Field notes / GPU Cloud

Notes from the control plane.

Short arguments about running accelerators responsibly: capacity, spend, model policy, and the small decisions that keep systems dependable.

The hidden cost of an unbounded GPU job

A practical argument for putting a maximum charge next to every launch button.

Capacity is a product surface

Why a list of GPUs is not enough when the decision is production-shaped.

Licenses belong next to the model

Serving policy is infrastructure policy when an endpoint can become a product overnight.