A frontier lab's lead earns a rent that open models chase within months; it lasts only where the lab is built into its customers' work
The frontier labs, OpenAI, Anthropic, Google DeepMind, xAI and Meta, train the most capable AI models: programs trained on huge amounts of text and data to write, reason and carry out tasks. They sell their use by the token (a fragment of a word) and through products such as chat apps and coding tools. David Ricardo, an English economist who asked in 1817 why landlords grew rich when bread grew dear, found that the owner of the scarcest input collects a rent: profit above what it takes to keep a business going, earned because something holds rivals off. Today the scarce input is the best model, and the lab that has it charges for it. This site judges that the rent on a lead is transitional, because open-weight models (whose trained settings, the weights, are published for anyone to run) arrive some months behind, and that it becomes durable only where a lab is woven into how its customers work or owns the place where users meet it.
- Getting cheap and interchangeable
- This site judges that answering a well-understood request, the kind a model from a year before could already handle, is a commodity: many labs and open models do it, and it is priced close to the cost of the chips and power that run it. William Nordhaus, an economist who measured how much of the value of new inventions their makers keep as profit, found that it is a small part; most goes to buyers as lower prices and better products. Kenneth Arrow, an economist who studied how markets handle knowledge, argued that information is hard to own because copying it is cheap; a model's behaviour can be observed and partly copied by a rival training on its answers.
- Staying scarce
- This site judges that the newest model at the frontier stays scarce for as long as it leads, and that training one costs so much in chips and money that only a handful of firms attempt it. Hamilton Helmer, a strategy consultant who studied why some firms keep earning more than their investors could expect elsewhere, found that a lasting advantage is a benefit behind a barrier; the one that clearly applies today is scale economies (the cost of training a model is fixed, so a lab with more users spreads it thinner, and a rival must match the users to match the cost). Switching costs (what a customer would lose by moving to a rival) are emerging, not proven: they exist where a lab's tools are built into a firm's systems, but this site cannot yet measure how often firms use several labs at once, so it does not count them in this group's powers.
- What would make it last
- Labs turn deployments into switching costs: their coding agents, business tools and staff inside customers become part of how firms work, so moving to a rival means rebuilding, not just repointing. Or a lab owns the place users meet AI, as a chat app with a habit behind it, which Helmer would call branding (a buyer pays more for the same thing because of its name) or network economies (the product grows more useful as more people use it). David Teece, an economist who studied why inventors often fail to keep the returns from what they invent, found that the firm holding the complementary assets (the distribution, sales and systems a product needs to reach buyers) often keeps the profit; Google and Meta already hold theirs, and OpenAI and Anthropic are buying theirs. The rent also lasts, without either, if open-ended uses where only the best model will do keep growing faster than bounded ones, the case Giovanni Cattani makes above. Rules could entrench it too: Sayash Kapoor and Arvind Narayanan, computer scientists at Princeton who study what AI can and cannot do, argued that licensing who may build models, combined with limits on sharing them, would leave only a handful of firms able to build the most capable ones; this site reads that as a route to a regulatory rent (a profit protected by a licence rivals cannot get), which would make the lead last.
- What would prove this wrong
- Labs keeping their price premium and their share of business spending on models that open ones already match, with no growth in deployments or products; or, the other way, business spending spreading across many labs and open models as freely as developers' routed tokens already do.
- Powers here today
- scale economies
- Binds on (how tight)
- Building frontier models (easing), Post-training (no reading; a model judges it probably moderate), Running models (easing), Deployment platforms (no reading; a model judges it probably easing), Capital (slack)
- The rival view
- The labs are durable businesses with high margins