There’s a new question showing up in pitch meetings that didn’t used to exist. it’s not “what’s your roadmap.” it’s “what happens when the model you’re built on ships your feature for free.”
Investors who once asked about your roadmap now open with questions about your data moat, your inference costs, and whether your product survives when a foundation model ships your feature natively. that’s not a minor shift in tone. it’s a completely different kind of fear being priced into every AI startup right now, and most founders are still pitching like it’s 2023, like the foundation you’re building on is bedrock instead of a moving target that occasionally decides to become your competitor overnight.
I think about this more than i’d like to admit. every AI product right now is a bet that a gap will stay open long enough to build a company inside it. the gap between what the base model can do out of the box and what a specific business actually needs. some of those gaps are wide and durable, workflow, trust, integration, the boring stuff that doesn’t get closed by a model card update. some of those gaps are one product announcement away from not existing. you don’t always know which kind of gap you built in until the announcement happens.
What’s strange is how little this gets said out loud at demo day. everyone’s pitch still sounds like they’re describing a fortress. nobody’s slide says “here’s the specific way we could become irrelevant in six months and here’s why we think we won’t.” that sentence would probably kill the round. it’s also the most honest sentence most AI founders could write right now.
The interesting question isn’t whether this risk is real, it obviously is, it’s what founders are actually supposed to do with it. you can’t out-engineer a foundation model’s roadmap. you don’t have visibility into it and you’re not supposed to. so the only real defenses are the unglamorous ones. the specific data you have that the model doesn’t. the workflow you own that lives outside the chat window. the trust a customer has already put in you that a better raw model doesn’t automatically transfer. none of that is exciting to say in a pitch. all of it is the only thing that actually holds up.
I don’t think “AI startup” is a stable category anymore, not the way “SaaS company” was a stable category for a decade. i think it’s closer to an arbitrage window. you’re building inside a gap that a much larger, much better funded lab could close whenever it decides your feature is worth shipping to everyone for free. that’s not a reason not to build. it’s a reason to stop pretending the gap is permanent and start being honest, with investors and with yourself, about how much of what you’ve built would survive the day it isn’t.
The founders who’ll actually make it through this era aren’t the ones with the best model integration. they’re the ones who already know exactly which parts of their product don’t depend on the model staying dumb.