In the era of AI, what is the truly reliable moat?

God Translation Bureau is a compilation team under 36Kr, focusing on fields such as technology, business, workplace, and life, with a particular emphasis on introducing new technologies, new ideas, and new trends from abroad.

Editor's note: When AI drives the cost of intelligence to zero, the hard assets built on atoms, licenses, and capital are entering the strongest compound - interest cycle in history. The text is from a compilation.

In a world where AI can develop any software, replicate any product, and automate any process, what exactly makes a company defensible? Most of the answers I've heard are wrong. These answers are all based on the same premise: that intelligence won't become a hundred times faster, a hundred times stronger, and a hundred times cheaper than it is now. But the fact is, intelligence will soon make such a leap.

Defensibility stems from two things: things that are difficult to do and things that are hard to obtain. AI is making the former worthless. Developing software, maintaining integrations, and embedding a product so deeply into a customer's system that it takes a year to replace? These are indeed "difficult", but they're becoming increasingly easy. However, having tens of millions of users, government licenses, a chip foundry, or a billion dollars in deployable capital? These are "hard to obtain" things. AI compresses the time required to "do things", but it can't compress the time required for "things to happen". This difference is now the most important screening criterion I use when investing. The following five points meet this criterion.

Privately - accumulated data over time. Not all data qualifies. A static dataset that exists only because it's expensive to collect will eventually be replaced or bypassed by synthetic data. That's "difficult to do", not "hard to obtain". The real moat is "living data": private information that's continuously generated through a defensible business operation. Orchard AI installs cameras on agricultural machinery, tracking billions of fruits on millions of trees across multiple growing regions and seasons. The data is updated daily, and with each pass through the orchard, the dataset becomes richer and the model becomes smarter. You can't replicate this by training a model on public data. You have to drive the same cameras through the same orchards for years.

Network effects. The addition of each new user makes the product more valuable to all other users. DoorDash is the clearest example: each additional driver speeds up delivery, each additional restaurant gives customers more choices, and each additional customer improves the economic efficiency for everyone. You can clone the app overnight, but you can't clone the drivers, restaurants, customers, and delivery density in ten thousand cities. Mutual funds like Cache show how this works outside of trading platforms: the larger the pool of funds, the higher the diversification, which attracts more participants. As AI makes it incredibly easy to create competitors, the cold - start problem may actually become more difficult, because now there are a hundred well - crafted alternatives vying for the start of the same network. Whoever has liquidity first can continuously accumulate advantages, while others can only scramble for the leftovers.

Regulatory licenses. The speed of government operations depends on politics, not technology. Anduril needs procurement licenses and classified contracts to sell products to the US Department of War, and no matter how powerful the AI is, it can't compress this time. Bank licenses take years, and FDA approvals take years. Moreover, the scope of regulation is expanding rather than shrinking, because as AI capabilities increase, so do the risks. Will the specific forms of regulation change? Almost certainly. But will the need for human approval disappear? I don't see that possibility.

Large - scale capital. This is something that almost everyone underestimates. The end - game of competition is in the physical world. A chip foundry costs $20 billion, a nuclear power plant costs $10 billion, and a satellite constellation costs billions more. There's a reason why Elon Musk says money may not matter in 15 years, yet he's raising $75 billion and taking SpaceX public. When the bottleneck shifts from software to atoms (the physical world), the ability to raise and deploy large - scale capital becomes one of the core advantages of this era. And the ability to access capital isn't just about money; it's also about institutional trust, past achievements, and personal connections that take decades to build.

Physical infrastructure. Factories, power plants, battery networks, data centers. Base Power has deployed thousands of battery units in homes in Texas and is also building its own manufacturing plant. Each installed unit is a physical asset that generates income in someone's yard. Soon, you'll be able to design this system with AI in a week. But you can't manufacture, install, and interconnect thousands of units in a week. Physics sets a lower limit on the timeline that intelligence can't break through. Whoever starts building first will see their lead grow with each month that competitors haven't started.

I've recently talked about "end - game positions", which are assets that strengthen as AI becomes more powerful. Each of them has this characteristic : they require years of real - world time to accumulate, and no amount of intelligence can compress this process. Network density takes years to spread among the population, regulatory approvals take years of political processes, infrastructure takes years to build, data takes years to accumulate, and capital relationships take decades to win. This "non - parallelizable" time is the common "meta - moat" under the above five points. Companies that have already occupied these positions are not only defensible; they're widening the gap every day, because the first - mover advantage itself is a moat.

For those items not on this list, the time can be compressed. Workflow embedding sounds solid until you realize that the switching cost is just engineering time in disguise. Ecosystem lock - in sounds like a fortress until AI can rebuild integrations as fast as you can describe them. Software scaling - spreading engineering costs across millions of users - becomes insignificant when engineering costs approach zero. These were moats built against "intelligence scarcity", and this scarcity is something we know is coming to an end.

There are still some things I'm not sure about. Will "trust" become an independent moat when AI takes on more work? Someone has to be responsible when things go wrong, and the institutions that take on this responsibility may become more valuable, not less. Is human attention a moat? When the cost of creating content drops to zero, brands that have already captured attention may have the most powerful compound - interest advantage. I think both of these are real, but I haven't figured out if they're independent categories or by - products of the five points I mentioned.

But the core screening criterion still stands: is it "difficult to do" or "hard to obtain"? If your moat's bottleneck lies in intelligence, you're just buying time. If its bottleneck lies in long years, human behavior, physical laws, political will, or capital, you may be building something that can stand the test of time.

Translator: boxi.