writing · 2026.06

What I Learned Cofounding an AI Startup

I spent the better part of a year inside an AI data infrastructure startup. From zero to product, from product to customers, through the messy middle of what building actually looks like. There are things I can't say publicly, but there's enough I can say that I think is worth putting down.

Mission alignment isn't a culture nice-to-have. It's a survival requirement.

Early stage is brutal. The product changes, the roadmap shifts, customers want things you didn't plan for, and half your bets won't pan out. The only thing that keeps a team moving through that is people who actually care about what they're building, not just what's on the cap table.

When that alignment breaks down at the leadership level, you feel it everywhere. Decisions start getting made for the wrong reasons and people optimize for how things look instead of whether they work. The work still gets done, but the why disappears, and that matters more than people admit.

AI work has long feedback loops. You're running experiments, iterating on evals, debugging pipelines that take hours to run. Anyone who's just here for the resume line checks out the moment it gets hard. You need people who actually want to solve the problem.

Your cap table will either attract or repel the talent you need.

AI research talent is expensive, equity-sensitive, and has options. Good researchers know what a healthy cap table looks like. They've seen enough term sheets to know when the math doesn't work for them.

If your equity structure is already diluted before you've built anything meaningful, you're going to struggle to compete for the people who will actually move the needle. They'll run the numbers, see the ceiling on their upside, and go somewhere else. Get the cap table right early. It's a hiring problem as much as a finance one, and in AI, hiring is everything.

Chasing fads will leave you permanently behind.

The AI space moves fast enough that by the time a trend is obvious, the first-mover window is mostly closed. If you're building to the current hype cycle, you're already late. You'll spend months on something the ecosystem commoditizes before you ship.

The better play is to find the real problem underneath the fad. The unsexy infrastructure problem. The evaluation gap nobody wants to talk about because it doesn't make for a good demo. The data quality issue that every frontier lab has but nobody's solved cleanly. That's where durable value gets built, not in chasing what's hot right now.

Customer intimacy has to be a founder job, not a delegation.

The gap between what customers say they want and what they actually need only closes through direct contact. There's no substitute for being in the room yourself.

When you filter that signal through BD or sales layers too early, you lose the texture. You get the sanitized version of the problem. You miss the offhand comment that reframes your entire roadmap. You miss the thing they're embarrassed to say to someone they see as a vendor but will say to a founder because it feels more like a conversation.

In AI especially, customers often don't have precise language for what they need. They know something is broken. They don't know what the fix looks like. Figuring that out together is the job, and you can't outsource it and expect the same output.

Evaluation infrastructure is product, not overhead.

If you can't measure what "better" means, you're guessing, and you're asking your customers to take your word for it.

A lot of teams build the model first and figure out how to measure it later. I think that's backwards. The eval is what tells you whether the last two weeks of work actually did anything, or whether you just moved a number around and convinced yourself it meant something. Without it you end up shipping changes on vibes and hoping the customer doesn't notice when they don't hold up.

Getting this right early paid off for us in a concrete way: when a customer pushed back on a claim, we could point to a metric that correlated with their own internal numbers. That's a very different conversation than "trust us, it's better." It's not glamorous work and it rarely shows up in a demo, but it's the difference between knowing where you stand and guessing.

Know the difference between a partnership and a distraction.

Early-stage BD generates a lot of heat. Intros, MOUs, exploratory calls, LOIs. All of it feels like progress because something is always happening. But not every warm handshake is a real opportunity, and the cost of chasing weak deals isn't just wasted time. It's focus, morale, and engineering cycles scoped to a customer who was never going to sign.

Get good at qualifying fast. What's their timeline? Who owns the budget? What does a yes actually look like on their end? The faster you answer those questions, the faster you can figure out whether something's worth your team's attention or just noise dressed up as signal.

Real partnerships are rare. Treat them that way.