CoreWeave Building Full AI Stack

These deals show CoreWeave trying to turn GPU supply into a full AI build environment, which makes it harder for customers to swap out the cloud without also redoing their training, evaluation, and deployment workflow. Weights & Biases brings experiment tracking and model monitoring. OpenPipe adds reinforcement learning fine tuning for agents. Marimo adds the notebook where developers write and run code. Monolith extends that stack into industrial engineering use cases.1 2 3 4 5 6

• Weights & Biases was the anchor acquisition. CoreWeave completed the deal on May 5, 2025, then used it a month later to launch new software products on top of its cloud. That moved CoreWeave from renting GPUs to owning the system of record for model experiments, evals, and production monitoring.1 2
• OpenPipe and Marimo fill the workflow on both sides. OpenPipe trains agents with reinforcement learning and was folded into a joint serverless RL product with Weights & Biases and CoreWeave. Marimo is the interactive Python workspace, so the stack now spans notebook, training, tracking, and serving.3 4 5
• This mirrors a broader market pattern where infrastructure companies climb into software to capture more budget and reduce churn. Databricks used MosaicML and MLflow to cover training through deployment, while GPU cloud rivals like Lambda still position more around developer friendly infrastructure rather than owning the full MLOps layer.6 7 8

The next step is for CoreWeave to bundle these tools into a default operating system for AI teams, especially labs and enterprises that want one vendor for notebooks, post training, observability, and compute. If that bundling works, CoreWeave competes less like a cloud capacity broker and more like an AI platform with deeper product lock in.2 3 4 5 6