The four moats, and what happened to each
The classic answers to "what stops someone copying you" were scale economics, network effects, switching costs, and brand. Every one of them is under pressure in AI, and three are close to gone at the stage most companies are competing at.
Scale economics. You rent your compute and you rent your model.
So does the company shipping your feature next month, at the same list price. Cost advantage arrives at a scale almost nobody reading this has reached.
Switching costs. An API call is a line in a config file. Where the product is a wrapper around a model, moving is an afternoon.
Real switching costs come from data gravity and workflow depth, and both take years you have not had yet.
Network effects. Genuine ones are rare and mostly claimed rather than held.
A tool that gets better as more people use it is a network effect. A tool that has more users is a customer list.
Brand. This one survives, and it is the only one of the four that a company can build deliberately inside a year. What people call brand at this stage is closer to recognition: whether the name arrives in the room before you do.
What replaced them
Strip out the three that decayed and what remains is recognition: whether the specific people who decide already know your name when it comes up.
That sounds soft until you price it. Enterprise buyers arrive at a category with two names already in mind, and the third company to be discovered is competing for a slot that is usually gone before the evaluation starts.
Investors take the meeting differently when they have already seen the company in their own feed. Engineers answer the message faster. Each of those is a compounding advantage that has nothing to do with the product.
The uncomfortable version: two companies ship the same capability in the same month, and the one the market has heard of wins. Higgsfield had no model advantage over Pika, Luma or Runway and went from launch to a five billion dollar valuation in about fifteen months. The difference was distribution.
The moats still standing
Four classic moats thinned. Six kinds of advantage still hold for AI companies, and the useful difference between them is how long each takes to build.
| Moat | What it is | Can a young company build it inside a year? |
|---|---|---|
| Data | A dataset or body of content a competitor cannot recreate by paying for it | Sometimes, when the product generates it as people use it |
| Community | Users who make the product more valuable for each other | Rarely, because it needs a crowd first |
| Distribution | Putting something in front of the people who decide, when you choose to | Yes, and it is the fastest of the six |
| Physical operations | Hardware, sites, supply chains and people on the ground | Seldom; it takes capital and years |
| Capital | Owning assets, including stakes in the companies taking a growing share of the economy | Only if you start with it |
| Regulation | Licences, approvals and legal firsts | Sometimes, and it lasts once won |
Distribution is the one that makes the others count, and the kind that compounds is trusted distribution. A regulatory first that nobody hears about protects nothing. Nolla Health became the first organisation in the US to have an AI issue a prescription; the approval is the moat, and the launch, 2.5M views from a founder account of 2,300 followers, is how the market learned it existed.
The same holds for the rest. A dataset defends you once buyers know you have it. A community needs a crowd before it has any effect, which is a distribution problem first. That is the order we would build in: distribution now, because it is the one you can have this quarter, and every other moat on the list is worth more once people know your name.
The objection, which is a good one
A moat is supposed to hold while you sleep. Attention does the opposite. Stop posting for two quarters and the market forgets, which is the argument against calling it a moat at all.
We think the objection is right about the stock and wrong about the capability. The attention itself is not the defensible thing, because it drains. What defends you is being able to refill it on demand, in a week, without buying it.
Most companies cannot. They have a product and a list of customers and no way to make ten thousand relevant people look at the same thing on a Tuesday.
That capability is rare, it takes time to build, and it does not transfer when a competitor raises more money. Which is the actual test of a moat.
What the data says about who gets attention
We track every AI launch on X we can find, with the engagement attached. Across the directory:
- The median launch did 688,938 views.
- The median account posting one had about 17,000 followers.
- That is a median of 26 views for every follower the account had.
- Of the launches posted by accounts under 10,000 followers, one in four passed a million views.
Read that again in terms of moats. If the typical launch reaches 26 times the following of the account that posted it, then audience size is not the defensible asset.
The following is a starting condition, and a weak one. What moved those numbers was the post, the timing and how many other accounts carried it inside the same window.
This is the good news and the bad news in one line. Attention cannot be hoarded, so a competitor's head start is smaller than it looks. It also means yours is.
Where it does compound
Attention behaves like preferential attachment. Each round of it makes the next one cheaper, because a name people recognise gets a longer first read, and a longer first read is most of what ranking systems reward.
So the asset is not the audience. It is the position: being one of the two names in a category, which raises the return on everything else you do.
That position is defensible in the only way that matters, because taking it from you requires a competitor to out-execute you repeatedly rather than outspend you once.
Why it cannot be bought
The obvious objection is that a budget solves this. It does not, for a structural reason we go through in the case against paid: the buyers who decide are disproportionately the ones paying to remove ads. Paid reach lands on everyone except them.
The second objection is that a company can publish its way there. Owned channels run into a discount that does not respond to budget, because every reader silently asks who benefits if they believe this, and on your own channel the answer is always known.
How to tell whether you have one
Four questions, answerable this week.
Can you name the two companies your buyer already thinks of? If you are not one of them, you are working from behind on every deal regardless of the product.
What happens if you announce something on a Tuesday? If the honest answer is a few hundred impressions and a quiet Slack channel, you have no distribution, whatever the follower count says.
Where does inbound come from? Companies with this capability get told "I saw your launch" unprompted. Companies without it get told nothing and assume the product is the problem.
Could you do it again next quarter? One good launch is luck. The moat is the second one.
Questions we get asked
Is attention really a moat if it decays?
The attention itself is not, because it drains within a quarter. The ability to rebuild it on demand is, because that takes time to develop and does not transfer to a competitor who raises more money. Treat the stock as perishable and the capability as the asset.
Do we need a large following before any of this works?
No, and the data argues against waiting for one. The median launch in our directory came from an account with about 17,000 followers and reached 688,938 people.
Of the launches from accounts under 10,000 followers, one in four passed a million views. Reach on a given post is decided by who carries it, rather than by how many people already follow you.
What about a genuine network effect? Is that not a stronger moat?
Much stronger, when it is real. Most claimed network effects are customer lists. If yours is genuine, distribution is how you reach the density where it starts working, which makes the two complementary rather than alternatives.
How long does a launch keep paying?
The reach is spent in about 48 hours. What persists is the recognition, which shows up months later as branded search, warmer first meetings and inbound that mentions something you posted. That gap between the traffic ending and the effect landing is why launches get judged too early and cut too soon.