Crowkis vs Cloudflare AI Gateway: the edge is the wrong place for trust decisions
Cloudflare's gateway adds caching at the CDN layer — exact-match, eventually-evicted, on someone else's network. Useful plumbing; not a reuse brain.
Cloudflare AI Gateway does what Cloudflare does well: put a proxy near your users with analytics, rate limiting, and a cache toggle. The cache is what edge caches are — keyed on exact request bytes, evicted on the edge's schedule, holding your customers' questions and answers in CDN infrastructure you don't control. For static assets that's the whole point. For LLM responses it's three problems wearing a feature's coat.
Exact-match keying means the paraphrase traffic that dominates LLM workloads sails through unmatched. Edge residency means your most sensitive corpus — everything your users ask — is cached outside your network boundary, with whatever multi-tenancy story the edge has that week. And there's no concept of reuse safety: no intent gates, no confidence floor, no poisoning defense. Whatever was stored gets served.
Five stages score every write before it can ever be served.
Crowkis runs where your data already lives, behind your own firewall, with the entire trust machinery between storage and serving: five read gates, five write stages, tenant walls, audit trails. Latency is sub-millisecond from local disk — the edge's one advantage evaporates when the cache is in the same rack as the app.
The bottom line
Use Cloudflare for the things edges are for. The decision about whether a cached LLM answer is safe to show a user belongs inside your perimeter, made by an engine that was built to make it.