Workers AI

The idea is compelling: serverless GPU inference on Cloudflare’s edge, available to a Worker through one binding and with no infrastructure to manage. When it works, edge latency meets the promise and the catalog covers popular open-weight models.

Reliability sinks the score. Whole models can become unusable for extended periods, sometimes without a status-page incident. This has continued for too long to build around confidently. If a product depends on one model responding, it needs another provider as a fallback. That fallback soon becomes the primary.

Pricing does not help. The mix of “neurons” and per-token rates makes comparisons unnecessarily difficult, without beating dedicated inference providers on price or speed in most cases. I would reconsider it if reliability improved. Today, I see no strong reason to choose it.