Hook
It was 9:03 a.m. Pacific time when the first request hit the new test harness at the CloudForge lab, and the response from a Vercel Edge Function arrived in 1.2 ms. A few seconds later, the same payload routed through AWS Lambda took 45 ms to warm‑up and return. The numbers were stark, the reaction immediate: developers across the globe were watching.
Here's the thing: for the first time since the term "serverless" entered the mainstream, we have a side‑by‑side, public benchmark that treats cold start, sustained load, and price per million invocations as equals. The result? Edge platforms are not just faster; they're cheaper for a growing class of workloads.
Context
On May 20, 2026, the Open Compute Initiative (OCI) announced the Edge‑First Compute Challenge, a three‑month contest inviting vendors to submit raw performance numbers for a set of 12 micro‑services. The rules were simple: each service must be deployed on the vendor’s default serverless or edge offering, then measured against identical traffic patterns using a 10 Gbps synthetic load generator located in three geographic zones – North America, Europe, and APAC.
But look, the timing was no accident. Last quarter, several high‑profile outages at traditional cloud data centers reminded enterprises that latency spikes can cripple revenue. Simultaneously, the rise of 5G and Web3 applications demanded sub‑10‑ms response times. The market was primed for a data‑driven showdown.
Let's be honest: most developers have been choosing between Lambda, Azure Functions, and Google Cloud Run based on familiarity, not hard data. This benchmark forces a conversation about real performance, not marketing hype.
Technical deep‑dive
The test suite consisted of three categories: stateless API gateway, image thumbnail generator, and real‑time chat relay. Each was written in Node.js 20, Go 1.22, and Rust 1.73 to capture language‑specific quirks. All functions were limited to 512 MiB memory and 2 vCPU equivalents where the platform allowed.
Cold‑start latency was measured by invoking a function after a 30‑minute idle period. Warm‑start latency used a 1‑second interval between calls. Throughput tests ran 5 000 concurrent invocations for 60 seconds, while cost calculations used the published per‑invocation pricing as of May 1, 2026.
"We wanted a test that mirrors what a real developer sees in production – not a synthetic micro‑benchmark that only a handful of engineers can reproduce," says Mira Patel, VP of Engineering at CloudForge.
Here are the headline numbers (average across the three regions):
- Cold start (ms): Cloudflare Workers 3 ms, Vercel Edge 1.2 ms, Fastly Compute@Edge 2.5 ms, AWS Lambda 45 ms, Azure Functions 38 ms, Google Cloud Run 30 ms.
- Warm‑start latency (ms): Edge platforms hovered between 0.8 ms and 1.5 ms; serverless stayed between 5 ms and 12 ms.
- Throughput (req/s per instance): Edge functions sustained 12 000 req/s, while Lambda and Cloud Run topped out at 4 500 req/s before throttling.
- Cost per million invocations (USD): Vercel Edge $0.25, Cloudflare Workers $0.28, Fastly $0.30, AWS Lambda $0.40, Azure Functions $0.38, Google Cloud Run $0.35.
What's interesting is the variance across regions. In APAC, Vercel Edge’s latency rose to 1.7 ms, still well under the 6 ms median for serverless. The cost advantage persisted because edge providers price per‑edge‑node execution rather than per‑GB‑second, which aligns better with bursty traffic.
From a developer’s standpoint, the biggest surprise came from the Rust functions. On Fastly, they hit sub‑1 ms latency consistently, while the same code on Lambda hovered around 8 ms. The difference stems from Fastly’s use of WebAssembly System Interface (WASI) that eliminates the need for a language runtime shim.
Impact analysis
The data reshapes three major groups.
Start‑ups building consumer‑facing apps can now justify moving the bulk of their API layer to the edge. The cost savings alone – roughly $0.15 per million invocations – translate to $150 k annually for a product that processes 1 billion requests a year.
Enterprises with legacy monoliths will feel pressure to refactor. While edge platforms excel at short‑lived stateless workloads, they still lack the deep integration with VPCs that large firms rely on. However, the throughput edge wins suggest a hybrid model: keep heavy data processing in the cloud, push latency‑critical paths to the edge.
Cloud providers are the ones most threatened. AWS’s 45 ms cold start is now a clear disadvantage. Even with its massive ecosystem, the numbers could accelerate the migration of developers toward newer players. Azure and Google are already rolling out “instant‑warm” options, but the benchmark shows they still trail edge‑native offerings.
But look, the story isn’t all about speed. Security teams note that edge runtimes isolate code at the node level, reducing blast radius. On the flip side, the distributed nature of edge introduces new surface‑area for supply‑chain attacks. Vendors responded with signed WASM modules, but adoption is still early.
My take
I'm convinced that 2026 marks the year edge computing stops being a niche for static assets and becomes the default for any latency‑sensitive API. The benchmark proves that the performance gap is no longer a rumor; it’s measurable, and the economics back it up.
That said, I don’t see serverless disappearing. For batch jobs, data pipelines, and workloads that need deep integration with managed databases, the classic cloud still offers unmatched convenience. The future will be a mosaic: edge for the front door, serverless for the back office.
Looking ahead, I predict three trends:
- Major cloud players will acquire or partner with edge specialists to close the latency gap.
- Tooling will evolve to let developers write once and deploy everywhere, with compilers that target both WASM and native VMs automatically.
- Pricing models will shift toward “pay‑for‑edge‑node‑seconds,” making cost predictability a selling point for edge providers.
Developers who ignore these signals risk building on a platform that will feel sluggish three years from now. The data is here; the choice is yours.
More from Dev Tools: Why the Dev Community Is Splitting Over the New GraphQL‑SQL Hybrid • Beryl Programming Language Surges as Edge AI Takes Off
