Why a five-person company can run Netflix's architecture for $30 a month
A five-person plumbing company called last month asking whether they needed "edge computing."
The honest answer is no, they didn't need it, but they did need what it enables: an AI voice agent that picks up at 9 PM without a server bill that eats the margin on the call. Five years ago that stack would have cost around $2,000 a month and required a person who knew what DevOps meant. Now it runs for less than what they pay for one phone line.
Vercel Hobby/Pro + Supabase Free/Pro + low-volume LLM calls for a small business voice agent or web product. Same infrastructure that runs much larger applications.
The path your request used to take
Imagine a homeowner in Kansas City pulls up your scheduling page at 7 PM. In the traditional model, the request travels from her browser to a data center somewhere — historically the East Coast or Virginia — gets processed, and travels back. The round trip adds roughly 50–100 ms of latency just from physical distance.
Edge computing means the code runs on a server that's physically close to the customer. Same request, same result, but the round trip drops to roughly 10–30 ms. Multiply across every interaction on a busy site and the difference is the gap between "feels snappy" and "feels broken."
You pay per request, not per server
Edge functions are billed per execution, in fractions of a cent. A small business processing a few hundred website interactions a day pays a few dollars a month for compute. Compare that to running a dedicated server 24/7 — a $50–$200/month bill whether anyone visits your site or not.
That pricing model is what makes the math work for a small business. You don't need to predict traffic. You don't need to overprovision. You don't pay for capacity nobody used.
You pay per request, not per server. That's the whole reason a five-person company can run the architecture a Fortune 500 used to run.
What edge isn't great for
Edge functions are stateless by design. Each request runs in isolation — no shared memory, no in-process caching, no long-running connections. For most small business use cases (form submissions, AI inference, database queries) that's exactly what you want. For workloads that need persistent state (long video transcoding, complex ML training), you want different infrastructure.
Modern edge platforms (Vercel Edge, Cloudflare Workers, Supabase Edge Functions) have eliminated most of what people used to complain about. Cold starts on the leading platforms are now sub-100 ms in most cases — fast enough that a typical user can't tell the function wasn't already running. Execution time limits vary: Vercel Edge streams up to 25 seconds, Cloudflare Workers gives 30 seconds of CPU time, Supabase Edge Functions wall-clock at 400 seconds. None of those are a real constraint for the work a small business is sending through them.
The infrastructure stopped being a wall
Five years ago, the infrastructure was a wall. Building a real-time AI feature for a small business meant negotiating with AWS, hiring a DevOps person, and signing up for a base bill of two thousand dollars a month before you ran a single transaction.
That wall is gone. The stack now runs on rented compute that scales itself, costs almost nothing at low volume, and stays cheap at high volume. The bottleneck is no longer infrastructure. The bottleneck is whether you've done the work to figure out what to actually run on it. Schedule a call if you're trying to figure out where AI fits inside your business and what the actual cost looks like.
Wondering what your stack would actually cost?
If any of this sounded like the week you just had, that's the conversation I have nearly every day with owners in your position.
Sources: McKinsey State of AI 2025 · BrightLocal Local Consumer Review Survey 2024 / 2026 · Invoca Home Services Call Analytics 2024 · Salesforce SMB Trends 2025 · Forrester B2B Buying Journey 2022 · Gartner B2B Buyer Behavior 2024 · Demand Gen Report 2024 · Apten / Hatch Speed-to-Lead Benchmarks 2026 · ReplyOnTheFly Google Review Response Benchmark 2026 · Panko (University of Hawaii) spreadsheet error research · SHRM AI in Onboarding 2024.
