Seed rounds got smaller, slower and considerably more sensible.

Anyone raising a seed round now will hear the same three questions. What is the gross margin once inference is paid for. Which existing budget line does the purchase come from. And what happens to the product if the underlying model improves by another generation.

None of these were standard two years ago. All three are reasonable.

The margin question

AI-native products carry variable cost in a way that classical software did not. A company with eighty percent gross margin and a company with thirty five percent are different businesses even at identical revenue, and the second one needs a different plan.

The budget question

Products that create a new budget category sell slowly. Products that replace an existing line item sell at the speed of a renewal cycle. Investors have relearned this and are pricing it.

The model risk question

If a general model improvement would make the product redundant, the company is renting a capability rather than building one. Founders who answer this well point to proprietary data, workflow depth or regulatory position. Founders who answer it poorly point to their prompt engineering.

The effect on the market

Fewer rounds, slower, at flatter valuations, with more of them surviving to a second round. Founders find it unpleasant and most of the ones who have raised in both environments prefer this one.