Weights
& Biases

Somewhere in the Mission District of San Francisco, behind a karate studio, three people decided in 2017 that machine learning teams were working in the dark. Experiments ran, models trained, numbers flickered across terminal windows - and then most of it disappeared. No log. No audit trail. No way to recall, two weeks later, exactly which hyperparameters produced the run that actually worked.

The people who noticed this most acutely were the ones building AI at OpenAI. Lukas Biewald - Stanford CS, former Yahoo engineer, founder of data labeling company CrowdFlower - had sold his first machine learning tools company for $300 million, locked himself in an Airbnb to implement backpropagation by hand, built robots on Raspberry Pis in his garage, and then talked his way into an internship at one of the most competitive AI labs in existence.

What he found there was a universal frustration. Tracking experiments was manual, slow, and unreliable. The infrastructure for doing rigorous AI research well simply didn't exist as a product. So he built it - with Chris Van Pelt (his co-founder from CrowdFlower) and Shawn Lewis (a veteran Google engineer) - and called it Weights & Biases.

The name is a reference to the two core parameters in any neural network. Simple, literal, and - if you've trained a model - immediately obvious why it matters. These are the numbers you're always chasing.

In March 2025, CoreWeave - the GPU cloud company that went public the same month - announced it was acquiring Weights & Biases for $1.7 billion. The deal closed in May 2025.

The strategic logic is clear: CoreWeave sells raw GPU compute. W&B is the software layer that makes GPU compute actually useful for AI developers. Together, they offer something neither could alone - infrastructure with intelligence baked in, from training run to model registry to deployment.

The first concrete product of the partnership was Mission Control, announced at the Fully Connected Conference in June 2025. It's an integration layer between W&B's tooling and CoreWeave's infrastructure - the beginning of a vertically integrated AI development stack that neither pure cloud providers nor pure software companies can currently match.

W&B Inference, launched in the same announcement, gives developers OpenAI-compatible API access to open-source models, with all the tracking and evaluation tools W&B is known for. For teams tired of black-box APIs, it's a genuinely different proposition.