Over time, it's clear that models have scaled across all of the attributes examined, particularly in terms of training compute required, number of trainable parameters, and training time. It is notably more difficult to examine patterns over time for power draw and training cost, as these metrics were generally not recorded until after 2010.

The deep learning era (2010-2025) contains the highest concentration of model releases, as expected, and also shows an increase in usage for the metrics examined. This implies the evident significance of the deep learning era as an inflection point in AI development, likely due to other factors such as the introduction of graphical processing units (GPUs) for general purpose parallel computing.

Seeing how overall model scale has grown over time, it is significant to note that the use of these resources required to train AI models has become generally more efficient. In many domains, scaling leads to increased expenses on a per-unit basis. Even in an algorithmic complexity sense, for example, more complex algorithms tend to have worse asymptotic constraints. The fact that larger AI models yield more efficient usage, e.g. lower cost per parameter and cost per FLOP, thus runs counter to traditional intuitions about scaling.

In other words, while most technologies show diminishing returns at high scale, there's no apparent saturation point yet for the large-scale AI models we're currently at. This could suggest that improving models by scaling up is a more efficient path.

Models have changed in how they can be accessed and used by the public, with more institutions offering open weights, i.e. where a model's trained parameters are publicly shared, yet not having access be a significant factor in the scale of a model's compute. This means that some of the largest-scale models in terms of compute that have been trained to date generally have some level of transparency and accessibility. However, given this large scale, it also means that users would need the necessary computational power to actually use these weights efficiently, which may not be feasible.

The United States and China lead global AI development in terms of the total number of AI models released over time. This is expected, given that many of the institutions that are key players in AI development are based in the United States, such as OpenAI, Google, Meta AI, and NVIDIA.

Also expected are the remainder of the countries which make up the top five. Namely, the United Kingdom is home to Google DeepMind, and Canada is home to academic institutions that led key methodological innovations particularly in the earlier stages of AI, such as backpropagation and the ReLU activation function.

Reflecting the top countries in terms of number of AI models released, it's interesting to note that while both the United States and China produce large models, the United States has greater variation in its export controls on semiconductors and intensity of AI-related publications. Whereas, China tends to have low export controls and a high intensity of AI-related publications.

Further, for the United States, the distribution of training compute is fairly comparable across all levels of export controls, indicating that export controls have not significantly limited available training compute.

At both a country-level and organization-level, compute is not significantly related to model accessibility, reflecting the earlier temporal findings.

Across the deep learning era, efficiency metrics have generally increased, also reflecting the earlier temporal findings. As expected, the models with the highest levels of training compute are frontier models, given that Epoch AI defines a frontier model in the dataset as models in the top 10 by training compute at the time of their release, a threshold that has increased over time as larger models are developed.

At the organization level, the data records xAI's 2025 Grok model as having the greatest efficiency in training compute FLOPs per training hour.

As training compute has grown, so have training costs. This is expected, and further seen where frontier models, i.e models leading in training compute, generally exhibit higher training costs than non-frontier models.

Several training metrics are highly correlated, particularly, power draw and compute, and power draw and cost. This is also expected given the nature of the data, as some of the values for these variables were derived directly from each other by Epoch AI, when values were not reported by the model's organization.