What Code Can’t Buy

In March, OpenAI spent several weeks negotiating with TPG, Advent International, Bain Capital and Brookfield. Not to raise money. To create a structure responsible for distributing its enterprise products inside the companies owned by those funds.

Anthropic did something very similar with Blackstone, Hellman & Friedman and Goldman Sachs.

I find these two deals much more interesting than they look.

The two most advanced labs on the planet, the ones building the models much of the industry works with, the ones probably concentrating more technical talent than any company could reasonably need, have just put a tremendous amount of money on the table to obtain something they cannot build themselves.

Access.

Companies they can get inside. CFOs who pick up the phone. Established relationships. Accumulated trust. Contracts, data, people who know whom to speak to and in what order.

OpenAI can train a better model. Anthropic can write a better agent.

They cannot vibe-code twenty years of business relationships.

The market hasn’t simply accelerated. It has changed shape.

One side of the problem is becoming extraordinarily cheaper. The other has barely moved.

Ten years ago, if you were a business person with a software idea, finding a technical cofounder was step zero. Not step one. The step before that.

Without one, you had a deck, a few mockups and a conversation that usually ended with: “Now we need to find someone to build the thing.”

That search could take six months. Sometimes twelve. And quite a few companies were probably started by two people who didn’t particularly want to spend ten years together, but were simply relieved to have finally found the other half of the puzzle.

That constraint is disappearing much faster than we realize.

Today, a good business founder can build a first version, put it in front of customers, see what breaks, fix it, start again, and get surprisingly far before a CTO becomes truly indispensable.

Not as far as building OpenAI in their bedroom.

But far enough to test an intuition, get users, learn and sometimes even find the beginnings of a market before assembling the technical team they would have needed on day one in 2019.

Cursor and coding agents are obviously the most visible manifestations of this phenomenon. But I think the interesting consequence isn’t that “everyone can code.”

It’s that producing working code is gradually losing some of its scarcity.

Let me clarify something, because this discussion gets stupid very quickly if we don’t.

If the technology itself is your moat, none of this applies in the same way. If you are building new models, infrastructure, robotics, biotech or proprietary research, you still need exceptional technical founders and engineers.

Probably even more than before.

When average tools become excellent, the gap between someone good and someone exceptional doesn’t necessarily disappear. It can become even more brutal.

But across an enormous part of enterprise software, the ability simply to build the product is becoming much less scarce.

Now look at the other side.

You still have to convince a large company to let you in. You still have to find the budget. Get through security. Compliance. Legal. Convince a buyer. Then their boss. Then sometimes their boss’s boss.

And answer a 200-line security questionnaire someone probably created in 2014 and nobody has dared delete since.

Enterprise sales cycles are still measured in months. Decisions still involve an absurd number of people. And the explosion in the number of AI products probably hasn’t simplified the problem. Mostly, it has given companies a few hundred more vendors to evaluate.

This is where the MIT figure gets interesting.

In their study of enterprise AI, 95% of the pilots analyzed had produced no measurable impact on the income statement.

95%.

And I think we badly misread that figure when we simply take it as proof that the models aren’t good enough yet. A large part of the problem isn’t the model.

It’s getting inside the house.

Because you cannot prompt a hundred enterprise customers into existence. You cannot vibe-code a dataset that took fifteen years to accumulate. You cannot generate the trust of a general counsel who has already been burned twice.

And you cannot synthesize overnight the domain knowledge of someone who has spent twenty years in an industry and knows precisely why the three obvious solutions to the problem don’t work.

That is the asymmetry that interests me.

AI is removing much of the business founder’s technical dependency on day zero. It has created no equivalent shortcut on the market side. Yet I still see a huge number of founders hiring for the company they would have needed to build in 2019.

They look for someone who can make what they cannot make. Whereas the interesting question is gradually becoming: What can’t I build, even with the best AI in the world?

Distribution. Proprietary data. Trust. Domain knowledge accumulated over twenty years. Original research. An established position in an industry.

The right to enter the room.

And if that is what is becoming scarce, something else changes with it.

The new cofounder may not be a person.

It may be an established player. I don’t mean the “strategic partnership” announced in a press release and added to a slide with two logos holding hands. I mean someone you genuinely tie yourself to because they already possess what would take you ten years to build: customers, data, trust, knowledge of the business or access.

Look at Abridge.

At first glance, it is an excellent healthcare AI company. Founded by a cardiologist, it now works with hundreds of hospital systems. But what I find much more interesting is what it has built around the technology.

Abridge partnered with UpToDate, then the New England Journal of Medicine and JAMA, to integrate their medical content into its tools. NEJM has been publishing since 1812.

Two hundred and fourteen years of legitimacy.

You can probably produce a better model next year. Good luck producing 214 years of institutional trust by the next sprint.

OpenEvidence tells the same story differently. The company has managed to enter hospital environments such as Mount Sinai, Cedars-Sinai and Sutter Health and integrate with the tools doctors already use. Some of its product’s features can obviously be copied. Its position inside the system, much less easily.

Harvey tells yet another version of the same phenomenon in law. Tens of thousands of lawyers now use it, and the product has managed to penetrate some of the hardest law firms in the world to get into.

Of course the product has to be good.

But imagine two companies tomorrow with exactly the same model. One starts with an excellent product and zero relationships. The other is already inside dozens of institutions that have entrusted it with their workflows, their habits, sometimes their data, and spent months approving it. These are not the same companies.

Having got inside becomes part of the product.

And that is precisely what private equity funds are selling OpenAI and Anthropic today. They aren’t merely selling introductions. They are selling an established position. Companies they already own. Executives they already work with. Relationships in which someone else has already absorbed the initial part of the risk.

In other words, they are selling time.

Twenty years of networks. Twenty years of trust. Twenty years of open doors.

And no amount of compute can compress twenty years into twenty seconds. There is obviously a trap in this strategy, and it is big enough to deserve spelling out: tying yourself to someone who owns distribution can also be an excellent way to become their subcontractor.

If they own the customer, the relationship, the data and the margin while you simply supply a technology layer, you haven’t built a moat.

You’ve signed a services contract with a nice story wrapped around it.

The alliance only becomes truly interesting if you retain something that accumulates on your side: measuring outcomes, data generated through use, an understanding of the workflow, a layer that improves because it is running. Something that means you are harder to replace after three years of partnership than you were on day one.

Otherwise, the partner learns.

You execute.

And one day they put your line item out to tender.

But despite that risk, I think the overall calculation has reversed. For fifteen years, the business founder’s reflex was to find someone who could build what they couldn’t build.

That reflex was perfectly rational.

The problem is that it may be outliving its reason for existing.

Scarcity has moved. And when scarcity moves, continuing to organize a company around the old scarcity is probably one of the surest ways to build the wrong company.

So the question is no longer just: Who can build this with me?

It becomes: Who already owns what I won’t be able to build?

OpenAI and Anthropic have some of the best engineers on the planet. They have the models. They have the capital. They have the brand. They probably have more technical capability than almost any company in history.

And they still wrote the check.

Because they knew exactly what they were buying.

Not code.

A door already open.

I am doing something fairly similar at my own scale with Naijm in the Emirates.

But that is another story for another time.