by Dwight Furrow
A moat is what protects a business from competition. The term comes from Warren Buffett’s image of a castle surrounded by water. The castle is the business; the moat is whatever prevents rivals from storming the walls. A moat might be a famous brand, a patent, a network effect, control over scarce resources, high switching costs for consumers, or a regulatory barrier that makes it difficult for competitors to enter the market. The deeper the moat, the easier it is for a firm to charge high prices, preserve margins, and survive imitation.
That is a problem for AI companies. They are spending staggering sums on chips, data centers, engineers, and research, yet it is not obvious that they have a durable moat. Among the top three frontier labs, the latest system from one company is only slightly better than the latest system from another and who is on top changes depending on who has released their latest model. And free-to-use, open-access models from China are only a few months behind the frontier labs in their development. If these trends persist, the AI business begins to look precarious. Enormous cost, rapid imitation, and uncertain customer loyalty does not make investors joyful. The castle is there but the water around it may be ankle-deep.
What is the solution for the AI labs? Anton Leicht’s Substack essay “Cut Off” suggests a solution, although his concern is not corporate profitability.
His argument is that access to the most advanced AI systems will become more limited, not less. The familiar assumption is that market pressure will make advanced AI abundant. Better models will proliferate; prices will fall; the future will belong to those who use AI most skillfully. Leicht thinks recent events point in another direction. “Access to frontier AI,” he writes, “will soon be limited by economic and security constraints.” That sentence should make investors, regulators, start-ups, and foreign governments pay attention.
His example is Anthropic’s decision to disseminate Claude Mythos only to a select group of companies and U.S. government security agencies. This is how the AI moat may be shifting from the technical superiority of models to controlled access. On the surface, the justification is security. Mythos can apparently identify serious cyber vulnerabilities that could threaten the software that runs the Internet if not patched—a disaster if put in the wrong hands. Because Mythos can identify these vulnerabilities, it was given first to trusted defenders allowing them to patch weaknesses before attackers could exploit them. But the same restriction also has a business consequence. By limiting who can use the model, Anthropic reduces the risk that competitors, foreign firms, or sophisticated users will use broad access to imitate and distill its capabilities. In that sense, selective release does double duty. It can be defended as responsible deployment, but it also helps protect Anthropic’s investment by making frontier capability scarce, vetted, and expensive. The model’s value goes beyond what it can do and includes who is permitted to use it.
According to Leicht, a convergence of pressures suggests this might be the future of AI. Frontier AI firms face security risks if their systems can be used for cyberattacks, biological weapons, or other dangerous purposes. They must confront theft risks if powerful models are hosted in insecure environments. Their business model is under threat if the practice of using more powerful models to train or improve open access models (a process called distillation) allows sophisticated users to reduce a billion-dollar breakthrough into only a temporary advantage. And they are limited by compute constraints because serving users is not like distributing ordinary software. Each query consumes scarce computational capacity.
This is where the moat begins to appear. Not a natural moat, exactly. A natural moat would arise from the product itself: the model is uniquely good, users cannot leave, rivals cannot imitate it, and the company’s advantage compounds over time. As noted above, that is not obviously the case with frontier AI. The leading models often seem to converge in their capacities. If that trend continues along with rapid development of free-to-use open access models, customers will go where price and convenience lead them.
If a company must recover its research costs in the short interval before another firm distills its model, the business is fragile. That fragility gives AI companies a strong reason to favor restrictions. They do not need to say, “Please regulate us so we can preserve our margins.” No one would put that in a press release. But they can say something more respectable: dangerous systems require careful deployment; customers must be vetted; access must be monitored; models must not be handed to criminals, adversaries, or careless operators. Much of that may be true. The trick is that true public reasons can also serve private economic interests.
Thus, regulations create the moat that technology alone cannot. If frontier AI becomes subject to strict access controls, then the competitive landscape changes. It is no longer a matter of who can build the best model. It is about who can satisfy security requirements, negotiate with government agencies, provide trusted infrastructure, and maintain relationships with the state. This is not the old Silicon Valley fantasy of frictionless innovation. It is closer to defense contracting, finance, telecom, or pharmaceuticals. These are industries in which permission matters as much as technical capacity. The winning firms are the ones who can function inside a dense web of compliance, procurement procedures, lobbying, security review, and government dependence. They know the language and can hire the right lawyers. They understand which doors open only from the inside.
Leicht’s account of compute scarcity strengthens this possibility. Frontier AI is not like traditional software, where the marginal cost of serving one more user is close to zero. Leicht notes that “providing access to AI models, especially those at the bleeding edge, takes massive amounts of computational resources.” That makes exclusion easier to justify. If access is scarce, someone must decide who gets it. That means politics enters the picture and, under those circumstances, incumbents usually benefit.
The result could be an access hierarchy. Government agencies and favored corporations get frontier models first. Less trusted firms get limited access. Start-ups, foreign companies, researchers, and ordinary users get product interfaces built by someone else. Everyone may eventually get the model, but by then the frontier has moved on. The public receives yesterday’s miracle while insiders work with tomorrow’s technology.
For frontier AI firms, this arrangement could simultaneously solve several business problems. It limits the ability of open source models to distill the frontier models. It restricts access to customers who can pay high prices and reduces the burden of serving less profitable users. It turns scarce compute into a premium resource rather than a commodity service. Most importantly, it changes the story investors can tell themselves. The company is no longer just one model provider among many, vulnerable to open-source competition and rapid imitation. It is part of the national infrastructure. This is not a bad place to be when going to the public stock markets with an IPO (Initial Public Offering), as both OpenAI and Anthropic have declared will happen soon.
But the costs of this competitive strategy are serious. Restricted access slows diffusion of AI. The broad experimental culture that makes new technologies useful depends on many people trying strange things, failing, adapting tools to solve local problems, and discovering uses no central authority could have predicted. If only a few trusted actors get meaningful access to frontier AI, then social learning is stifled. Innovation becomes less chaotic, less surprising, but also less democratic.
We could probably benefit from less chaos and fewer surprises. This new world is approaching too rapidly for most people, if current polling is to be believed. But the anti-democratic implications are disturbing to say the least. Leicht imagines a future in which unlimited access becomes “the exception, not the norm.” The public interface remains but access to intelligence is rationed.
The geopolitical consequences may be even more significant. If frontier AI becomes economically and strategically central, then unequal access divides countries into insiders and outsiders. The insiders get better tools for defense, administration, science, logistics, and commerce. The outsiders get delayed, filtered, or second-best systems. This becomes a world-order problem. AI would become another way in which technological advantages harden into political hierarchy.
The difficult point is that this outcome may not require a conspiracy. It can emerge from reasonable decisions made under pressure. Firms worry about the technology being misused, and governments worry about adversaries getting ahold of it. Security agencies worry about vulnerabilities. Companies worry about recovering costs, and data centers cannot serve everyone. Each decision can be defended on its own terms, but together they build a wall. AI safety advocates get slower and more controlled deployment while national security officials get leverage over a strategically important technology. Frontier labs get protection from commodification. Large corporate customers get privileged access. The losers are start-ups, weaker states, independent researchers, and the general public, all of whom may be told that access is coming later, in safer form, through approved channels.
So has AI found its moat? Perhaps.
Maybe AI will evolve in a different direction. Perhaps the better angels of Messrs. Amodei, Altman, and the Google hierarchy (the leaders of the frontier labs) will prevail to preserve whatever latent democratic forces remain buried in the business and technical architecture of these powerful LLM’s. (The thought that the current occupant of the White House might have undue influence over the development of perhaps the most powerful technology in history ought to concentrate the mind.)
But the incentives do not appear to be aligned with anything that looks like democracy. And one should never ignore incentive structures if prediction is your game.