Workers AI

The idea is incredible: serverless GPU inference on Cloudflare’s edge, one binding away from your Worker, no infrastructure to manage. When it works, latency from the edge is exactly as advertised, and the model catalog covers the popular open-weights bases.

The reliability is what sinks it. Entire models regularly go effectively offline for extended periods — sometimes acknowledged on the status page, often not — and the timeline for improvement has been too long to build against. If your product depends on a specific model responding, you need a fallback provider anyway, at which point the fallback tends to become the primary.

The pricing doesn’t help its case either: a confusing mix of “neurons” and per-token rates that makes cost comparison harder than it should be, without actually being cheaper or faster than the dedicated inference providers in most cases. I’d revisit this instantly if the reliability story changed; today there’s no compelling reason to choose it over the competition.