Serverless-Only

The "I refuse to operate any servers" path. Possible end-to-end in 2026 across multiple providers, with real tradeoffs.

The platforms

Cloudflare-flavored (the "all CF" stack)

• ★ Cloudflare Workers — V8 isolates; fast cold starts; truly global. The default for new edge-first apps.
• Cloudflare Pages + Workers — combined static + server.
• Cloudflare D1 — SQLite at the edge.
• Cloudflare R2 — S3-compatible, no egress fees.
• Cloudflare KV / Cache API — fast reads at the edge.
• Cloudflare Queues / Workflows — async work.
• Durable Objects — single-instance stateful actors; multiplayer / chat / counters.
• Vectorize / Workers AI — AI without leaving CF.

AWS-flavored (Lambda + DynamoDB)

• ★ AWS Lambda — the originator; 1M req/mo free tier.
• DynamoDB — serverless NoSQL with on-demand pricing.
• API Gateway / Lambda Function URLs — HTTP entry.
• EventBridge / SQS / SNS — async pipelines.
• S3 — object storage.
• AWS Step Functions — durable workflows.
• SST — TS-flavored IaC for serverless on AWS.

GCP-flavored

• Cloud Run — serverless containers (cold-start + autoscale).
• Cloud Functions — function-as-a-service.
• Firestore — serverless NoSQL.
• Cloud Tasks / Pub/Sub — async.
• Workflows — durable orchestration.

Vercel / Netlify-flavored (frontend-driven)

• Vercel Functions — Edge + Node serverless.
• Vercel KV / Postgres / Blob — bundled stores.
• Netlify Functions / Edge Functions — same shape.
• Pair with Neon / Supabase / Convex for the DB layer.

Other

• Deno Deploy — serverless Deno.
• Bun on Workers patterns — emerging.
• Modal / Replicate / Beam — serverless GPU compute for AI.

What you give up

• Long-running processes. Background jobs need workflow engines (Inngest, Trigger, CF Workflows).
• Disk. Tmpfs at best; persistent data lives in object stores or DBs.
• WebSockets at scale — Workers Durable Objects help; raw Lambda struggles.
• Cold starts. Workers are great here; Lambda VPC + Java is not.
• Egress / data costs — every byte leaving your serverless code can cost money.

What you gain

• No machines. No SSH, no patches, no apt update.
• Bursty cost model. Pay for actual usage; idle is free.
• Global distribution. Workers / Vercel Edge / Lambda@Edge.
• Compliance — most platforms come with SOC 2 / ISO / HIPAA.

Common stacks

• All-Cloudflare: Workers + Pages + D1 + R2 + Queues + Durable Objects.
• Vercel + Neon: Vercel Functions + Neon Postgres + R2 (or Vercel Blob) + Resend.
• AWS-native: Lambda + DynamoDB + S3 + EventBridge + Step Functions, deployed via SST or CDK.
• Convex stack: Convex (DB + functions) + Vercel for frontend + Resend.

When serverless is wrong

• Heavy CPU / sustained workloads (training models, big batch jobs).
• Apps with many long-lived WebSockets and real session state (Durable Objects help).
• Strict latency to a non-CDN-cached database.
• Predictable steady traffic where reserved instances dominate.

Pick this if…

• Bursty / unpredictable traffic.
• Side projects, indie SaaS, MVPs.
• Apps where you'd rather pay AWS than a sysadmin.
• You want global low-latency without thinking about regions.

Skip if your workloads are CPU-heavy / sustained, if you have predictable steady load, or if your data has compliance constraints that bite serverless data residency assumptions.