Building got cheap. Knowing what to build did not.

The hardest part of starting an AI company used to be building the product. It is now the cheapest part. What stayed expensive is knowing which problems deserve a company, and finding out before the money is spent. We built EMAR Intelligence around that second problem, because what you learn solving it carries to the next company.

A feature is easy to build and easy to lose

In 2023 a wave of startups sold one capability: upload a PDF and ask it questions. The products worked. Small teams built them in weeks, most on a model they rented from OpenAI by the call. On October 29 of that year OpenAI added PDF upload to ChatGPT itself, and what had been a company became a menu item.

Nothing went wrong in the engineering. Those teams built what they set out to build, quickly and well. The trouble sat upstream of the code. Every competitor had the same model, at the same price, on the same day, so the ability to build bought no advantage that outlasted a release cycle. When anyone can build the thing, the thing is not where the value sits.

Most startups still spend as if building were scarce

A new AI company still tends to be organized the way software companies were organized when building was the bottleneck. Raise money, hire engineers, ship fast, and look for the customer afterward. Each step made sense when a working product took a specialized team a year. It makes less sense when a working product takes a month, because the month was never the risk. The risk is the question the product was supposed to answer: does anyone have this problem badly enough to change how they work?

That question is slow and unglamorous to answer. It means finding the people who have the problem and learning what it costs them and what they do about it today. Then it means asking them to commit something real before there is anything to demo. Time, data, or money. Polite interest does not count. Most ideas fail this test, and a company that builds before running it spends its runway learning what a few weeks of conversation would have told it.

What a single startup cannot keep

Discipline alone does not fix this. Suppose a team does the slow work. They learn how to find the people with a problem, how to run a proof that produces a commitment instead of a compliment, and when to stop. Along the way they build things: research, a memory layer that carries a user’s context, agents that do real work, the engineering underneath, a way to reach buyers, a way to operate.

If the company fails, nearly all of that is lost. The team scatters and the code is archived. The next founder starts from zero, with the same rented models as everyone else. If the company succeeds, the knowledge stays inside it and serves one product. Either way, the most valuable thing the effort produced gets used once. That thing is the ability to tell a good problem from a bad one quickly.

A studio is a way to use it more than once

A venture studio is the structure that keeps what a single company would lose. We run every idea through the same five stages: opportunity, prove, build, launch, learn. The last stage feeds the first. A venture we stop at the proving stage still leaves its research and its lessons behind. A venture we launch leaves its technology, its agents, and its operating experience. The next one starts from there.

That is the sense in which company-building compounds. The models will keep changing, and anything tied to one model is worth less each time they do. What carries over is the part that was never the model: knowing who has a problem, a method for proving it, and working pieces a new product can stand on. Some ventures we originate ourselves. Others we co-found with operators who bring a problem we could not have found on our own, which widens the set of problems we see without lowering the bar for proof.

What would prove us wrong

This is a thesis, and it should be read as one. We have two ventures in stealth and one area we are exploring. We have not launched a company yet, so the claim that each venture starts further along than the last is a prediction. It is not yet a record.

Studios also have a mixed history. The usual failures are well known: attention spread across too many ideas, founders who own too little to care enough, and shared infrastructure that becomes a tax on the companies it does not fit. The structure does not protect us from any of those. What it gives us is a short loop, so we find out sooner when we are wrong.

So here is how to check us. Watch whether our second and third ventures reach a proven problem faster than the first did. Watch whether we stop things, and whether we say so. If every idea we start becomes a company, we are not proving anything. And if a later venture ends up rebuilding what an earlier one already built, then the compounding was a story we told ourselves.