Walk through what an AI model company's income statement probably looks like. Where does training compute belong, and what revenue does it need to break even?

Revenue of Rs 3,000 crore, inference cost of Rs 1,350 crore and an operating loss of Rs 1,850 crore. Treat training as research and development, not cost of revenue: it builds the next model rather than serving today's customers, and the gross margin is then 55%; in cost of revenue it would be minus 12%. With inference at 45% of revenue and Rs 3,500 crore fixed, breakeven revenue is about Rs 6,364 crore, more than double today.

The way any software company does, then ask what is different. Revenue by line, subscriptions and API. Cost of revenue, which is the cost of serving those customers today: here inference compute, the electricity bill of answering every question. Gross profit is what serving customers leaves; everything below it, research, people, selling and overhead, is the cost of being in business and building what comes next. The unusual feature is that one research line, training compute, is larger than inference.

Think of a film studio. The cost of making the next film is not the cost of selling tickets to the current one. Training builds the next model, which may earn revenue for a year or two; it does not serve today's customer, so it belongs in research and development below gross profit. That gives a gross margin of 55%. Put it in cost of revenue and the same business reports minus 11.7%. The economics have not changed; the classification has, which is why the first question about any reported gross margin is what is in it.

There is a third choice worth naming. Some accounting frameworks allow development costs to be capitalised when strict conditions are met, and then amortised over the asset's life. If training were capitalised, EBITDA would be Rs 150 crore, positive, while operating profit would carry the amortisation. Whether a model qualifies, and over what life, is a judgement for the applicable standard and the auditor, so confirm the current treatment before relying on any reported EBITDA in this sector.

Separate costs that move with revenue from those that do not. Inference is 45% of revenue, so each extra rupee of revenue contributes 55 paise. Training, staff, selling and overhead total Rs 3,500 crore. Breakeven revenue is Rs 3,500 crore divided by 0.55, about Rs 6,364 crore, a little over twice today's revenue.

Then state what that breakeven assumes, because it is fragile in both directions. Training is fixed only if the company stops racing to build bigger models; in practice it tends to rise with ambition. Inference cost per unit of revenue tends to fall as models and chips become more efficient, which lowers the breakeven. A good answer gives the number and then says which of those two forces the business plan is betting on.