Stanford HAI’s 2025 AI Index research chapter identifies three distinct scaling rates. It estimates that training compute for notable AI models doubles about every five months, datasets used to train large language models double about every eight months, and the power required for frontier training doubles about once a year.
Those rates are considerably faster than the cadence associated with Moore’s Law, which describes transistor counts doubling roughly every two years. Measured only by the frequency of the doubling, Stanford’s reported rates are 4.8 times as fast for training compute, three times as fast for datasets and twice as fast for training power. They measure inputs, however, and should not be treated as equivalent rates of improvement in model capability.
The five-month doubling
Stanford’s compute figure covers “notable” models, meaning systems selected for technical significance, influence or strong industry attention. The report says the increase has been especially pronounced over the five years preceding its 2025 publication. It does not mean every model or laboratory doubles its training budget on an individual five-month schedule.
The scale of the change is nevertheless substantial. IEEE Spectrum reported from the AI Index and Epoch AI estimates that Google’s original 2017 Transformer cost about $930 to train, while Stanford’s 2025 chapter places estimated costs for several later frontier models in the tens or hundreds of millions of dollars. At a constant five-month doubling rate, the estimated compute attached to the trend would multiply 64-fold over 30 months.
Dataset growth creates a different constraint
The eight-month dataset rate describes the number of tokens used for LLM training, not the number of unique documents available online. Stanford’s chapter traces the rise from roughly 2 billion tokens for the original Transformer to hundreds of billions for GPT-3 and about 15 trillion for a major 2024 model. A mechanical continuation from 15 trillion would produce about 42 trillion tokens after 12 months and 120 trillion after 24 months.
That arithmetic explains why data supply has become a constraint, but it does not establish a precise exhaustion date. Stanford cites an Epoch AI projection placing full utilization of the available training-data stock somewhere between 2026 and 2032, with a wide confidence interval. The timing depends on filtering, repeated use of data, access to private material and whether developers deliberately overtrain models to reduce later inference costs.
Synthetic data may expand the supply, but Stanford’s 2025 chapter says its effective use still requires further research. The 2026 AI Index update goes further, reporting that synthetic data had still not replaced real data in pre-training even as data curation and post-training methods improved.
Training power is doubling annually
The power trend moves the scaling problem from software into physical infrastructure. Stanford reports that the total power draw required to train frontier models has been doubling annually, with its historical series extending through 2024. Power is measured in watts and describes the rate at which electricity is drawn, while total energy consumption also depends on how long the training run lasts.
The same chapter estimates that AlexNet training produced 0.01 tons of carbon dioxide equivalent, compared with 588 tons for GPT-3, 5,184 tons for GPT-4 and 8,930 tons for Llama 3.1 405B. Using the report’s comparison of 18.08 tons per average American per year, the Llama estimate is approximately equal to the annual emissions of 494 people. These remain estimates because developers do not consistently disclose every input needed for direct accounting.
Hardware efficiency is improving at the same time. Stanford reports annual gains of about 43% in 16-bit hardware performance and 40% in energy efficiency, but those improvements have not prevented the estimated power draw of the largest training runs from rising. Efficiency and total system demand describe different layers of the problem and cannot simply be subtracted from one another.
Who can build at this scale
The capital and infrastructure requirements help explain why industry dominates notable-model development. Stanford reports that U.S.-based institutions produced 40 notable models in 2024, compared with 15 from China and three from Europe, while nearly 90% of that year’s notable models originated in industry. The industry share had risen from about 60% in 2023.
The 2026 update extends that pattern: industry produced more than 90% of notable models in 2025, while U.S. institutions produced 59 and Chinese institutions produced 35. That shows continued U.S. leadership alongside rapid Chinese growth, rather than a simple claim that geographic concentration is tightening everywhere.
Inference is getting cheaper as training gets bigger
The economics look different once a model has been trained. Stanford found that the price of querying a model at roughly GPT-3.5 performance on the MMLU benchmark fell from $20 per million tokens in November 2022 to $0.07 by October 2024. That is a reduction of more than 280 times in approximately 18 months.
The Index estimates that inference prices fell between nine and 900 times per year, depending on the task and performance threshold. Frontier training can therefore become more expensive while access to an established level of capability becomes dramatically cheaper, allowing the same class of performance to spread into far more products.
Education shows the downstream effect
The spread is already visible in education. Stanford’s 2026 education chapter reports AI use for schoolwork among four-fifths of U.S. high school and college students. It also says only half of middle and high schools have an AI policy and that just 6% of teachers consider their school’s policy clear.
The same chapter reports a 22% increase in new AI PhDs across the United States and Canada between 2022 and 2024. Stanford says the numerical growth went into academia while the share entering industry stayed flat, reversing the direction of the previous decade’s employment trend.
Transparency has not kept pace
Larger training runs have not produced fuller public disclosure. Stanford’s 2023 introduction to the Foundation Model Transparency Index explains that researchers assessed 10 major developers across 100 indicators. The highest scores ranged from 47 to 54 out of 100, while the lowest was 12.
The 2025 AI Index also explains that Epoch AI sometimes estimates training compute from hardware specifications and usage patterns or from model architecture and training data. Stanford’s 2026 update says training code, parameter counts, dataset sizes and training duration are no longer disclosed for several of the most resource-intensive systems. The headline curves are therefore researched estimates, not a complete ledger supplied by the companies themselves.
What the curves imply for infrastructure
The figures describe related pressures, but they should not be collapsed into one measure. Compute tracks arithmetic operations, dataset size tracks training tokens, and power tracks instantaneous electrical demand. Their interaction influences demand for accelerators, data centers and electricity, but one curve does not automatically force the other two to retain the same rate.
Stanford’s 2026 report places global AI data-center power capacity at 29.6 gigawatts and lists 5,427 data centers in the United States. Those totals include infrastructure for training and inference, so they should not be presented as the footprint of frontier training alone. They do show why model scaling increasingly depends on facilities and power systems outside the laboratory.
The comparison that puts the rates in scale
Against a two-year Moore’s Law interval, the reported five-month, eight-month and 12-month doubling times occur 4.8, three and two times as frequently. If those rates continued unchanged for 36 months, the arithmetic would imply roughly 150 times as much training compute, 23 times as many training tokens and eight times the training power. Those are mechanical extrapolations, not predictions that laboratories will obtain the necessary capital, data and electricity.
The Stanford findings are most useful as a picture of pressure rather than destiny. Compute growth can slow, model architectures can become more efficient, datasets can be reused or curated differently, and infrastructure limits can alter investment decisions. What the 2025 Index establishes is that the three principal inputs to frontier AI were rising on unusually short clocks, while disclosure about those inputs remained incomplete.
