Stanford's 2026 AI100 report says the scaling era is under pressure and AI is 'not yet paying its way'

Every few years a panel of academics tries to say what AI actually is, on the record, in one document. On November 10, 2026, the One Hundred Year Study on Artificial Intelligence, known as AI100 and administered out of Stanford's Human-Centered AI institute, released its third report. Its framing is blunt: AI is transforming almost everything faster than anyone predicted, and the method that got it here is running into walls. For a title that tracks the race to superintelligence, that combination is the interesting part.

Who wrote it and why it carries weight

AI100 was launched in 2014 with an endowment from Eric and Mary Horvitz. Every five years its standing committee, currently chaired by Sheila McIlraith of the University of Toronto, commissions a study panel. This year's panel was led by Michael Wooldridge of the University of Oxford and draws on computer science, philosophy and sociology, according to a summary of Stanford's release.

That makes it an unusual counterweight. It is not written by a lab with a model to sell or a fundraising round to justify. It is also not a prediction market. It is a long document with footnotes, and its key takeaways are open on the AI100 site.

The scaling claim

The first takeaway makes the headline argument. The report says AI "owes much to scaling," meaning building systems by using data and compute on an unprecedented scale. It continues that "the scaling paradigm is now under pressure because new sources of high-volume, high-quality data are scarce and compute costs for frontier AI models are prohibitive."

The panel expects the next phase to be defined less by repeated training of ever larger models and more by smarter algorithms and smaller models rebuilt less often, given the right context and memory, more time to reason at inference, specialized training for that, and tools or agents for complex tasks. In its section on general AI, the panel adds that frontier models may soon cost billions of dollars to train and that, with models living perhaps 18 months, such costs "might stretch even the resources of the Silicon Valley giants."

The money claim

The fourth takeaway is the one most likely to be quoted. The report calls capital investment in AI infrastructure "unprecedented in scale and speed," propping up parts of the U.S. economy and fueling fears of a financial bubble. It states that Nvidia had, at the time of writing, a market valuation of over $5 trillion, and that over 15% of U.S. stock market gains in 2025 can be attributed to Nvidia alone, with 60% of market gains from technology and communications.

Then comes the line our readers should hold onto: "the economic impact of AI is not yet evident in aggregate macroeconomic data or in companies' returns on investment: AI is not yet paying its way." The panel warns that a burst bubble would ripple through the global economy and through AI development itself. We reported the market side of that risk in our coverage of the November 5 selloff.

What the report says about general AI

The standing question on general AI is more cautious than any lab roadmap. The panel notes that definitions vary widely, that betting markets place a 50% chance of AGI by 2030, and that a 2023 letter urging a pause in developing more powerful systems drew over 700 tech leaders and 33,000 people in total. It records that the debate returned in September 2026 after "rogue AI" incidents, with industry leaders suggesting a slowdown was appropriate, but that no pause or obvious slowdown followed.

It lays out the optimist and pessimist cases. Optimists point to physical AI, world models and agentic systems as the missing step. Pessimists argue the transformer will not give general intelligence even with those additions, and ask where the data for bigger systems will come from, with synthetic data an unproven answer. The panel also lists limits: no continual learning in frontier models, nondeterministic outputs, and costly retraining. For how forecasters line up, see our AGI timeline tracker.

Cyber, science and the rest

A summary of the release highlights two other areas. In science, the report says AI is compressing timescales from years to months or weeks and that AI-powered labs will be the norm within a decade, but that validating machine-generated discoveries needs new methods. In cybersecurity, it warns that systems have already shown they can find vulnerabilities and carry out attacks using teams of AI agents, and that attacks launched deliberately or inadvertently are a tangible threat.

What we would watch

The report's most testable claim is that the next phase leans on reasoning, memory and agents rather than ever larger training runs. That is checkable. If the 2027 frontier releases keep landing with bigger parameter counts and bigger compute bills, the panel will have misread the field. If labs keep emphasizing inference-time reasoning and tool use, it was right.

Our position: AI100 is useful precisely because it is unexcited. It does not say progress has stopped, and it does not say superintelligence is near. It says the cheapest route to more capability is getting harder, and that the economic payoff is still unproven in the data. Both claims cut against the pace implied by today's capital spending, and both are worth checking against the next two quarters of lab revenue and capex reports. One caution on the evidence: the report is a synthesis by a panel, so its claims about economic returns rest on the aggregate data available to the authors at the time of writing.