CoreWeave launches CoreWeave Forge for AI inference stack

CoreWeave's move reflects a shift in where AI workloads are concentrated. Training built the first wave of GPU clouds, but according to Urvashi Chowdhary, vice president of product and AI services at CoreWeave, serving models faster and cheaper will define the next phase. That is pushing specialized cloud providers beyond raw GPU capacity into storage, networking and software.

The company's answer is to tune every layer above the hardware, from the vLLM engine to quantized models and custom speculative decoders. Chowdhary said CoreWeave has been intentional about leveraging open-source tools and contributing back to open systems so customers keep flexibility, while building its own services on top of each other. CoreWeave RL Rollouts addresses a specific pressure point: when customers train agentic models with rewards and verifiers, inference becomes the bottleneck during rollouts, and the new capability lets those checkpoints roll into a live inference setup so it can scale independently.

Those pieces now sit inside CoreWeave Forge, launched at the Fully Connected event. Chowdhary framed the goal as creating accessibility for leading technologies, so that even an individual developer signing up alone gets the same performance and reliability.