Are AI Companies building sand castles, not moats?
It's been another week where the status quo is constantly changing
In the 2010s, social networks had an obvious and impenetrable moat: the network effect, which meant that the more friends you had on one service, the harder it was to switch to another. In the 2020s, moat logic is a lot less certain.
This has the risk of sounding like one of those AI-written koans that show up on LinkedIn (“It’s not a moat — it’s a sand castle”). But I think the analogy aptly describes some of the weird stuff going on. Media writer John Hermann has a good piece in this week’s New York Magazine about how AI companies seem to be unable to forecast their own future, and I’d argue that extends to their own business defensibility.
Moats can have a few grooves (this SEO-optimized post by Charles River Ventures digs them all out). Let’s get sand between our toes and see what moats AI really has:
Network Effects: The 2010s were defined by these moats, in which friends on a service beget more friends on a service.
These strike me as being really low for AI. There are attempts at getting network effects — the first dot biz video was about forward deployed engineers, which sends a hive of pseudo-consultants into a company to try to create some lock in. OpenAI attempted group chats at one point, but the effort seemed to fizzle. Their video social network, Sora, was shut down. For AI, the Network Effects piece hasn’t fallen in place.
Switching Costs: How hard is it to switch your AI?
This strikes me as the ultimate AI paradox. The goal is to build a series of increasingly intelligent computers that can understand anything. The more they can understand, the easier it becomes to transfer data between domains. To add to this, much of the knowledge that an AI accumulates for individual power users is stored in laughably simple personal “markdown” files that describe, in plain English, how to replicate an AI’s function. Another AI can just…read those notes.
This is especially acute when IDEs — the “integrated development environments” programmers use — can easily switch not only between AI models, but also between the companies providing those models.
It seems unfair to claim there’s no lock-in at all — I know that as a personal user of a $100 plan, there’s a certain amount of laziness that keeps me from switching, especially when I presume that the next model drop will catch me up with the competition. But if that upgrade schedule weren’t predictable within the week? I’d jump ship immediately.
Google has tried to increase switching costs by bundling their plans with storage, YouTube Premium, and a full suite of AI services. That heft may work in the long term, but it’s currently unclear.
Intangible Assets: This strikes me as the strongest moat for any company which would, in turn, worry me. It’s the squishiest. Looking at branding, each company has some intangible assets: OpenAI has the best consumer awareness and is synonymous with “AI” for many; Anthropic has safety vibes and gorgeous brand direction; Google is Google; and XAI has the edgelords on lock.
These intangible assets can also extend to relationships with the government. Sam Altman and OpenAI have close ties to the Trump Administration, while Anthropic has a feud. Jensen Huang of NVidia and Trump are close as well. The list goes on. This type of crony-capitalist relationship is an untraditional measure of intangible assets, but that’s 2026 for you.
Cost Advantage: Daily updates about inference optimizations — what actually helps the AI answer questions — and chip crunches are too frequent to catalog here. To me, that says that no company has created an enduring cost advantage. Like the AI models themselves, companies can easily catch up to competitors.
This cost advantage includes the constant business threat of open source models, which can be run locally or in the cloud for a fraction of the cost of models from the big players. Without judging the “hacking threat” new open source models may pose, open source is constantly nipping at the heels of the big companies, especially as customers become more budget-conscious.
Efficient Scale: In his interview with Dwarkesh Patel, Anthropic CEO Dario Amodei mentioned Cournot competition, which basically means markets can be stable if they’re super pricey to get into. Big training runs may have those characteristics, because they can cost hundreds of millions of dollars of computing power. But open source may show that these big expenditures aren’t necessary.
It’s another possible case of an AI paradox: the more the creation of models is optimized, the easier it gets for competitors to make new ones.
This analysis is highly subjective, and the billions of dollars flowing into AI indicate that it’s not shared by a lot of big money players. But a moat is supposed to be hard to cross, and these moats seem pretty shallow, narrow, or dry. A sandcastle? It can be beautiful, and it can be amazing for a day. But you can also copy it a few feet over on the sand. And it can be washed away.