Wayve Signs Multi-Year Deal to Train AI Driver on CoreWeave Cloud

CoreWeave announced on November 9, 2026, a multi-year agreement under which Wayve will train its AI Driver foundation models on CoreWeave Cloud using NVIDIA Blackwell GPUs. The partnership covers the full development loop — training, evaluation and simulation — for a system that learns from real-world driving data collected in more than 70 countries.

The deal matters because autonomous-driving development has become a compute problem as much as a robotics problem. Wayve's approach, often described as end-to-end learning (models trained directly on driving experience rather than hand-coded rules), demands sustained training runs, verifiable progress between software versions and enough observability to compare models and validate safety at scale. CoreWeave now supplies a portion of that infrastructure.

How did Wayve's infrastructure evolve before this deal?

Wayve began its machine-learning tooling with Weights & Biases, the experiment-tracking platform, and has since scaled into CoreWeave Cloud. The expansion adds CoreWeave's cloud compute to a development process that, according to the announcement, must show measurable progress between software versions, sustain long training runs and provide the observability needed to compare models and validate safety at scale.

Wayve is developing the AI Driver for a range of deployments, from personally owned cars to robotaxis. Silvius Rus, Wayve's senior vice president of engineering, described the company's development loop in a statement: "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."

That quote carries the core engineering discipline of the program. Deployment is not automatic. A new model ships only when it beats the version already operating on public roads, a gating mechanism that turns evaluation infrastructure into a safety control. Rus expects access to CoreWeave's AI cloud to accelerate commercial deployment of the AI Driver.

Why does simulation come before road deployment?

A CoreWeave technical overview of physical-AI infrastructure identifies three autonomous-vehicle workloads: world-model training, sensor simulation and closed-loop evaluation. The numbers behind those workloads are large. Autonomous vehicles generate terabytes of sensor data per vehicle per day, and handling that flow requires storage, networking and training pipelines to operate efficiently together.

Physical AI development (AI systems that act in the physical world) does not stop when a vehicle ships. New edge cases appear after deployment, which makes repeated retraining, checkpointing (saving intermediate model states during training) and fast data pipelines operational requirements rather than optional extras. Because failures can have physical safety consequences, most model development runs through simulation and staged validation before real-world use.

CoreWeave published a reference demonstration that quantifies the scale of simulated testing now possible. 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. CARLA is an open-source driving simulator. The runs varied weather, traffic density, maps, intersections and random seeds, illustrating the range of simulated conditions that can be generated for model development. For a company like Wayve, which must prove a new model outperforms the current one before deployment, that kind of parallel simulation throughput directly compresses the iteration cycle.

What keeps long training runs reliable?

Long training jobs fail for prosaic reasons, and hardware reliability becomes a first-order concern. 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. The engine is designed to identify potential faults early, localize problems and aid maintenance — capabilities CoreWeave identifies as directly relevant to the reliability requirements of long training runs.

Across its broader physical-AI platform, CoreWeave says its Kubernetes Service orchestrates training and simulation jobs, while distributed and object storage holds petabyte-scale sensor and simulation data. Networking is engineered around the GPU, not the CPU. InfiniBand and RoCE (RDMA over Converged Ethernet) links move data directly between GPUs across servers, bypassing CPUs to reduce idle waiting while distributed jobs exchange sensor data, simulation state and gradients (the gradient is the signal used to update model weights during training).

Where does traceability enter the picture?

Training capacity alone does not validate a safety-critical driving system. The experiment tooling adds version lineage: Weights & Biases Models tracks experiments, Registry connects a checkpoint to the exact dataset and training run that produced it, and Weights & Biases 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 driving system, that traceability is not bookkeeping. It is the mechanism that supports comparison and validation alongside raw training capacity. When an evaluation flags a failure mode, engineers can reconstruct precisely which data and which training run produced the behavior — a prerequisite for the staged validation process that autonomous-vehicle development demands.

What does the agreement change for Wayve's roadmap?

The agreement makes the entire iteration cycle — gathering driving experience, training on it, evaluating it and deploying improvements — part of Wayve's infrastructure expansion. The announcement frames the stakes plainly: 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.

The implication for the market is clear. Autonomous-driving companies are increasingly competing on infrastructure: who can run more simulations per hour, retrain faster on fresh edge cases and prove, with auditable records, that each release is safer than the last. Wayve's multi-year commitment to CoreWeave's Blackwell-based cloud signals that the company intends to compete on that cycle time as it moves the AI Driver from development toward personally owned cars and robotaxis.