CoreWeave built its reputation selling raw AI compute. Now it wants to sell the software layer on top of it. At its Fully Connected conference, the company unveiled CoreWeave Forge, a development platform that folds model serving, observability, post-training, and evaluation into a single environment — and it comes with a striking performance claim: a preview feature called RL Rollouts that improved model reload latency by 15x in testing.
Forge is aimed at a problem anyone who trains large models knows well. In reinforcement learning, a model generates attempts that get scored and used to update it, and each update creates a new checkpoint that must be reloaded into the inference servers producing the next round of attempts. Slow reloads hold up training. RL Rollouts, built on Nvidia's Dynamo framework, hot-loads new checkpoints into a live inference deployment so the training loop never has to stop and wait.
The 15x figure, attributed to CoreWeave's Chetan Kapoor in conversation with SiliconANGLE, measures reload latency against a baseline configuration — not total training speed — but for teams running RL at scale, checkpoint reloading is one of the real bottlenecks. Forge wraps that capability with tools teams otherwise stitch together themselves: managed Python notebooks, Weights and Biases experiment tracking, OpenPipe post-training work, a new Agent Lens observability tool, and serverless supervised fine-tuning and reinforcement learning, which CoreWeave says trains 1.4x faster at 40% lower cost than a self-managed setup.
The platform comes in Free, Pro, and Enterprise editions — Pro starts at $60 a month with a 30-day trial — a deliberate bid to reach individual developers, not just the giant customers that made CoreWeave's name. MasterClass and Canva are already building on it, and more than 4,500 attendees showed up for the launch event. The company says Forge is open across the models, frameworks, and clouds teams use, with existing Weights and Biases credentials working there today.
The business logic is clear. CoreWeave reported $35.6 billion in total debt at midyear and spent $14.1 billion on property and equipment in the first half of 2026. Infrastructure that expensive needs high-margin software on top. Forge is CoreWeave's answer to the industry's margin question — and a bet that the next bottleneck in AI is not getting GPUs, but making them useful.