Microsoft AI Diffusion Report Shows Adoption Rising as Regional Gaps Widen

Microsoft’s latest AI diffusion report records worldwide usage at 18.8%, despite a widening gap between the Global North and Global South. Almost every measured economy gained users during the second quarter of 2026. However, adoption rose faster in already connected and wealthier markets.

The result complicates the familiar claim that generative AI is spreading everywhere at equal speed. Global adoption is expanding, but access to infrastructure, connectivity, digital skills, and locally useful models still shapes who benefits first.

That tension matters beyond the ranking of individual countries. Governments are treating AI access as an economic priority, while businesses are moving from chatbot experiments toward routine workflows. Microsoft, OpenAI, Google, Meta, and a growing group of open-model developers are competing to supply that demand.

The Microsoft AI Diffusion Report Finds Growth Almost Everywhere

The headline is broad growth, but the more important development is where that growth accumulated.

Microsoft published its AI diffusion report for the second quarter of 2026 on September 21. It estimates that 18.8% of the global working-age population used a generative AI product during the reported period.

That figure increased from 17.8% in the first quarter and 16.3% in the second half of 2025. The gain reached nearly every economy in the dataset, suggesting that generative AI has moved beyond a small collection of early-adopter markets.

Microsoft defines AI diffusion as the share of people ages 15 through 64 who used a generative AI product during a reporting period. Its estimates draw on aggregated and anonymized Microsoft telemetry.

The company adjusts those observations for operating-system and device-market share, internet penetration, and national population. The result is called AI User Share, a population-normalized estimate designed for comparison across countries.

This is not the same as measuring corporate AI purchases, data-center capacity, or the number of locally developed models. It attempts to measure actual product use among working-age people.

The United Arab Emirates remained the leading economy, with estimated AI usage of 73.3%. Singapore followed at 64.3%, while Ireland reached 49.9%, France reached 49.6%, and Norway reached 49.4%.

Those numbers show that model development does not determine adoption by itself. Neither the UAE nor Singapore operates the same concentration of frontier model laboratories found in the United States or China. Both have still produced high usage through connectivity, digital services, government coordination, and internationally connected workforces.

South Korea recorded the largest absolute quarterly increase among the leading economies. Its AI User Share rose 3.5 percentage points to 40.6%, moving the country from sixteenth to twelfth place.

That continued a longer acceleration. Microsoft says South Korea added nearly 15 percentage points over the previous 12 months, the largest gain among major economies across three consecutive reporting periods.

Saudi Arabia made the largest move in the ranking. Its user share rose from 29.4% to 31.4%, helping it climb five places and enter the top 25.

Japan registered the largest relative increase. Usage grew from 22.5% in the first quarter to 24.7% in the second, an increase of roughly 10% relative to its earlier level.

The United States rose from 31.3% to 33%, but remained twenty-first in the ranking. That position highlights the difference between producing AI systems and diffusing them across an entire population.

The country-level methodology originally covered 147 economies and found a strong relationship between AI usage and national income. The underlying research reported a Spearman correlation of 0.83 between AI User Share and gross domestic product per person.

Correlation does not establish that wealth directly causes AI adoption. However, the relationship fits the infrastructure pattern visible across the report. Electricity, internet access, suitable devices, skills, and affordable services all affect whether a person can use generative AI regularly.

The new figures therefore describe two developments at once. AI is reaching more people, yet the markets with stronger digital foundations continue to convert new products into routine use more quickly.

The Global AI Adoption Gap Is Still Getting Wider

Worldwide growth has not closed the divide because the Global North continues to add users at a faster rate.

AI usage in the Global North reached 28.8% during the second quarter, up from 27.5%. The Global South reached 16.2%, rising from 15.4%.

Both groups grew, but their gains differed. The North added 1.3 percentage points, compared with 0.8 points in the South.

The resulting gap widened from 12.1 percentage points to 12.6 points during the quarter. It had stood at 10.6 points in the second half of 2025.

That direction is more consequential than one quarter’s global average. If stronger markets keep gaining faster, the economic and educational advantages associated with AI use can compound before slower markets establish comparable access.

Infrastructure is one part of the explanation. AI products generally require dependable electricity, internet connectivity, compatible devices, and affordable data. A model may be available online without being practically accessible to someone facing unreliable service or high connectivity costs.

Language is another constraint. Many widely used systems perform best in languages with extensive digital text, evaluation data, and commercial demand. Users working in lower-resource languages may encounter weaker answers, limited interfaces, or fewer locally relevant tools.

Digital skills add a third barrier. Opening a chatbot is simple, but productive use often requires users to judge sources, protect sensitive information, structure requests, and connect outputs to real work.

Organizations face related challenges. A business needs suitable data, governance, staff training, and a defined workflow before occasional experiments become dependable operations.

This is why adoption cannot be inferred from product availability alone. The presence of a free web interface does not eliminate the broader costs of connectivity, devices, training, verification, and integration.

The Global North-South categories also require caution. They combine countries with sharply different income levels, demographics, languages, political systems, and technology markets.

The UAE and Singapore, for example, lead the entire ranking despite often appearing within broad geographic discussions of the Global South. Their results warn against treating billions of people as a single technology market.

Researchers interviewed about the divide have also questioned whether the label encourages overly broad assumptions. A useful regional average can reveal inequality while still obscuring differences within each group.

The same issue appears inside wealthy countries. Microsoft’s separate United States research found substantial differences between urban and rural communities. National averages can hide divides related to education, broadband access, employment, and local industry.

That makes the global gap a distribution problem rather than a simple race between two fixed blocs. The central question is whether people in each country can translate available AI into useful, repeatable activity.

Microsoft’s earlier research provides an important clue. In countries with low internet penetration, adoption was often much higher among the connected population than among the total working-age population.

The technical study found that Zambia’s AI User Share rose from 12% of its population to 34% among connected people. Pakistan’s estimate increased from 10% to 33% when measured against internet users.

Those comparisons do not prove that connectivity alone will erase the gap. They do show substantial latent demand among people who can get online.

The report’s core reversal follows from that evidence. Slow national adoption does not necessarily reflect low interest in AI. It can reflect a shortage of the conditions needed to turn interest into use.

AI Adoption by Country Also Reveals Different Uses

The divide concerns not only how many people use AI, but also what early adopters ask it to do.

Microsoft examined a subset of consumer Copilot conversations with a privacy-preserving classifier. The classifier grouped conversations according to their primary intent, such as learning, writing, coding, search, or shopping.

The analysis found that self-directed learning and skill development represented a 29.2% larger share of conversations in the Global South than in the worldwide mix. Schoolwork and academic support held a 32.3% larger share.

Creation and editing tasks also appeared more frequently. These included work involving text, code, and images.

The Global North leaned more toward shopping, product research, information retrieval, and feedback. These are relative shares, not direct measures of the total number of conversations.

The education pattern offers a counterweight to the adoption gap. People in less-connected markets who already use generative AI appear especially interested in knowledge and skill development.

That distinction matters for schools, governments, and product teams. A market with lower overall penetration can still have strong demand for tutoring, translation, writing assistance, and technical learning.

The pattern also has implications for knowledge workers. Useful AI adoption rarely ends with generating an answer. People need to connect responses with reliable documents, previous decisions, and subject-specific context.

A searchable personal knowledge base can help users preserve that context. However, software cannot replace connectivity, source evaluation, or local-language support.

Microsoft warns against extending its conversation findings to the wider population. The data describes early adopters using consumer Copilot, not every person in the Global South.

Demographics may also influence the result. The Global South has a younger population, which can increase the relative importance of schoolwork and skills training.

The dataset does not reveal why users chose particular tasks. A higher learning share might reflect unmet educational demand, younger users, job-market pressure, or the absence of other digital services.

Nor does a conversation category measure learning quality. A student can use AI to understand a concept, draft an assignment, or obtain an unreliable answer. The classification records intent rather than educational outcomes.

Even with those limits, the comparison challenges a narrow view of adoption. Markets with fewer users are not simply delayed versions of wealthier markets.

Their use cases may develop around different needs. Education, multilingual access, public services, and small-business assistance can matter more than shopping or workplace software.

That creates pressure on model providers to compete on relevance, not only benchmark performance. A system that handles local languages, curricula, laws, and cultural context can be more valuable than a stronger general model with limited regional fit.

It also changes the policy question. Expanding AI adoption without improving information quality and user skills can spread unreliable assistance as quickly as useful tools.

The opportunity is therefore tied to responsibility. Educational demand can support wider access, but it also raises questions about accuracy, privacy, academic integrity, and age-appropriate design.

Open-Weight Models Could Lower One Barrier, Not Every Barrier

Open-weight models can improve affordability and local adaptation, but they cannot supply electricity, connectivity, or trained users.

An open-weight model releases the numerical parameters learned during training. Developers can download, modify, and deploy those weights, although the training data and development code may remain closed.

That distinction matters because open-weight does not always mean fully open source. Still, access to model weights gives organizations more control than a system available only through a proprietary service.

Microsoft argues that these models are becoming an important route for global AI diffusion. They can be adapted to local languages and specialized domains without requiring an organization to train a frontier model from the beginning.

They also allow developers to choose among hosting providers or run a model on their own infrastructure. That flexibility can reduce dependence on a single vendor and support deployments with stricter data requirements.

Open-weight usage is difficult to count through consumer chatbot traffic. People can encounter these models inside applications, coding agents, public services, or company systems without knowing which model produced the output.

Self-hosted deployments create another blind spot because their traffic may never reach a centralized provider. Microsoft therefore examined application programming interface activity as a complementary signal.

The report draws on OpenRouter, a service that routes developer requests across many open and proprietary models. Microsoft says its analysis of public data found open-weight systems handling 75% of token usage on that platform.

That number is striking, but it is not a worldwide market share. OpenRouter attracts developers who are unusually willing to compare providers and switch models. It also excludes much direct usage of proprietary systems and privately hosted open models.

Earlier OpenRouter analysis covering more than 100 trillion tokens found open-weight systems at roughly one-third of volume by late 2025. Releases from DeepSeek, Kimi, Qwen, and OpenAI reportedly produced increases that continued after their launch periods.

The persistence matters because launch traffic can represent curiosity rather than dependable adoption. Sustained token volume offers stronger evidence that developers are placing the models inside recurring workflows.

Open-weight systems have also narrowed parts of the capability gap with proprietary models. The report says leading open releases often approach the closed frontier within several months, although performance varies by task.

That pattern encourages model routing. An application can reserve a high-capability proprietary system for difficult requests while assigning routine work to a less costly open model.

For organizations processing large workloads, that choice can influence whether an AI service remains economically viable. Agents can generate repeated model calls, making inference efficiency more important than a single impressive response.

The implications are particularly relevant in lower-income markets. Universities, nonprofits, governments, and local companies can adapt an existing model instead of financing an original frontier-scale system.

Microsoft cites its Bring Your Own Language research as one example. The project improved average performance for Chichewa and Māori by about 12% across a collection of benchmarks through data refinement and model adaptation.

That result does not mean every low-resource language can receive the same improvement. It demonstrates why editable weights and carefully prepared local data can matter.

DeepSeek provided an earlier example of accessibility affecting adoption. Its R1 release helped increase usage in China and attracted users across developing markets, according to reporting on Microsoft’s previous findings.

The developing-market response showed how a widely available model can shift adoption outside the markets served most directly by American platforms.

However, open models cannot remove the foundational bottlenecks identified by the report. A downloadable model still needs hardware, electricity, technical staff, security controls, and an application that people can access.

Local adaptation also demands suitable data. Communities with limited digitized text may have the greatest need for language support and the least material available for training or evaluation.

Open weights therefore address only part of the main opponent in this story: the promise of universal AI access versus the uneven conditions of actual use.

They can reduce model-level costs and improve local relevance. They cannot independently close a gap produced by infrastructure, skills, institutions, and income.

Microsoft’s Measurement Has an Important Blind Spot

The figures provide a useful directional benchmark, but they do not represent a complete census of global AI activity.

Microsoft acknowledges that no single indicator can fully measure AI adoption. Its telemetry offers timely behavioral evidence, yet it naturally reflects the products, devices, and traffic visible to Microsoft.

The company adjusts for differences in operating systems, device access, internet penetration, and population. Those corrections make comparisons more meaningful, but statistical adjustments cannot observe every AI service.

The problem becomes especially visible in China. Domestic applications from Chinese companies can attract extensive use without appearing proportionally in Microsoft’s current product coverage.

Microsoft says its next report will expand the list of measured AI tools. It expects the change to increase estimated adoption in almost every economy, with a larger effect in China.

That future revision will mostly represent better measurement of existing use, not a sudden adoption surge. Readers should therefore avoid treating the next increase as purely organic growth.

The number of available applications is also rising quickly. Hugging Face Spaces, which hosts runnable AI demonstrations and applications, grew from approximately 124,000 at the end of 2023 to 1.46 million by September 2026.

Microsoft estimates that the total increased by about 83% since its first diffusion report in late 2025. These Spaces range from personal experiments to more production-oriented projects.

They should not be counted as 1.46 million distinct commercial products. Their growth still shows why a fixed list of monitored tools becomes less representative over time.

The public report repository explicitly lists other limitations. AI User Share measures usage, not capability, productivity, safety, or economic impact.

A person who briefly tests a chatbot and a professional who uses several models daily may both appear as users. Their economic exposure and practical benefit can be entirely different.

The metric also does not prove that AI improves work. Usage can include productive assistance, low-quality outputs, entertainment, schoolwork, experimentation, or automated activity embedded in another service.

Microsoft has a commercial interest in expanding AI adoption. It sells cloud infrastructure, model access, productivity software, and consumer AI services.

That interest does not invalidate the research, especially because the company publishes its methodology and downloadable data. It does make independent comparison important.

Surveys, API traffic, workplace studies, download data, and national statistics can each test a different part of the picture. None offers a complete view on its own.

The regional labels introduce another measurement challenge. The Global North and Global South are analytical groupings, not uniform economic or political units.

Averages can reveal that access is unequal while hiding leading markets, fragile regions, and large differences within countries. Reporting should keep both levels visible.

The fairest interpretation is that Microsoft has produced a broad and frequently updated signal. It is more useful for comparing direction and relative adoption than for declaring an exact count of every AI user.

The finding that the gap widened remains meaningful within that framework. The same methodology shows both regions growing, with the North advancing faster across successive reporting periods.

Still, the planned coverage expansion can change country rankings and regional estimates. The next dataset will be a methodological test as well as an adoption update.

Three Signals Will Show Whether the Divide Starts Closing

The next phase depends on whether broader measurement, cheaper model access, and foundational infrastructure begin moving in the same direction.

The first signal is Microsoft’s expanded tool coverage. The next report should show how much usage the present method misses, particularly in China and markets dominated by domestic applications.

A large upward revision would weaken country comparisons based on the current rankings. It would not necessarily erase the North-South gap, but it could show that part of the divide reflects visibility rather than behavior.

A modest revision would strengthen the current interpretation. It would suggest that unequal infrastructure and access remain the dominant explanation even after more products enter the measurement.

Readers should separate methodological revisions from quarter-to-quarter growth. Microsoft has already warned that broader coverage will capture activity that exists today.

The second signal is sustained open-weight usage outside developer routing platforms. OpenRouter’s token mix points toward rapid adoption, but its users do not represent the entire market.

Evidence from cloud providers, enterprise deployments, universities, and public-sector services would show whether open models are becoming durable infrastructure. Growth in localized applications would be especially relevant.

The strongest evidence would connect model availability with new users or better services in lower-adoption markets. Token volume alone cannot show whether the benefits are reaching people who previously lacked access.

A reversal would also be informative. If open-model use concentrates among well-funded developers in wealthy markets, improved model economics may reinforce existing advantages instead of closing them.

The third signal is progress in connectivity and practical skills. The internet access data maintained by the International Telecommunication Union gives essential context for any adoption ranking.

Watch whether countries with low population-level AI use narrow the difference between total and connected-user adoption. That would indicate that access programs are reaching beyond an already connected minority.

Skills programs require similar scrutiny. Enrollment totals are weaker evidence than continued use, completed training, workplace application, or measurable educational outcomes.

For business buyers, the report offers a straightforward warning. A global AI strategy cannot assume identical infrastructure, language support, or employee readiness in every market.

For developers, localization is not a secondary feature. It can determine whether a model that performs well in a benchmark becomes useful in a classroom, clinic, government office, or small business.

For knowledge workers, rising adoption increases the value of source verification and durable context. More AI access means more generated material, not automatically more reliable knowledge.

The Microsoft AI diffusion report shows that generative AI is still spreading quickly. It also shows that growth alone does not distribute opportunity evenly.

The next question is not whether worldwide adoption will rise again. It is whether connectivity, skills, localized models, and broader product access let lower-adoption markets grow faster than the leaders.

Watch the next methodology update, real-world open-model deployments, and connected-population usage together. If all three improve, the widening gap can begin to narrow. If they separate further, a globally available technology will keep producing sharply regional outcomes.