AI is killing friction

I built our fund’s internal tooling myself. I am not a particularly good engineer. Even a year ago, that sentence would have been absurd. Today it is a throwaway line, and the throwaway quality is the point. Claude Code killed the external software agency for us, the same way it is quietly killing a lot of other things.

That is the trivial end of what is happening. The interesting end is what it means for the companies we back.

There is a type of AI startup that, right now, looks genuinely impressive. A domain-specific model, a clean interface, a workflow that used to take hours compressed into minutes. Architects get a 3D model of a house without spending two days building one. Lawyers get a draft contract. Accountants get reconciled books. The product is beautiful, the demo converts, the users love it.

I admire these companies. I also think the software-only version of them is an endangered species.

The reason is simple. The only thing they are really selling is a software layer over a known workflow. The moat is the head start of being the team that built it first. But the frontier is coming for them. Whatever those models cannot do today, they will do in twelve to twenty-four months, and they will do it inside the general-purpose assistant the customer already pays for. “Faster than a human” stops being a moat the moment the alternative is “faster than your startup, and already open in another tab.”

This is not a prediction about any one company. It is a structural claim. If your product is a thin software layer on top of a workflow a foundation model will eventually learn, you are racing the frontier, and the frontier is better funded than you are.

If the software layer is commoditizing, defensibility has to live somewhere the model cannot just regenerate overnight. Physical assets help: factories, fleets, installed infrastructure, hardware in the field. So does regulatory surface, the kind that takes years of permits, certifications, and slow-built relationships with regulators. Captive operations count too, the sales force or install crew or logistics network that actually moves atoms. And domain expertise fused into the product tightly enough that copying the software does not copy the business. There are others, and the list is not the point. The point is that the remaining moats are things the model cannot simply relearn. None of this is new. What is new is the kind of founder who can credibly go after it.

The same force that is commoditizing pure software is the thing that makes hard, vertically integrated companies finally buildable by small teams.

Inside a company, coordination has always been the tax you pay for doing different things at once. Software, hardware, logistics, regulatory, sales, operations. Each domain wants its own people, its own processes, its own vendors, its own language. The bigger the range, the heavier the tax. Classical economics says this is why firms have boundaries at all, and it is why the 2010s playbook for a hard company was: raise a fortune, hire an army, spend five years stitching the army together.

This is the friction AI is killing. Not by making any one function ten percent faster, which is the boring version of the story, but by expanding what a single person or a tiny team can credibly cover. The architect can also be the factory programmer. The regulatory lead can also draft the software. The founder without a CS degree can ship production code. The range of a small team now approaches the range that used to require a hundred people.

Speed falls out of this almost for free. The bottleneck in hard companies was never ambition, it was the time between knowing what to build and having the right people in place to build it. That gap is collapsing.

A friend of mine founded Gropyus. They build entire apartment buildings, modeled in software, manufactured in a factory, assembled on site. It is exactly the kind of vertically integrated, real-world-footprint company I am describing.

Gropyus was and still is tough to build. Not because the thesis was wrong, but because every one of the domains above had to be stood up from scratch. The modeling software. The factory. The robots. The logistics partnerships. The permits. Each one a project, each one requiring specialists, each one taking years. Katerra and a graveyard of other full-stack construction companies tried similar things in the last cycle, and most of them did not make it.

The honest version of my thesis is not that Gropyus was easy. You still need exceptional founders to build a company like that. What is different is the type of exceptional. Five years ago, it took founders with almost mythical range, operators who could personally hold software, hardware, regulatory, and operations in their heads at once, and the institutional money to put an army of specialists under them. AI gives some of that range to a deep domain expert with a small team. The bar is still high. The shape of the bar has changed.

This does not mean every hard company will work. Most will not. Unit economics still matter, regulation is still slow, atoms are still stubborn. But the set of people who can credibly try is now orders of magnitude larger than it was, and the time to first physical output is shorter. That shifts where the opportunity lives.

If you are a founder sitting on an idea for a pure software play over a known workflow, the honest question is whether you are building something that survives the next two generations of frontier models landing inside the general-purpose assistant. For most of these ideas, the answer is no.

The inversion is that the hard stuff, the stuff that needed a factory or a fleet or a decade of regulatory relationships, is more buildable than it has ever been. Not easy. More buildable. That is a different thing, and it is where I would point anyone starting a company today.

Build hard things. Software alone will not save you.