Edge Computing and Serverless: The 2025 Picture

The line between “serverless” and “edge” blurred in 2024. Here is what each platform actually offers and how to choose.

The two big models

Function-as-a-Service (FaaS) — AWS Lambda, Google Cloud Run, Azure Functions. Container-based, single-region, generous CPU and memory, slower cold starts.

Edge runtimes — Cloudflare Workers, Vercel Edge, Deno Deploy, Fastly Compute. V8 isolates or Wasm, multi-region by default, sub-millisecond cold starts, strict CPU limits.

Comparison

Concern	Lambda / Cloud Run	Edge runtime
Cold start	100–500 ms (Node), 1–3 s (Java)	< 5 ms
Max execution	15 min	30 s wall, 30–50 ms CPU on free tiers
Memory	up to 10 GB	128 MB typical
Filesystem	/tmp	none / virtual
Native modules	yes	rarely (Wasm only)
Database access	direct VPC peering	via HTTP / Hyperdrive / connection poolers
Multi-region	manual	automatic

Where edge actually wins

1. Auth and routing — Set-Cookie, JWT verification, A/B test buckets.
2. Static-site dynamism — geolocated content, redirects, image transforms.
3. API caching layer — store-and-revalidate close to users.
4. Personalization — read a header, swap a fragment, ship the response.

Where edge struggles

• Heavy CPU work (image processing, ML inference) — limits hit fast.
• Long-lived database connections — most edge runtimes don’t allow raw TCP.
• Workloads needing big native dependencies (FFmpeg, headless Chrome).

The hybrid pattern

Modern stacks combine both:

• Edge in front: routing, auth, cache, personalization.
• Region behind: heavy compute, database queries, background jobs.

Frameworks like Next.js, SvelteKit, Astro and Remix make this split declarative — you just label routes edge or nodejs and the platform handles deployment.

Cost models worth knowing

• Cloudflare Workers: pay per request + per CPU ms — predictable, no per-GB-second.
• Lambda: pay per request + per GB-second — hostage to memory configuration.
• Cloud Run: pay per request + container time — scales to zero, generous concurrency.

A common surprise: a busy Lambda can be 3–10× more expensive than the same workload on Cloud Run, because Lambda bills per-invocation memory time.

TL;DR

Use edge runtimes for the request perimeter and regional FaaS or containers for the heavy lifting. Measure cold starts, P99 latency and bill — not vendor claims.