Anguilla is a British Overseas Territory in the Caribbean with a population of roughly 16,000. In 1995 it was assigned the two-letter country code ai, and for the following twenty-five years the resulting top-level domain was an obscure curiosity — a few tens of thousands of registrations, mostly people who liked the letters. Then a chatbot became the fastest-adopted consumer product in history and every company in the world needed to signal that it was doing something with machine learning.
The result is one of the cleanest natural experiments in namespace economics available. A fixed supply of short strings, a demand shock arriving over about eighteen months, an unusual registry structure with a two-year minimum term, and a buyer population with venture funding and a strong preference for a specific suffix. Whatever you think about domain investing as an activity, the price behaviour here is worth understanding, because the same dynamics show up in every constrained-namespace market.
The registry structure is unusual and it matters
Three features distinguish .ai from a typical ccTLD and each has a direct effect on price.
Two-year minimum registration. You cannot register a .ai domain for one year. The minimum term is two, which roughly doubles the entry cost relative to a comparable gTLD and, more importantly, doubles the carrying cost of a speculative portfolio. At around $70 to $100 per two-year term at retail depending on registrar, holding a thousand speculative names costs $35,000 to $50,000 a year. That is a meaningful filter on low-conviction speculation, and it partly explains why .ai has less obvious junk inventory than gTLDs where names can be held for under $10 a year.
No registration restrictions. Unlike ccTLDs with local-presence requirements, anyone anywhere can register. That removed the friction that keeps most country codes local, and it is why .ai could function as a global industry namespace rather than a Caribbean one.
Government-owned, revenue-material. The registry is operated on behalf of the Government of Anguilla, and registration revenue went from a footnote to a headline line in the territory's public finances — public reporting has put it in the tens of millions of dollars annually, a substantial share of total government revenue. In 2024 the government entered a partnership with Identity Digital to modernise registry operations. The relevant point for anyone valuing a name: this registry has become fiscally important to a sovereign entity, which is a reasonable argument for continuity, and also a structural reason to expect pricing and policy to be actively managed rather than left alone.
What the demand curve actually did
Registration counts moved from the low hundreds of thousands before the generative AI wave to well over half a million within a couple of years — the kind of multiple that in any other namespace would be described as a bubble and here was substantially driven by end users rather than speculators.
That distinction is the important one and it is where most commentary goes wrong. A speculative bubble in a namespace is characterised by names changing hands between investors with no development. What happened in .ai is different in a measurable way: the proportion of registered names resolving to an actual website with actual content is high relative to comparable gTLD land rushes. Companies were not buying .ai names to flip them. They were buying them because .com was taken, because the suffix communicated category membership at zero explanatory cost, and because investors and customers had started reading it as a signal.
That is real utility demand, and it produces a different price floor than speculation does. Speculative demand evaporates when sentiment turns. Utility demand persists as long as the category exists.
Valuing a specific name
Public comparable sales data in this market is thin, because most significant transactions are private and the ones that are reported are reported selectively. Treat any single quoted figure with suspicion. What is more reliable is the structure of what drives price, which is consistent across namespaces:
| Factor | Effect | Notes |
|---|---|---|
| Length | Very strong, non-linear | Single letters and two-character names are a separate market; 3–6 characters command large premiums over 10+ |
| Dictionary word | Strong | A real word beats an invented one at equal length, most of the time |
| Category fit | Strong | A term that describes what buyers in the category actually do |
| Pronounceability | Moderate | Survives being said aloud on a podcast without spelling |
| Existing use / traffic | Moderate | Type-in traffic and inbound links have measurable value |
| Trademark conflict | Can be fatal | A name colliding with a registered mark has a buyer pool of one, or zero |
| Hyphens, numerals | Strongly negative | Roughly halves comparable value |
The rough tiers observable in the market, as broad ranges rather than quotes:
- Registration-fee names. Long, multi-word, or awkward. The overwhelming majority of the namespace. Worth the renewal and no more.
- Low four figures. Reasonable two-word combinations with clear meaning. A liquid market with many buyers and many alternatives.
- Five figures. Short, clean, single-concept names with obvious commercial application. This is where most funded startups transact.
- Six figures. Short dictionary words, strong category terms, memorable single concepts. Thin market, long time to sale, buyer is usually a funded company with a specific need.
- Seven figures and above. One-, two-, and three-character names and top-tier generic terms. Effectively a private negotiation market with no reliable public comparables.
The valuation mistake that recurs
Sellers anchor on the top of the range for the tier above theirs and then wonder why nothing sells. The correction is to think in terms of buyer population rather than intrinsic quality.
Ask: how many organisations exist for which this specific name is meaningfully better than the alternatives available to them at a tenth of the price? For most names the honest answer is a single-digit number. A name with three plausible buyers is not worth what a name with three hundred is worth, regardless of how good it sounds, because the probability of finding a motivated one in any given year is low and the carrying cost accrues regardless.
The corollary is that time-to-liquidity should be priced in. A name that will sell within six months at $20,000 and one that will sell within four years at $60,000 are close to equivalent once you account for renewal costs and the opportunity cost of capital — and the second carries far more risk that the category cools.
What could break the market
Three risks, in descending order of how much they should worry a holder.
Category dilution. Not that AI stops mattering — that it stops being a differentiator. When every company does something with machine learning, saying so in your domain communicates nothing, the same way .io stopped signalling anything about developer tooling once every company used it. This is the most likely scenario and it is gradual rather than sudden. The names that survive it are the ones whose value comes from the second-level string being good, not from the suffix being fashionable.
Registry policy change. A sovereign registry with material revenue dependence has both incentive and ability to change pricing. Premium tiers, renewal increases, or reserved-name policies applied to categories of short strings would all directly affect holder economics. This is not speculative concern — registries do this.
Namespace competition. There is no shortage of alternative suffixes, and the constraint on any of them displacing .ai is convention rather than availability. Conventions can shift quickly in this industry.
The risk that gets cited most and matters least is the ccTLD-delegation question — the theoretical possibility of an administrative change to Anguilla's status affecting the code's assignment. It is worth knowing about, the historical precedents (notably .su and the ongoing .io discussion around the Chagos Archipelago) show these processes take many years and are handled with substantial notice, and the revenue dependence makes disruption strongly against every party's interest.
Why this belongs in an infrastructure publication
Because it is the same analysis. A premium domain is an asset with a fixed acquisition cost, a recurring carrying cost, a utility value that depends on how it is used, and a resale value that depends on a market you do not control. That is structurally identical to a three-year Reserved Instance, a GPU cluster purchased ahead of demand, or any other capacity commitment — you are pricing an option under uncertainty about your own future requirements and about a market price you cannot forecast well.
The disciplines transfer directly. Price the carrying cost explicitly rather than treating renewal as noise. Distinguish utility value (what it is worth to you in use) from market value (what someone else would pay), and do not let the second justify a purchase the first cannot. Compare against the cheapest alternative that meets the actual requirement, not against the best available option. And decide the exit condition before you buy, because the failure mode in both markets is identical: an asset held indefinitely because selling would crystallise a loss, accruing carrying cost the whole time.
That is the same reasoning applied to infrastructure commitments and unit economics, and the same reason buying accelerator capacity ahead of demand is a harder decision than it looks. Constrained supply, a fashionable category, and a two-year commitment window produce recognisable behaviour whether the asset is a string or a rack.