The AI gold rush examined¶

The Patrician has observed many enthusiasms sweep through Ankh-Morpork. There was the year everyone became convinced that swamp dragons were the future of personal transportation, which ended predictably with scorched eyebrows and renewed appreciation for horses. There was the brief period when pyramid schemes were considered innovative financial instruments, which concluded with several entrepreneurs acquiring unexpected familiarity with the city’s correctional facilities. And there was the infamous tulip incident, which we do not discuss in polite company.

The current AI enthusiasm bears certain family resemblances to these episodes. Vast sums are being deployed based on projections ranging from “optimistic” through “wildly speculative” to “have you actually met reality?” The people deploying these sums are not fools, which makes the situation either brilliantly prescient or expensively mistaken, and we will not know which for some years. Some of the evidence, however, has begun arriving ahead of schedule.

What is actually happening¶

Large language models can generate remarkably coherent text, write code that frequently works, and answer questions with confidence that occasionally correlates with correctness. These are genuine capabilities. They are also considerably narrower than the hype suggests because the models work through pattern matching on enormous training datasets rather than understanding in any sense that would survive philosophical scrutiny.

Training frontier models was believed to cost hundreds of millions and require the resources of a hyperscaler. In December 2024 a Chinese laboratory called DeepSeek reported that the final training run for its V3 model had taken 2.788 million GPU hours, which the paper priced at about five and a half million dollars, a figure covering that run alone and not the research and abandoned experiments behind it. Its R1 model followed in January 2025, and Nvidia lost five hundred and ninety-three billion dollars of market capitalisation in a single day, the largest one-day loss in the history of the US market. The Patrician notes that the ratio of the number that went in to the number that came out is instructive, and that “you need to be a hyperscaler to compete” has become an assumption in need of urgent revision.

The enterprise return on investment has been measured, at least once. An MIT report published in August 2025 found that ninety-five per cent of corporate generative AI pilots delivered no measurable return, against enterprise spending it put at thirty to forty billion dollars, which is consistent with what tends to happen when requirements are vague and data is worse than advertised. That figure travelled a good deal faster than its method, and has been contested since. The other set of numbers is less ambiguous: OpenAI has committed about 1.4 trillion dollars to data centre capacity over eight years against an annualised revenue run rate its chief executive puts above twenty billion. The arithmetic does not resolve cleanly. The Patrician has seen many sets of numbers that did not resolve cleanly. He knows what generally happens next.

AI expertise commands seven-figure compensation packages. The definition of AI expertise has also expanded dramatically. Someone who completed an online course can now market themselves as an AI engineer through a process of credential and certification inflation that makes assessing actual capabilities increasingly difficult. This is familiar from previous enthusiasm cycles and resolves in the familiar way.

The bubble indicators and the base case¶

Several characteristics suggest bubble dynamics: prices disconnected from fundamentals, investment driven by fear of missing out, infrastructure spending racing ahead of clear demand. In early 2026, Azure grew at 39 per cent with demand still exceeding available supply, and Microsoft’s shares fell anyway, because capital spending was outrunning the growth it funded. Not a shortage of customers, then, but a question about the price of serving them. These do not prove a bubble exists. They do suggest that the gap between enthusiasm and return is becoming harder to defer discussing.

Despite these indicators, the case for genuine transformation has merit. Productivity improvements from AI-assisted coding are measurable and meaningful. Developers produce code faster. Research progress is genuine. The integration into workflows is proceeding directionally across industries. The internet was absolutely a bubble in the late 1990s and also absolutely transformed society. Both things were true simultaneously, which is the difficulty.

The DeepSeek moment, in which a competitive model appeared at a fraction of the assumed cost, cuts in two directions. It undermines valuations premised on enormous training costs as a moat. It also suggests that inference costs will fall faster than expected, which makes applications cheaper to run and potentially accelerates adoption. Whether falling costs primarily benefit users or primarily destroy the business models of those who invested assuming costs would remain high is the question that the next few years will answer.

The Patrician notes that the geopolitical competition between major powers over AI leadership ensures sustained funding regardless of commercial returns. Governments view AI as strategically important, which provides a floor under development that pure commercial dynamics might not. Whether this is reassuring depends on your views about government-directed technology investment, which is a question with a long and instructive historical record.

The Patrician’s assessment¶

AI is genuine technology with real capabilities that will have meaningful impacts. Current investment levels are probably excessive relative to near-term commercial opportunities, and this probability has increased as the evidence accumulates. The technology will matter more over the long term than current sceptics predict while mattering less than current enthusiasts claim. This analysis is not exciting, but it has the advantage of being consistent with how every previous significant technology has developed, which is slowly, expensively, and differently from what anyone predicted.

The next few years will likely see disappointing returns for many AI investments. Some companies will succeed spectacularly. Many will fail quietly. The infrastructure will find uses, though perhaps not the uses that justified building it. The capabilities will improve gradually, possibly faster and cheaper than the infrastructure investors assumed. The economic value will eventually emerge through productivity improvements that are individually modest and cumulatively significant.

The Patrician has seen many enthusiasms come and go. Most proved to be somewhere between genuine innovation and expensive mistake, with the technology eventually settling into usefulness that its early advocates could not quite specify and its early critics refused to credit. AI will likely follow this pattern. The gold rush is real. The evidence that reality is beginning to make its presence felt is also real. Whether the participants strike gold or merely acquire expensive experience in prospecting will be determined by events still unfolding. He recommends keeping receipts.

He notes, with mild interest, that the receipts are beginning to arrive.