shorthand.
Each version was checked in Weights & Biases Weave on word error rate (the share of recognized words that are wrong) and LLM-judged quality, and the system ran at the Qatar and Abu Dhabi races before full rollout. Problems that used to take engineers several minutes are now handled within a pit-stop window of under 30 seconds. The interview gave no error rates, so “75” and “30 seconds” describe process and outcome and say nothing about accuracy.
Validate the simulation before trusting synthetic data
Critical real-world moments are rare and not easy to record, so the industry often uses simulators or “world models” that generate realistic scenes to produce synthetic data. Ahlfeld’s first rule: validate the simulation extensively against controlled real-world measurements. He sorts synthetic data into two categories:


His example is a humanoid robot in training: its raw data had only a few low-light clips and no direct sunlight, and the task failed. The team used the NVIDIA Cosmos world model to add those lighting conditions; after retraining, the robot handled both. He stresses that it worked because lighting is a world-model strength; problems of contact, material behaviour or wear still need real testing.
What it means for Taiwan

Taiwan’s energy transition is in a phase of many demonstrations and test runs, with CCS, hydrogen and new generation equipment building up their first field data. The anomalies seen during test runs are the stretch that the 100,000 hours of normal data would be missing. Three practices to consider:
Keep anomaly data: record anomalies, shutdowns and edge-of-envelope conditions from test runs systematically, not just steady-state data.
Validate the simulation before using it to fill data: use real measurements to establish where the simulation can be trusted. Areas such as subsurface fluids and material corrosion, where simulation is less accurate, still need field testing.
Keep humans in the loop: Ahlfeld says every AI recommendation for a physical system should be reviewed by a person and leave a traceable record, which applies to power plants and demonstration sites too.

Sources (5)
Sources
Superintelligence, "The most common failure isn't a bad model": CoreWeave's Richard Ahlfeld on physical AI (published October 4, 2026): https://read.getsuperintel.com/p/the-most-common-failure-isn-t-a-bad-model-coreweave-s-richard-ahlfeld-on-physical-ai
CoreWeave, CoreWeave Launches Physical AI Field Engineering (September 10, 2026): https://www.coreweave.com/news/coreweave-launches-physical-ai-field-engineering-to-turn-proprietary-data-into-production-ai
CoreWeave, AI Agents Decode F1 Radio in Near Real Time: https://www.coreweave.com/blog/how-ai-agents-on-coreweave-help-process-f1-radio-in-near-real-time
CoreWeave, CoreWeave's Stack for Physical AI: https://www.coreweave.com/blog/wayve-decart-neura-robotics-nissan-all-running-physical-ai-on-coreweave
NEURA Robotics, NEURA Gym: https://neura-robotics.com/neuragym/

Leave a comment
Comments are reviewed before publication. Email is private and used only when contact is necessary. Privacy notice