Vocabulary,
then a worksheet.

By the end you will be able to put any software cost into a defensible box, and you will know the three classical and three behavioural concepts every NPV calculation depends on.

Every software cost lives in a 2 × 2 × 2 box.

A line item like "annual Datadog subscription" is Fixed × Indirect × Recurring. Saying so out loud already reveals what raising it or moving it would mean.

The cost question that scales —
and the one that doesn't.

In modern SaaS, the trend is unmistakable: variable costs grow as a share of total cost. A 2010 software product was 80% fixed cost. A 2026 AI product can be 70% variable — token spend dominates the P&L.

If your costs are mostly fixed, every additional user is nearly free. If they are mostly variable, every additional user has a price tag — and your business model must reflect it.

The cost you can blame on a product —
and the cost that everyone pays.

Allocation is half art, half politics. The same dollar can be made direct or indirect by where it sits on the org chart, and that changes which manager has the incentive to control it.

One-time vs ongoing —
and the great migration of the 2010s.

The cloud era was, in accounting terms, a giant CapEx → OpEx migration. The AI era now extends it down to the request: you no longer rent a server by the hour, you rent a forward-pass by the token.

The most important cost engineers never name.

If your senior engineer spends a week refactoring a logging library, the opportunity cost is not zero — it is whatever the next-best use of that week would have produced.

The trap that keeps dead projects alive.

A sunk cost is money or effort already spent that cannot be recovered. The economically correct response to a sunk cost is to ignore it.

In practice we don't. We keep funding the failing migration because "we've already paid for the licences." We hold onto the rewrite because "we're 70% done." The sunk-cost fallacy is the most expensive behavioural pattern in software organisations.

If you wouldn't start this project today knowing what you know now — the right answer is to stop it, regardless of what you've already spent.

For pricing and capacity, marginal is the only thing that matters.

Engineers pricing an API often quote the average cost. Customers buying access incur the marginal cost. Mismatch here is a familiar trap — quote a low average price, watch your margin disappear when the customer hits the variable-cost tail.

Five phases. Five buckets of money.

Most software cost lives in ③ and ④. Most engineering attention lives in ②. That is the central misalignment of our profession.

Sixty to eighty percent of a software product's lifetime cost is incurred after 1.0 ships.

The dollar of effort you spend writing tests today is buying you 3–4 dollars of maintenance later. Refusing to write them is borrowing at a usurious rate.

Total Cost of Ownership (TCO).

A TCO calculation is honest only if it includes every category we have named so far — fixed and variable, direct and indirect, non-recurring and recurring — across all five lifecycle phases, including the opportunity cost of the engineers' time.

It is the standard unit of comparison whenever we choose between alternatives. Lecture 5 will turn this into a numerical decision. Today we just learn to construct the worksheet.

If two alternatives look similar but you can only compare one phase, you are not comparing TCO — you are comparing convenience.

The new line items.

A startup's pitch deck that lists "infrastructure" as a single number is not yet thinking about the AI cost stack.

Anatomy of a 100,000-task-per-month AI agent.

Prompt caching alone could cut the input-token line by 80% — saving $1,600/month, or 11% of TCO. We will quantify caching in Lecture 13.

Five years, ten categories, one decision.

Pair up. Pick a project from your job — real one preferred. Fill in any cells you can defend. We will not discount today (that's Lecture 3); just get the cash flows down.

Tomorrow we discount.

Today you have learned to name every cash flow. But $1 spent in Year 5 is not the same as $1 spent today. Tomorrow we make that explicit.

In Lecture 3 we introduce Present Value and Future Value, the time-value of money, and the discount-rate machinery that turns the worksheet you just built into a single comparable number.

Bring your TCO worksheet to Lecture 3 — we will discount it together as the first warm-up.

Build a complete TCO.

1. Choose a software product or feature you have personally worked on (or are working on). It can be commercial, internal, or a side project.
2. Build a complete 5-year TCO worksheet. Include every category from today's lecture. Show your assumptions explicitly — they will matter more than the numbers.
3. In one paragraph, state which category surprised you the most when you wrote it down — and why.

Submit as a single PDF + the spreadsheet to /submissions/HW2/<your-name>/ in the course Git repository.

To read this week.

What today bought you.

1. Every software cost lives in a fixed / variable × direct / indirect × non-recurring / recurring cell.
2. Opportunity cost is the cost you never see on an invoice — and the one NPV is built on.
3. Sunk costs are irrelevant to forward-looking decisions. Don't let them be.
4. Software TCO is dominated by phases ③ and ④ — 60–80% lies after launch.
5. A modern AI product has new line items: tokens, GPUs, vector DBs, inference caches. Budgets that don't name them are wrong by construction.