Stanford AI100 Finds Data and Compute Costs Pressure AI Scaling

Stanford’s One Hundred Year Study on Artificial Intelligence (AI100) released its third five-year review Tuesday, Nov. 10, as scarce data and frontier-model computing costs challenge the scaling approach behind recent generative AI gains.

Scaling means expanding the data and computing resources used to train a model. That strategy helped drive major advances in generative AI over the past five years. High-quality material is getting harder to find, while the computing expense of frontier AI models is high, putting pressure on the economics of repeating ever-larger training runs. The 2026 review, titled AI Goes Mainstream: The Scaling Era and the Rise of Generative AI, covers the period when generative AI entered broader use.

A broad calculation made during the panel’s September writing period puts a frontier training run at tens of trillions of tokens and upward of 100,000 high-end GPUs working without interruption for several months. The estimate is not attached to a named system, and precise requirements for newer models have not been made public.

Possible routes to further gains include more efficient algorithms and potentially smaller models. Better use of context and memory, along with reasoning while a model responds, may also contribute. Specialized training and tools or agents designed to handle complex tasks are among the other approaches identified for future development. The emphasis is shifting toward how a model uses information and performs after its initial training, alongside the size of the training run.

The economics extend beyond model training. Companies spend billions of dollars on infrastructure for state-of-the-art AI, and data-center computers are often replaced after several years. New models are often developed through additional training rather than by refining existing systems, raising questions about whether returns will justify continued investment.

Frontier AI work is concentrated among a few developers in the United States and China. That concentration puts governance in focus, including the public’s role in decisions about how AI is developed and deployed.

AI100 began in 2014 and reviews AI development in five-year increments; earlier reports came out in 2016 and 2021. The latest review considers technology’s economic, social and geopolitical effects as it moved into mainstream use. It does not establish that scaling has stopped, leaving open both whether the era is ending and when.