Hard to Make Is Dying. Hard to Get Is Not.

Every major technological shift creates a bubble.

The internet did. Mobile did. Crypto did. And AI is doing the same.

Inside a bubble, it becomes difficult to separate speculation from structural change. Companies get funded because they have “AI” in their pitch deck. Products appear that are little more than wrappers around the same underlying models. Valuations rise faster than revenues.

Some of these companies will disappear.

AI won't.

AI is not interesting because every company suddenly needs a chatbot. It is interesting because it changes one of the fundamental assumptions on which technology companies have been built:

Being difficult to build is becoming less valuable.

Hard to make

Historically, technical complexity itself could create defensibility.

If building a product required 30 engineers, millions in funding and years of development, relatively few people could compete with you.

That barrier is collapsing.

A teenager with an AI coding agent can already build things today that would have required an entire development team only a few years ago.

And models will get better.

Software development, design, research, analysis and content creation are becoming increasingly accessible.

This doesn't mean technology becomes irrelevant. Quite the opposite.

It means technology becomes abundant.

And abundance rarely makes for a good moat.

If your competitive advantage is simply that your software is difficult to build, you have to ask yourself:

For how much longer?

Hard to get

There is another kind of defensibility that AI does not eliminate nearly as easily.

Things that are not hard to make, but hard to get.

Licenses.

Distribution.

Relationships.

Physical access.

Regulatory approvals.

And, perhaps most importantly:

proprietary data.

AI models are incredibly capable generalists. They have absorbed enormous amounts of publicly available human knowledge.

But they cannot know what they cannot access.

Imagine a bakery.

The baker might know that every Friday afternoon, a particular pastry sells out. He knows what happens when it rains, when the nearby school closes for holidays, or when a local event takes place.

None of this knowledge is particularly sophisticated.

But it is valuable because it is local, current and private.

A general AI model may know almost everything about how a bakery works.

It still doesn't know what happened at that specific bakery at 4:17 PM yesterday.

That distinction matters.

Data is the kerosene of AI

If intelligence becomes abundant, unique context becomes scarce.

And scarce things become valuable.

This changes how I think about AI businesses.

The interesting question is no longer:

What can I build with AI?

Almost everyone will be able to build increasingly sophisticated things with AI.

A better question is:

What can I know that everyone else — including the underlying AI models — cannot?

This is where proprietary data becomes interesting.

But simply owning a dataset is not enough.

The strongest businesses will create their own data as a consequence of being used.

Imagine software for thousands of repair technicians.

A technician encounters a broken machine. The AI predicts the cause. The technician performs the repair. The software then learns whether the prediction was correct.

Suddenly the company doesn't just have manuals or maintenance records.

It has something much more valuable:

Problem → Prediction → Action → Outcome.

Repeat this millions of times and something interesting happens.

The product gets better because more people use it.

And because the product gets better, more people want to use it.

Usage creates proprietary data. Proprietary data improves intelligence. Better intelligence creates more usage.

That is a real flywheel.

The real opportunity

This makes me particularly interested in industries that still operate with large amounts of hidden information.

Construction sites.

Factories.

Repair shops.

Agriculture.

Procurement.

Logistics.

Local businesses.

There is an enormous amount of knowledge in these environments that has never made it onto the internet.

A mechanic knows that a particular sound usually means a particular component is about to fail.

A procurement manager knows the price a supplier will actually accept after three rounds of negotiation.

A construction manager recognizes the combination of small problems that usually results in a three-week delay.

A farmer knows that a certain part of a field behaves differently after several days of rain.

Much of this knowledge exists in people's heads, WhatsApp conversations, spreadsheets, machines and everyday decisions.

AI does not automatically have access to it.

And that may be precisely where some of the biggest opportunities are.

The goal could be to build products that sit directly inside these workflows and, through providing immediate value, continuously capture information that previously disappeared.

Over time, the software itself may become the least interesting part of the company.

The accumulated knowledge becomes the asset.

From building to accessing

For a long time, startups asked:

Can we build this?

AI increasingly makes the answer:

Yes.

So perhaps the more important questions of the next decade will be:

Can we access this?

Can we learn something others cannot?

And does every interaction make that advantage stronger?

AI makes intelligence cheaper.

It makes building cheaper.

It makes copying easier.

But it doesn't make everything abundant.

The companies I find most interesting are therefore not those that simply use the best AI.

They are the ones that use AI to create access to something others don't have.

Because in a world where almost anything can be built,

hard to get might matter much more than hard to make.