The OECD.AI Index 2025 update brings with it some exciting new data. The Index is reassessed each year to ensure the inclusion of the best available AI-related data and information. This year brings with it several notable changes.
New indicator: GPU availability
A new indicator has been added in the Index for 2025 regarding the availability of AI-capable GPUs by country. The data was developed in partnership with Oxford University Innovation by identifying the availability of public cloud-based compute resources relevant to AI across cloud regions. The underlying data can be explored on OECD.AI, where further information on GPU types by region and shares by country.
The Index uses the data by giving a score of 1 to a country if it has at least one cloud compute region capable of Tier 1 workloads (high AI capability), 0.5 for countries with GPUs capable of only Tier 2 or Tier 3 workloads (medium and low capabilities, respectively), and a score of 0 for countries without cloud compute regions or with cloud regions with no AI compute capability. GPUs are categorised based on expert classification and are documented in the methodological note.
Google trends dropped from the Index, Wiki-data added
Data on Google trends were used in 2023 and 2024 to signal a country’s knowledge and awareness of AI concepts. While the data was valuable in assessing the relative search volumes on AI topics, the data collection method was discontinued, and maintaining the data source proved impossible. The data were therefore swapped with Wiki-data on site visits to AI-related topics.
In the future, OECD.AI is continue their partnership with Google, including by participating in the review of an alpha version of a Google Trends API. Once the API progresses to a point where the relevant data becomes readily available, OECD.AI will assess whether this indicator can be re-inserted into the Index.
To ensure this dynamic is still captured, a new indicator was developed that finds the visits to Wikipedia AI-related webpages. AI pages were identified by finding related pages through the Wikipedia internal relationship structure. The data are now available on OECD.AI, and you can find more details on the data in the methodological note.
The new indicator was incorporated into the Index using the average monthly share of visits to AI-related Wikipedia pages relative to total visits to Wikipedia across all pages in each country.
AIKOD dropped, Artificial Analysis added
Data collection for the AI knowledge on Demand (AIKOD) database has ceased, given the extensive manual data collection process required. In its place, OECD.AI has established a data partnership with Artificial Analysis, an industry leader on AI benchmarking and one of the main inputs into the original AIKOD database. The underlying data can now be further explored on OECD.AI, including models by modality, by whether the model has open-weights or closed, by developer country, and data on quality of models. See the methodological note for more information about the data.
The data for Artificial Analysis was used in the Index by taking the cumulative sum of models developed in each country up to the end of each year.
Methodological changes to existing indicators
A new source for languages spoken per country will be used for all years of the Index for the Hugging Face indicator on datasets by language. Previously, the World Factbook was queried to find the top two languages, and then those languages were normalised such that the percentages added up to 100. However, the World Factbook often did not display percentages, and would list the official languages first. This indicator is meant to measure whether the local populations are being served in their spoken languages, so this misalignment on population shares often led to bias. To address this issue, the source for languages by country was changed to the open-source project CLDR, which provides exact percentages based on the most spoken language.
A second change was made to normalise supercomputers by population. Previously, the decision had been made not to normalise supercomputers to account for the network effects these large computing environments can have. However, this indicator was the only one in the Index that was measured by counts and not normalised. The decision was made, in consultation with national delegates, to normalise this indicator for consistency.