Stop 6 of 10 · Weekly
Cost model
Usage assumptions written down and sourced, a spreadsheet built around the levers the agency actually controls, a sensitivity table, and every rate typed in by Fadi from the public calculator.
The number the board remembers
The Calverton County board will hear one figure about this migration and remember it for three years. It will be wrong, because it was produced in an afternoon from a workload nobody had measured, and it will be quoted back to Fadi at every quarterly review until the programme ends.
A cost model does not become defensible by being more precise. It becomes defensible by being explicit about what it assumed, where each assumption came from, and how much the answer moves when the assumption is wrong.
Assumptions before arithmetic
Every model starts with a table nobody wants to write: the assumptions, each with a source and a confidence.
| Assumption class | Example on this account | Where the figure comes from |
|---|---|---|
| Volume | Records stored and growth rate | County capacity report |
| Shape | Read to write ratio, peak concurrency | Database team estimate, unmeasured |
| Schedule | Non-production environments running hours | Design decision, a lever |
Mark the unmeasured ones. They are what the sensitivity table exists for.
From the Calverton County architecture document and discovery notes in this Project, extract every assumption a five-year cost model would depend on. I am the cloud architect on the account. Give me a table: assumption, the value or range, where it came from in the attached material, and a confidence of measured, estimated or guessed. Anything you cannot source from the attached material goes in as guessed, with the question I should ask the county to turn it into a measurement. Do not produce any costs yet.
Now build the five-year model as a spreadsheet structure I can paste into Excel. Break every line out by metered dimension and quantity, with an empty rate column and an empty monthly cost column, and formulas that roll quantity times rate up to a monthly and annual total. Organise it so the levers the county actually controls are their own input block at the top: commitment term, non-production running schedule, storage tiering policy, instance right-sizing, and data transfer pattern. Every lever changes the totals through the formulas rather than by me editing lines. State the metered dimension for every line precisely enough that I can find it on the public pricing calculator.
Fadi then opens the calculator and types the rates in himself, with the date he read them in the header. That is fifteen minutes and it is the fifteen minutes that make the spreadsheet his.
Add a sensitivity table underneath the model. Take each assumption marked estimated or guessed and show the five-year total when it moves down by a quarter and up by a half, one assumption at a time, with the other assumptions held. Then rank the assumptions by how much they move the total, and tell me which two the model is really resting on. Those are the two I will go and measure before this goes to the board.
The ranking is the point. Most models turn out to rest on one number that was invented in a meeting, and finding that out costs one prompt.