CoreWeave announced on November 9, 2026, a multi-year agreement for Wayve to train its AI Driver foundation models on CoreWeave Cloud with NVIDIA Blackwell GPUs. The work spans training, evaluation and simulation for a system that learns from real-world driving data collected in more than 70 countries.
Wayve began with Weights & Biases and has since scaled into CoreWeave Cloud. The expansion adds CoreWeave’s cloud compute to a development process that must show progress between software versions, sustain long training runs and provide enough observability to compare models and validate safety at scale.
A Data Flywheel for Driving Models
Wayve is developing the AI Driver for uses ranging from personally owned cars to robotaxis. Silvius Rus, the company’s senior vice president of engineering, described how its global driving data feeds model development: “Our AI Driver gets better by running this data flywheel: gathering real-world driving experience, training on it, evaluating it, and deploying only what outperforms what is already on the road.” He expects access to CoreWeave’s AI cloud to help accelerate commercial deployment of the AI Driver.
Simulation Before Road Deployment
A CoreWeave technical overview of physical-AI infrastructure identifies world-model training, sensor simulation and closed-loop evaluation as autonomous-vehicle workloads. Autonomous vehicles generate terabytes of sensor data per vehicle per day. Handling that flow requires storage, networking and training pipelines to operate efficiently together.
Physical AI development continues after deployment as new edge cases appear, making repeated retraining, checkpointing and fast data pipelines operational requirements. Since failures can have physical safety consequences, most model development runs through simulation and staged validation before real-world use.
In a CoreWeave reference demonstration outside the Wayve agreement, the company reports running 1.25 million CARLA autonomous-vehicle simulations in approximately 12 hours across independent shards using one GPU each. The runs varied weather, traffic density, maps, intersections and seeds, illustrating the range of simulated conditions that can be generated for model development.
Reliability Across the Training Stack
For the hardware layer, NVIDIA says Blackwell’s Reliability, Availability and Serviceability Engine continuously monitors thousands of hardware and software health data points to predict and intercept sources of downtime. It is designed to identify potential faults early, localize problems and aid maintenance, capabilities relevant to the reliability requirements for long training runs described in CoreWeave’s announcement.
Across its broader physical-AI platform, CoreWeave says its Kubernetes Service orchestrates training and simulation jobs, with distributed and object storage holding petabyte-scale sensor and simulation data. InfiniBand and RoCE networking move data directly between GPUs across servers, bypassing CPUs to reduce idle waiting while distributed jobs exchange sensor data, simulation state and gradients.
The experiment tools add version lineage: Weights & Biases Models tracks experiments, Registry connects a checkpoint to the exact dataset and training run that produced it, and Weave supports traces, evaluations and monitoring. Those records give engineers a route from a problem found in testing back to the model checkpoint, dataset and training run involved. For a safety-critical driving system, that traceability supports comparison and validation alongside training capacity.
The agreement makes the entire iteration cycle part of Wayve’s infrastructure expansion. Faster commercial deployment is tied to how efficiently integrated training, evaluation and simulation can turn driving experience into validated improvements while maintaining the reliability and traceability needed for safety-critical driving software.
Credit: Source link
