CoreWeave's AI improvement loop: Forge, RL rollouts,and the inference tipping point
Key points
Key takeaways from a theCUBE interview with CoreWeave VP of Product for AI Services Urvashi Chowdhary (October 2026):
From training to inference. A CoreWeave survey of 15 customers (7 heavy users, 6 recent users and 2 that chose not to go with CoreWeave) found inference share rising from around 10% of workloads in year one to a roughly 60/40 training-to-inference split by year two, with customers expecting about 90% inference going forward. Training still grows; inference is just growing faster.
Forge is free to get started. Forge offers API-key access to leading open-source models with no contract required, plus tooling to distill and fine-tune models,track experiments,and observe agents. Arya, an automated research assistant, helps developers move through the improvement loop faster, completing CoreWeave's "AI improvement loop".
Reinforcement-learning rollouts. Working with Nvidia's Dynamo stack,a CoreWeave customer improved model quality by 23%and cut tool calls by 30% during RL rollouts,and new checkpoints now ship into scaled inference about 15x faster. RL rollouts are becoming a major workload for agentic use cases.
An AI-factory stack,not just compute. CoreWeave positions itself as AI infrastructure:training, post-training,and inference built on high-throughput hardware with fast storage,and networking,with optimisations spanning inference engines,KV-cache-aware routing,prefill/decode disaggregation,quantisation,and custom speculative decoding for popular open-source models.
Observability for fleets of agents. As customers run more agents,and more tool calls,CoreWeave added agent lenses for tracing which tools are called,plus secure sandboxed execution,so teams can spot bottlenecks,and iterate.
Open source from the start. CoreWeave builds each service on top of the others,and deliberately contributes back to open ecosystems,contrasting the early AWS criticism over open-source contribution,and argues this keeps customers flexible.
Speed is the CoreWeave way. Chowdhary,a former AWS executive,says CoreWeave values customer obsession,ownership,and taking on hard problems,and that "moving with speed" lets it ship at an Nvidia-like pace.
Read more: theCUBE interview · theCUBE
