Global GenAI Usage Reaches 18.8%; South Korea Ranks 12th Worldwide at 40.6%

The AI Economy Institute, Microsoft’s think tank, said in the "Global AI Diffusion: Q2 2026 Trends and Insights" report released on the 28th that it compared generative AI usage levels around the world and by country in the second quarter of 2026. The report disclosed the current state of AI use by country, regional usage gaps, and trends in the spread of open-weight models.

According to the report, the global generative AI usage rate reached 18.8% in the second quarter of 2026, up 1.0 percentage point from the previous quarter. South Korea’s share of AI users came in at 40.6%, placing the country 12th in the world. South Korea also recorded the largest increase among major economies for the third consecutive time.

The report was based on aggregated and de-identified Microsoft telemetry data. The population measured was the global working-age population between 15 and 64, and the metric used was the share of generative AI product usage. The calculation reflected operating system market share, device market share, internet penetration, and differences in population size by country.

Global generative AI usage in June 2026 stood at 18.8%, up 1.0 percentage point from 17.8% in the first quarter. Over the same period, South Korea’s usage rate rose from 37.1% to 40.6%, an increase of 3.5 percentage points. South Korea’s global ranking also climbed from 16th to 12th, up four places.

Microsoft said South Korea’s share of AI users has increased by about 15 percentage points over the past 12 months. It also said South Korea recorded the largest increase among major economies three times in a row, and that the pace of AI adoption in South Korea continues to accelerate.

In the country-by-country comparison, the United Arab Emirates ranked first with an AI usage rate of 73.3%, while Singapore ranked second with 64.3%. Ireland’s AI usage rate was 49.9%, and France’s was 49.6%.

Saudi Arabia recorded the largest gain, with its ranking rising five places from 30th to 25th. Japan’s usage rate increased from 22.5% to 24.7%, up 2.2 percentage points. Japan posted the highest relative growth rate from the previous quarter.

The pace of AI diffusion varied by region. In the second quarter, AI usage stood at 28.8% in the Global North and 16.2% in the Global South, leaving a regional gap of 12.6 percentage points. The gap widened from 10.6 percentage points in the second half of 2025 to 12.1 percentage points in the first quarter of this year and 12.6 percentage points in the second quarter, and the disparity continues to expand.

Microsoft pointed to differences in foundational infrastructure such as power, data centers, and internet access as a backdrop to these gaps. It also cited differences in digital skills and language support.

Regional differences also appeared in how AI is used. According to an analysis of consumer Copilot usage purposes, the share of conversations in the Global South for self-directed learning and skill development was 29.2% higher than the global average, while the share for schoolwork and academic support was 32.3% higher than the global average.

The main AI use cases in the Global South were learning, image generation and editing, text generation and editing, and code generation and editing. By contrast, the Global North showed a relatively stronger share of consumer-oriented uses such as shopping, search, and feedback.

Microsoft said that if AI is not to widen the digital divide between countries and regions, technology diffusion must be accompanied by basic usage conditions. The core foundations it identified were electricity, data centers, internet access, AI utilization capabilities, and local language support.

A key trend linked to these accessibility issues was the growth of open-weight models. Open-weight models are characterized by the release of trained weights, the ability to run in a developer’s own environment, and the flexibility to adapt them to specific purposes.

The report cited independent benchmark studies and said the performance gap between open-weight models and closed models is narrowing rapidly. After a new closed frontier model sets the top performance mark, a leading open-weight model reaches near the previous best level within months, the report said, noting that this does not mean open-weight models are now on par with the latest closed frontier models.

The report said performance gains in open-weight models have translated into greater real-world usage. OpenRouter provided analysis of more than 100 trillion tokens, and said that as of the end of 2025, open-weight models accounted for about one-third of total token usage.

Microsoft further analyzed publicly available OpenRouter data. The additional analysis found that the share of open-weight models had risen to 75%, and that usage continued to increase after the release of major models such as DeepSeek, Kimi, Qwen, and GPT-OSS.

However, the figures are based on usage within developer-oriented platform OpenRouter. The report said caution is needed in interpreting these numbers as an expansion of global AI market share.

Microsoft said open-weight models could become a path to improving AI access in the Global South. However, it said open-weight models alone cannot directly solve structural constraints such as power, internet access, and digital skills. Instead, it said open-weight models can lower the cost of using AI and ease the burden of adapting models to local languages and industry needs.

It is noteworthy that for nonprofits, companies, governments, and universities, fine-tuning existing state-of-the-art AI models rather than building new ones from scratch could be a practical alternative. This approach skips initial training and adapts existing models for use in specific languages or fields. As a result, analysts said it could expand options for developing AI services tailored to local needs in regions with limited data and computing resources.

The applicability of this approach was also demonstrated in cases involving low-resource languages. Microsoft AI for Good Lab used data cleaning and adaptation of existing language models through the "Bring Your Own Language" framework. As a result, performance for Chichewa and Maori improved by an average of about 12% across several benchmarks, laying the groundwork for application to languages such as Inuktitut, which have limited available data.

At the same time, not only models themselves but also AI application development is showing rapid growth. Hugging Face Spaces is a platform that provides runnable AI applications and demos. Its scale grew from about 124,000 at the end of 2023 to about 1.46 million at the end of August this year.

Since the first AI diffusion report was released in November 2025, the increase has been about 83%. The development scope has expanded from personal experiments and research projects to applications close to real services, and Microsoft analyzed this as rapid diffusion into the application stage of model use.

Accordingly, the AI Economy Institute plans to expand the scope of its measurements in the future. Starting with the next report, it plans to incorporate newly emerging AI tools into its metrics and to analyze country usage rates as well as usage patterns and changes in the model ecosystem more broadly.

Source: TECHWORLD · Kim Seung-ki
Original: https://www.epnc.co.kr/news/articleView.html?idxno=407406

References

• https://www.epnc.co.kr/news/articleView.html?idxno=407406