✦ Weights & Biases & the Software LayerThin moat

The subtlest bet is an attempt to stop being a commodity. CoreWeave acquired Weights & Biases — the experiment-tracking and model-development platform used across the AI research community — in 2025, and has been wiring it into the infrastructure since: Mission Control cluster health data surfaced directly inside W&B so engineers can correlate a failed training run with the node that caused it, plus inference and evaluation products layered on top¹.

The strategic point is switching costs. Renting GPUs is a price comparison; a team whose experiments, model registry and production evaluations live inside a platform faces a migration, not a purchase order. It is the same logic that makes the hyperscalers sticky — the compute is replaceable, the tooling around it is not — and it is the only bet on this page that could create a moat rather than a scale advantage.

The honest caveat is that the tooling is deliberately portable: Weights & Biases works with any cloud, which is why researchers adopted it in the first place, and making it CoreWeave-exclusive would destroy the thing that was bought. So the payoff has to come from integration being genuinely better rather than from lock-in. Watch for platform revenue disclosed separately from compute, and watch whether customers cite the software when they renew. Until then this is a good product attached to a rental business.