you guys have to read this because it changes the entire math for anyone building on other people's models! adam mosseri just told techcrunch that we should expect capped ai token budgets per engineer — not a theoretical prediction but something already happening at companies like grok and google gemini where they've already hit usage ceilings. think about what that means: every llm-powered app you use, from coding copilots to chat agents, is riding on someone else's infra and that infrastructure provider can decide when your allowance runs out. this isn't a small detail for startup founders because it turns the business model upside down overnight. if perplexity or poe built their whole stack on unrestricted api access and suddenly get capped they have to completely re-engineer how they serve users, which is exactly what they're doing right now by building in tiered systems and caching layers.
and the underlying economic reason is actually kind of fascinating because it ties back to scaling laws that predict training costs grow exponentially with model size! we keep pushing for bigger models but each jump up in parameter count multiplies compute requirements so soon everyone will be running on a token-per-user quota instead of free rein. mosseri's point is essentially saying the era of unlimited llm usage through an api is already over and anyone who builds without planning for it gets crushed by their own scale — which sounds like every startup in this space should have seen coming but nobody did because everybody was chasing training data growth instead. i can't stop thinking about how many well-intentionated ai apps are going to become unplayable overnight when the capping actually hits, and mosseri is just pulling back the curtain on a cliff that's already beneath everyone's feet. read the full thing if you want all the details because this probably affects your work more than you realize right now.
Source: https://techcrunch.com/2026/07/14/metas-adam-mosseri-says-ai-token-budgets-could-soon-be-capped-per-engineer/
and the underlying economic reason is actually kind of fascinating because it ties back to scaling laws that predict training costs grow exponentially with model size! we keep pushing for bigger models but each jump up in parameter count multiplies compute requirements so soon everyone will be running on a token-per-user quota instead of free rein. mosseri's point is essentially saying the era of unlimited llm usage through an api is already over and anyone who builds without planning for it gets crushed by their own scale — which sounds like every startup in this space should have seen coming but nobody did because everybody was chasing training data growth instead. i can't stop thinking about how many well-intentionated ai apps are going to become unplayable overnight when the capping actually hits, and mosseri is just pulling back the curtain on a cliff that's already beneath everyone's feet. read the full thing if you want all the details because this probably affects your work more than you realize right now.
Source: https://techcrunch.com/2026/07/14/metas-adam-mosseri-says-ai-token-budgets-could-soon-be-capped-per-engineer/