Every AI company is underwriting a price cut it never negotiated.

a16z calls it LLMflation. For a model of equivalent performance, inference cost has been falling roughly tenfold a year. What was expensive in 2023 is close to free now.

Sounds like good news. It is only good news if you are built to collect it.

Price your product on what inference costs today and one of two things happens.

Costs fall, your pricing holds, and your margin quietly doubles.

Or a competitor passes the saving to the customer and you find out how loyal yours actually are.

Nothing in your positioning decides which one you get. Your architecture does.

Locked to one frontier model and you pay frontier prices forever, whatever happens underneath. Able to swap the model beneath your product and you inherit every price drop the day it lands.

Most founders treat model choice as a quality decision. It is a margin decision with a quality constraint, made once at build time with nobody from finance in the room, and then billed every month for years.

Are you priced on what inference costs today, or on what it will cost eighteen months from now?

Krishna Lakamsani | Building A Foundery | We build cash-positive companies, not valuations| I write at the intersection of AI, capital, and the future of human work: lessons, reflections, and honest takes from the founder-investor seat.