One machine

Gives way next

Someone really did this

Source:

Frozen at users

person a day

Requests, busiest second: –
Machines: –
Cloud bill, a month: –
Per user, a month: –

Assumptions

Every number the model uses. Change any of them.

This is a model built from common rules of thumb, not a quote for your system. If your product is heavier or lighter than these, type your own figures and everything redraws.

Prices

AWS on-demand list prices, US East (N. Virginia), checked 11 November 2026. Real bills at scale are lower: large customers commit ahead and negotiate.

About

Will it scale, and what will it cost?

Drag from one user to a billion. The diagram is the system a careful team would be running at that size; the orange dots are requests moving through it. Press Freeze the system and keep dragging to see why each new part appears: the queue builds, responses slow, and requests start to fail.

What is real, and what is a model

The model. Machine counts come from the assumptions (how many requests a user makes, how many a server can serve). They are standard rules of thumb, listed and editable under Assumptions.
The slowdown. Response times under load use the M/M/1 queue: mean time is the service time divided by one minus the load. That is exact for that queue and is why a system at 95 percent is not “nearly fine”.
The prices are AWS public list prices, linked under Assumptions.
The companies on the rail are real, placed where their own engineers said they were, each with its source. They often did it with far fewer machines than the rules of thumb suggest. That is the point: architecture and care matter more than the rules.

Keys

←→ fewer and more users · Shift with an arrow jumps a whole zero · Space grow or pause · F freeze · Esc close

Made by Pratyush Shivam. Typeface: Geist, SIL Open Font License.