Marginal thinking and opportunity cost

Marginal cost is what the next unit costs to make, ship or serve. Not the total divided by the count, which is the average, but the extra: total cost with the unit minus total cost without it. Marginal revenue is what the next unit brings in, which is the price if selling one more does not move the price, and less than the price if it does, because the price cut applies to every unit you were already selling. Foundations session 2 gave you the tool for that second case: on an elastic product the price cut needed to sell one more costs more on the old units than it earns on the new one.

The rule is one comparison. Make the next unit if its marginal revenue is above its marginal cost. Stop at the unit where they meet. Every unit before that point added to profit; every unit after it subtracts. A firm that stops earlier leaves money on the table; a firm that goes further pays for the privilege. Nothing in the rule mentions the average, and that is deliberate: the average cost of a run can be above the price while the next unit is still worth making, and it can be below the price while the next unit is a loss.

Where the average misleads. A product with a large up-front cost and a small per-unit cost has an average that falls with every unit sold. Early on the average is far above the price and a leader reading only averages will call the product a loser. At the margin it was a winner from the first sale, because each sale earned more than it cost to make. The opposite trap is a product whose average is fine but whose next unit is dear, because the team is at capacity and the next order needs overtime, a new courier contract, or a second shift. The average says yes; the margin says the next order is the one that loses money.

Money already spent is unrecoverable on every path, which is exactly why it cannot rank the paths. The reason people count it anyway is that it feels like a debt the product owes them, and stopping feels like admitting the debt will never be paid. It will never be paid either way. Keeping the product running does not recover the listing fees or the photographer's invoice; it only adds next month's costs to the pile. The test that clears the fog is to write both paths on one page with only the costs and earnings that lie ahead, as the figure does. If the pile appears on both, it cancels, and what remains is the decision.

Two things that look like sunk costs are not. A recoverable asset is not sunk: stock that can be returned, a licence that can be resold, a deposit that comes back. Its resale value is a real earning on the "stop" path and belongs in the comparison. Information already bought is not sunk in the useful sense: three months of sales data cost money to collect, and that money is gone, but the data itself changes what you expect the next month to earn. Use what the spend taught you; ignore what it cost.

Opportunity cost is the value of the next best alternative you gave up. It is the honest price of an hour, a real, a server, a desk, and it is nearly always higher than the price on the invoice, because the invoice records what you paid and not what you could have done with it. A meeting with eight people has an invoice cost of zero and an opportunity cost of eight hours of whatever those people would otherwise have done. That is why "it is free" is a claim about the invoice and never about the cost.

Comparative advantage is what opportunity cost says about who should do what. In the figure, Ana writes 4 listings an hour or answers 8 tickets; Bruno writes 1 or answers 4. Ana is better at both, which is called an absolute advantage, and it tempts a manager to give Ana everything. But look at what each gives up. When Ana writes a listing she gives up 2 tickets; when Bruno writes one he gives up 4. Ana's listings are cheaper in tickets, so Ana should list. When Ana answers a ticket she gives up half a listing; when Bruno answers one he gives up a quarter. Bruno's tickets are cheaper in listings, so Bruno should answer. Split that way, the two of them together produce more of both than under any other split, and Bruno has a job even though Ana is faster at it.

The rule generalises far past two people. It is why a country that is worse at making everything still exports something, why a platform does the storefront and the seller does the sourcing, and why a senior engineer who is faster at writing tests than the junior should still not be the one writing them. The question is never "who is best at this"; it is "who gives up the least to do this".

Add one more unit of an input while holding the others fixed and, past some point, each added unit produces less than the one before. One more packer in a warehouse of fixed size adds a lot of throughput; the tenth adds a little; the twentieth gets in the way. That is diminishing marginal returns, and it is the reason marginal cost eventually rises: once the cheap capacity is used, the next unit needs the expensive input, the overtime hour, the rented van. It is a statement about the short run, when something is fixed. Given time to add warehouse as well as packers, the constraint moves.

The data habit it gives you: when a team says "we doubled the ad budget and got 40 percent more sales" the interesting number is not the 40, it is the shape. Was the first half of the budget earning far more per real than the second half? A spend that has hit diminishing returns has a marginal return below its average return, and a report that quotes only the average will keep recommending the next increase long after the next increase stopped paying. Any number in this paragraph is illustrative; the shape is the point.

A physical unit has a marginal cost that is real and visible: the item from the supplier, the packaging, the courier, the commission. It is roughly flat while there is spare capacity and rises when there is not. Pricing it is Foundations session 2's job: the price sits above the marginal cost by an amount the elasticity decides.

A software copy has a marginal cost close to zero. The first copy cost the whole development budget; the millionth costs a few seconds of bandwidth. Shapiro and Varian's argument, as reported, is that this is why information goods are priced by what they are worth to the buyer and not by what they cost to make, and why the same product gets sold in versions at different prices to different buyers. Deep-dive session 3 takes that apart. Tonight the point is only the shape: a fixed cost that is enormous, a marginal cost that is nearly nothing, and an average cost that falls forever.

A prediction from a model sits between the two, and the position matters. Training the model is a fixed cost, like writing the software. But each answer the model gives is not a copy; it is a computation, metered per token in and per token out, and it costs something every single time. Small, but not zero, and multiplied by every request. That means the average cost of an inference does not fall to nothing the way a software copy's does; it falls towards the per-token floor and stops. Every decision about whether to build, rent or switch off a model comes down to that floor times the volume against the fixed cost, and deep-dive session 4 puts dated list prices on it. No price is quoted here on purpose: the shape is the lesson, and the numbers move.

Economies of scale are what a large fixed cost does to the average: the more units it is spread across, the lower the cost per unit, until the marginal cost stops falling and takes over. They are the reason big producers can undercut small ones and the reason a market with a large fixed cost tends towards a few big players. They are also the reason "our cost per unit fell 30 percent" can be entirely a statement about volume and not at all a statement about efficiency; an illustrative figure, and a common confusion.

Coase, in "The Nature of the Firm", Economica 4(16) 386-405, 1937, asked a question that sounds naive and is not: if markets coordinate so well, why do firms exist? Why not hire every hour and buy every part on the open market? His answer is that using a market has costs of its own, finding a counterparty, agreeing a price, writing and enforcing a contract, and a firm exists wherever doing the thing inside, by instruction, is cheaper than buying it outside, by contract. The boundary of the firm sits where those two costs meet. It is opportunity cost applied to organisation: every task is either coordinated by a price or by a manager, and the cheaper coordinator wins. Deep-dive session 9 uses exactly that line to decide whether to build, buy or rent a data platform.