While the industry argues about cloud vs. on-prem, the most interesting workloads are moving to neither — they’re running at the edge, closer to the user than any region can reach.
Edge is not a buzzword anymore
Three things changed: V8 isolates made cold starts disappear, WASM made polyglot runtimes viable, and developers demanded sub-20ms response times globally. The result is a three-way split in the compute stack: edge (latency-critical, stateless), regional (transactional, stateful), and archival (cheap, cold).
The numbers that matter
- Cold start: V8 isolate ~5ms vs Lambda ~200ms
- Global p99: Cloudflare Workers ~30ms vs us-east-1 ~150ms transatlantic
- Cost per million requests: edge ~40% lower than regional, once cold starts dominate
What breaks at the edge
State. Transactions. Strong consistency. Anything that assumes a single source of truth breaks when your compute is running in 300 places simultaneously. The most successful edge architectures either keep state regional (Cloudflare D1, Turso, Neon) or embrace CRDTs and per-user durable objects.
The sleeper trend
GPU at the edge. Real-time inference for recommendation, moderation, and personalization increasingly runs at the edge — not because the cloud can’t, but because latency budgets don’t allow a round-trip. Watch the first edge-native vector database ship. That’s when this becomes obvious.