The math behind edge compute selection

Cloudflare versus Vercel seat models

Cloudflare Workers uses V8 isolates to run JavaScript and WebAssembly. Vercel Edge Functions use a JavaScript environment with a subset of Node.js APIs. Cloudflare provides a $5 monthly plan that includes 10 million requests and 30 million CPU-milliseconds. This plan does not charge for data transfer. Vercel Pro starts at $20 per month per user. Vercel also bills for invocations, active CPU time, and data transfer. A five-person team on Cloudflare pays $5 for the account. That same team pays $100 for Pro seats on Vercel. Cloudflare wins on scale. Cloudflare’s free tier provides 100,000 requests per day and 10 ms of CPU time per invocation. Vercel’s Hobby tier is limited to non-commercial use. Vercel integrates tightly with Next.js to provide framework-aware routing. Cloudflare requires the OpenNext Cloudflare adapter to run Next.js on its runtime. Vercel also provides automatic framework detection and git-based preview deployments.

AWS Lambda and high compute workloads

AWS Lambda charges for the total duration of execution. This duration includes both the time the code uses the CPU and the time spent waiting for I/O. Cloudflare Workers only bills for active CPU time. This difference creates massive price gaps for high-compute workloads. A Worker that serves 15 million requests per month, with an average of 7 milliseconds of CPU time per request, results in an estimated monthly cost of $21 when you include the $5 monthly subscription. For a workload with 5 million image processing requests and 2.5 seconds of CPU time per task, AWS Lambda (x86) costs $202.47. Cloudflare Workers costs $254.40 for the same task. Lambda costs 25% less. For a mobile app gathering 10 billion requests with 15 ms of CPU time per task, Cloudflare costs $5,969.40. AWS Lambda (x86) costs $6,556.68 for that volume. Cloudflare costs 10% less for this high-volume, low-processing workload.

Experiments by Cascadeco show that switching to Arm-based Lambdas provides performance uplift and cost savings. Python workloads benefit significantly from a move to Graviton2. In every configuration tested, work per dollar favors Graviton2. However, memory-heavy workloads that require absolute performance might perform better on x86-based Lambdas.

Consider a logistics company with 10,000 vehicles emitting telemetry once an hour. To perform inference on 7.44 million messages per month with 1 GB of memory and 2 seconds of execution, AWS Lambda (x86) costs $263.92. This total includes $1.49 in request charges, $248.00 in compute charges, and $14.43 in SnapStart charges. AWS SnapStart provides up to a 90% reduction in cold start latency for Java functions at no additional cost. Lambda allows memory configuration from 128 MB to 10,240 MB. Cloudflare Workers fixes memory at 128 MB per isolate.

Latency and memory limits

Cold start latency varies by architecture. Cloudflare Workers cold starts typically stay under 5ms. Vercel and Lambda cold starts often range from 100ms to 800ms. I would skip Cloudflare if your application needs massive memory. Cloudflare limits memory to 128 MB per isolate. Lambda allows users to allocate up to 10,240 MB of memory. Cloudflare limits subrequests to 50 per invocation on the Paid plan. Vercel and Lambda have different limits.

Cloudflare R2 provides $0.015 per GB-month storage with zero egress fees. AWS S3 charges for outbound data transfer. For a workload shipping 40 TB of large objects per month with 60 million reads, R2 costs about $164. A specialist CDN might offer better delivery guarantees for large-file distribution. What is the actual p95 CPU-ms per route?

Memory is fixed.