Edge runtimes are out. Fluid compute is in. The post-mortem nobody asked for.

Vercel's documentation in 2024 was an extended argument for the Edge Runtime. Run your code at the user's nearest POP, cold start in single-digit milliseconds, billed per request. The pitch was tight and the demos were good.

The documentation in 2026 is quieter on the topic. Edge Functions still exist. The "Why Edge?" page is gone. The default for new Next.js applications is Fluid Compute, which is Node.js, which is not the edge in any meaningful sense.

What happened

Three things, in roughly this order:

1. The runtime restrictions broke too much. Edge Runtime is V8 with no Node APIs. Half the npm ecosystem assumes Node. Every framework spent a year shimming, every framework eventually gave up on shimming everything.
2. Cold starts on Node got fast enough. Fluid Compute reuses instances across requests, which means the cold-start problem the Edge Runtime was built to solve is mostly gone for high-traffic workloads. Low-traffic workloads still cold-start, but they always will, on any platform.
3. The latency win was smaller than advertised. For most apps, the bottleneck is the database round-trip, not the function-to-user hop. Putting your function at the edge while your Postgres is in us-east-1 is theater.

Why this matters

The serverless edge story sold a generation of developers on a deployment model that turned out to be a niche optimization rather than the default. That's fine — niches matter — but the marketing implied "the default," and that's the wave you're seeing recede.

The contrarian read

Edge runtimes are not dead. They are the right choice for: middleware (auth checks, geo redirects, A/B routing), genuinely stateless computation (signed URL generation, image transformations), and latency-sensitive read-through caches. What changed is that they are not the right choice for most request handlers, and the platforms are tired of pretending otherwise.

The platforms moving people back to Node-class runtimes are not capitulating. They're admitting that the simplest tool is usually the right one. Which is the same lesson the industry learns every five years, in the dialect of whichever vendor is currently teaching it.