The cost of an answer is a product decision

There is a sentence that has shaped a decade of Indian consumer software, and it is almost right.

"India cannot be monetised."

The evidence for it is real. Indian consumers do not pay what American consumers pay, the gap is large, and it closes at the speed of income growth rather than at the speed of product improvement. Meanwhile a frontier model stack costs exactly the same to run in Delhi as it does in Denver, because the bill is denominated in tokens and tokens do not care where anyone lives.

Set a consumer subscription at what the Indian market will bear, put a frontier stack behind it, and you are structurally loss-making from the first user. Not at low volume. At every volume. Scale makes it worse.

So the industry concluded that the category does not work, built for audience instead of revenue, and waited for advertising or for model prices to fall.

The narrower version

The correct statement is one clause longer. India cannot be monetised at that cost base.

That sounds like hair-splitting. It is the entire company.

There are two numbers in the equation. What a user pays, and what it costs to answer them. The first is a market fact. It is set by incomes, by competition and by habit, and no amount of cleverness moves it much in the short term.

The second is not a market fact. It is an engineering decision, and it is the only variable on the board that a founder actually controls.

Almost everybody treats it as fixed and waits. We treated it as the product and spent three years on it.

What that actually meant

Four things, none of them glamorous.

Our own small models rather than rented large ones. Not for independence as a principle, but because the alternative is renting your cost structure from somebody whose incentives point the other way. We build the models. When we make one cheaper, we keep the saving.

A router that escalates only when it must. Every request is scored for difficulty before anything answers it. The overwhelming majority resolve on a small model and never touch a large one. The reflex of sending everything to the biggest available model is the most expensive possible way to be wrong.

Structured results instead of long-form generation. This is the one people miss. A chat-first product generates paragraphs, and output tokens are where the bill lives. A search-first product returns a structured result with a summary on top. The output has a hard ceiling instead of compounding through a conversation.

Heavy reuse of what repeats. Job queries are not personal in the way conversations are. A large share of them are the same question asked by different people, and a large share of traffic never needs recomputing at all.

None of these is a research breakthrough. Together they put us roughly one order of magnitude below where a chat-first stack sits, and that is the difference between a business and a subsidy.

The part that compounds

Here is the consequence that took us longest to appreciate.

When answering a question costs almost nothing, adding a surface costs almost nothing.

Signal did not add inference cost. Neither did Luma, or Mail. They all run on the same models, the same router and the same infrastructure, which was already built and already paid for. Each new surface is a fixed engineering cost and then approximately free to serve.

For a company whose costs scale with usage, a sixth product is a sixth set of bills and a reason to be careful. For us it is a reason to ship. The newest surfaces are the most efficient things we run, not the most expensive, and that is a structural inversion of how a software company normally ages.

Why nobody else did it

Because it is three years of unglamorous work that produces no announcement, in service of a market everybody had already written off.

If you believe India cannot be monetised, optimising the cost of an Indian query is a strange thing to spend three years on. If you believe the cost base is the only thing wrong with the sentence, it is the only thing worth spending three years on.

We spent them. That is the whole disagreement.