AI Labs: AGI or Bust
Why AI is the most influential technology of our lifetimes and yet its creators might not see any upside.
The true meaning of life is to plant trees, under whose shade you do not expect to sit.
Let me start by saying that I am neither a doomer nor a Luddite. I believe it is undeniable that AI will transform almost all aspects of our lives; some of those transformations will be positive, others negative. The belief that this technology will just go away again, which is held by some, is not only naive but also misguided. Artificial intelligence will deliver prosperity and advancement in ways we cannot yet imagine. Every technological revolution so far has always delivered more than people could anticipate at the time, and living standards have always improved. One should not be too averse to short-term pain in exchange for long-term gain. But conviction about the technology is not the same as conviction about the companies at the forefront of the technology. The past is littered with examples of companies that pioneered transformative technologies that redefined not just their niche but the whole trajectory of humanity. This raises a more interesting question: not whether AI will reshape the world (or destroy it), but whether the economics at the frontier actually work for the labs.
My thesis is that frontier labs face an "AGI or Bust" economic problem. In recent years, the cost of frontier training runs has risen at an extraordinary rate, with estimates from Epoch AI suggesting that the amount of compute used in training has grown 4-5x per year, while the dollar cost of frontier training runs has risen by roughly 3.5x per year. This kind of growth can only be justified if each successive generation creates enough additional value to support ever-larger investments. If this scaling continues all the way to systems capable of recursive self-improvement or some other form of persistent, self-reinforcing capability advantage, then the enormous expenditures of frontier labs may ultimately be rational. A world in which the frontier labs reach recursive self-improvement will be one in which they hold a widening lead over their competitors in capability, and open-source models or smaller distillations could never catch up. This is the AGI scenario in the "AGI or Bust" dilemma. The other side of the dilemma is the one in which capabilities plateau somewhat before such a breakthrough occurs. This is a scenario in which training costs eventually stop producing commensurate gains in capability or economic value, forcing the frontier labs to slow or halt the relentless increase in R&D spending. It also narrows the gaps between the frontier, the runners-up, open-source and distilled models over time. The frontier model may remain the best model, but it may no longer be sufficiently better to justify large and persistent pricing premiums, at which point model capabilities begin to resemble a commodity rather than a defensible monopoly.
This dynamic differs from traditional software businesses in a fundamental way. A company might spend tens of millions of dollars annually on software platforms such as CRM systems, collaboration tools or databases. Even if cheaper or open-source alternatives exist, the savings-to-risk trade-off does not look advantageous. AI spending in the future, however, could easily be 2 to 3 orders of magnitude larger for the largest companies, reaching billions if not tens of billions per year, which creates an overwhelming incentive to reduce cost however possible. As open-source models approach frontier performance, large enterprises will increasingly choose to self-host or deploy models on-premises through clouds or their own clusters rather than pay premium prices to frontier labs. A 10-20% reduction in costs may be insignificant if measured against a traditional SaaS contract, but when one measures it against multi-billion-dollar AI spend, it becomes economically important. The infrastructure required to serve models at scale is already becoming commoditised through hyperscalers and specialised hardware providers. Consequently, once performance differences narrow sufficiently, cost minimisation rather than model quality becomes the dominant purchasing criterion. The result is that artificial intelligence may become one of the largest industries in history while frontier model providers capture only a fraction of the long-term value they created. In a world without a persistent capability moat, the economic rents are likely to flow both up and down: down to the chip manufacturers, data centre operators, cloud providers, etc., and up to the companies using AI to improve their productivity, ship more features and unlock more value for their customers. This points to an interesting conclusion: the business model of the AI labs is only viable if scaling ultimately produces a durable and widening capability advantage that does not diminish over time. Without such an outcome; whatever you call it: AGI, superintelligence, recursive self-improvement or something else, the frontier layer risks being competed away, leaving the labs with the industry's largest costs but no lasting claim on its profits.
The key implication is that AI can be enormously valuable while AI labs still fail to capture the value they created. In a non-AGI plateau scenario, the labs function less like durable software monopolies and more like entities financing the discovery process for the rest of the economy (ironically, this is what a lab is). The long-term rents would likely accrue at the infrastructure layer and at the product layer where AI helps reduce costs, increase productivity and create value for customers.
AI will change the world.
Betting on the labs is a bet on AGI.