Cloudflare AI Platform adds unified inference layer

This is Cloudflare pushing from hosted models into genuine AI middleware. The pitch is less about raw model quality and more about operational control for agentic workloads, where cost, latency, and retries matter as much as output quality. A one-line switch between Cloudflare-hosted models and third-party models is the kind of abstraction teams actually want once they start juggling multiple providers, and automatic failover plus streamed-response buffering address real agent failure modes rather than demo-tier chat apps. Bringing Replicate’s catalog into the stack makes Cloudflare more of a distribution layer for models, and the BYO-model path via Cog suggests it wants to own both inference routing and custom-model hosting for production workloads.