Serverless is a billing and operations model: you run code without managing servers, usually in a cloud region. Edge computing is a location: code runs close to users or devices. Edge functions such as Cloudflare Workers combine the two, and in 2026 the bigger question is where AI inference should run.
Updated October 2026
What changed in 2026
- Vercel now recommends migrating functions from its edge runtime to Node.js, with both running on its Fluid compute model (Vercel docs).
- AWS stretched serverless in both directions at re:Invent 2025: Lambda Managed Instances run functions on EC2 capacity in your account, and Lambda durable functions can pause for up to one year without paying for idle compute (AWS Compute Blog).
- GPU inference moved to the network edge. Akamai launched Inference Cloud with NVIDIA RTX PRO 6000 Blackwell GPUs on 28 October 2025, starting in 20 locations (Akamai).
- Cloudflare's network now lists 348 cities (Cloudflare), and paid Workers can use up to 5 minutes of CPU per request (Cloudflare docs).
Serverless and edge are not opposites
The old framing set these up as rivals. They answer different questions.
- Serverless answers "who runs the servers?" The provider does. You ship functions and pay per use.
- Edge answers "where does the code run?" Close to the user or the device.
So you can have serverless in a region (AWS Lambda), serverless at the CDN edge (Cloudflare Workers, Lambda@Edge), or non-serverless at the edge (a GPU box in a factory).
The three edges in 2026
| Edge type | Where it runs | Typical latency to user/device | Good for | Examples |
|---|---|---|---|---|
| CDN edge functions | Provider points of presence worldwide | Low, near the user's city | Auth, redirects, A/B tests, personalisation, light APIs | Cloudflare Workers, CloudFront Functions, Lambda@Edge |
| Network-edge GPU inference | Provider sites with GPUs | Low, regional | LLM and vision inference for apps | Cloudflare Workers AI, Akamai Inference Cloud |
| On-site physical edge | Your plant, store, vehicle or robot | On the device or LAN | Real-time control, vision, robots, offline operation | NVIDIA Jetson, industrial PCs, Azure Local, AWS Outposts |
Most "serverless vs edge" articles only cover the first row. If your data comes from machines rather than browsers, the third row is the one that matters.
Serverless vs edge: side by side
| Regional serverless | Edge functions | On-site edge | |
|---|---|---|---|
| You manage servers | No | No | Yes, or a platform does it for you |
| Billing | Per request and duration | Per request and CPU time | Hardware plus software licences |
| Cold starts | Possible | Usually small (isolates) | None; always on |
| Runtime limits | Longer runs; durable functions can wait up to a year | Tight CPU and memory caps | Set by your hardware |
| Works offline | No | No | Yes |
| Data stays on site | No | No | Yes |
Edge function limits you should know
Edge functions trade power for proximity. AWS's own comparison shows the gap (AWS docs):
| CloudFront Functions | Lambda@Edge | Cloudflare Workers (paid) | |
|---|---|---|---|
| Runtimes | JavaScript | Node.js, Python | JavaScript, TypeScript, WebAssembly |
| Max execution | Sub-millisecond | Up to 30 s | Up to 5 min CPU per request |
| Memory | 2 MB | 128 MB viewer / 10,240 MB origin | 128 MB per isolate |
| Network access | No | Yes | Yes |
Cloudflare figures from Cloudflare docs. These caps are fine for routing and personalisation. They are too small for most AI models, which is why inference uses separate GPU services.
Where should AI inference run?
This is the real 2026 decision. Use three questions.
- Where is the data born? Browser and app data can go to a network-edge GPU. Camera and sensor data from a plant should usually stay on site.
- What if the network drops? If a missed decision stops a line or harms a person, run it on site.
- How big is the model? Large language models need regional or network-edge GPUs. Small vision and language models fit on edge devices like NVIDIA Jetson.
Many teams end up with a split: a small model on site for fast, safe decisions, and a larger model in the cloud for hard cases and retraining.
When to choose serverless, edge functions or on-site edge
| Choose | When |
|---|---|
| Regional serverless | Spiky back-end workloads, event processing, long-running workflows |
| Edge functions | Logic that sits in front of a web app and must be fast for global users |
| Network-edge inference | AI features in consumer or SaaS apps that need low latency without your own GPUs |
| On-site edge | Machines, robots, cameras, vehicles; anything that must keep working offline or keep data local |
Where it's heading: 2026 → 2030
| When | What it looks like |
|---|---|
| Around 2020 | Serverless means AWS Lambda in a region. "Edge" means CDN caching plus small edge functions. |
| 2026 (now) | Vercel recommends moving from its edge runtime to Node.js; AWS adds Lambda Managed Instances and durable functions; Akamai and Cloudflare sell GPU inference at the network edge. |
| By 2028 (expected) | One deployment tool places each function or model in region, network edge or on site based on latency, cost and data rules (our view). |
| By 2030 (expected) | Most physical AI decisions run on site, with the cloud used for training, fleet policy and audit (our view). |
How NexaStack fits
NexaStack covers the on-site edge that CDN platforms do not reach. It deploys, updates and monitors AI models and agents across plants, stores and vehicles, and keeps them running when the link drops. Each decision is logged and governed, so you can see which model version acted and why. Read our guides to MLOps for edge AI and edge vs fog vs cloud computing, or see the NexaStack Edge AI solution.
Frequently asked questions
Is serverless the same as edge computing?
No. Serverless describes who manages the servers. Edge describes where code runs. Edge functions such as Cloudflare Workers are both.
What is an example of serverless computing?
An AWS Lambda function that resizes an image when it lands in a storage bucket. You pay only while it runs, and AWS handles the servers.
Is serverless better than Kubernetes?
Neither is better in general. Serverless suits spiky, event-driven work with little ops effort. Kubernetes suits steady workloads, custom runtimes and on-site edge clusters where you need control.
Why are some teams leaving serverless?
Common reasons are cost at steady high volume, runtime limits and cold starts. AWS's Lambda Managed Instances, launched at re:Invent 2025, target the steady-volume case (AWS).
What are the latest serverless and edge computing trends in 2026?
GPU inference at the network edge (Akamai, Cloudflare), Vercel steering developers from its edge runtime to Node.js, longer-running serverless workflows on AWS, and AI models running on site for physical operations.
Find the right place for your AI inference
Bring one AI workload and we will map whether it belongs in a region, at the network edge or on site, and how NexaStack would run the on-site part. Book a free edge AI inference placement assessment →
