How the U.S.–China AI Rivalry Is Being Decided by Everyday Users Around the World

A RAND study finds that while U.S. AI models still dominate global use, their lead is fragile, as users can easily switch to competitors once performance gaps narrow. China’s DeepSeek showed how quickly market shares can shift, especially in developing countries, when a credible alternative emerges.

The competition between the United States and China over artificial intelligence is often framed as a battle of research labs, computing power, and national strategy. But a new report from the RAND Corporation’s Center on AI, Security, and Technology suggests the real contest is unfolding somewhere far more mundane: in the daily choices of users around the world. Drawing on global website traffic data collected by Similarweb across 135 countries, the study tracks how people actually use large language models (LLMs), offering a rare view of AI rivalry as it plays out in real life rather than policy speeches or technical benchmarks.

U.S. Dominance, but on Shaky Ground

By the numbers, the United States still leads decisively. American models such as ChatGPT, Gemini, Claude, Copilot, and Perplexity account for roughly 93 percent of global LLM website visits as of August 2025. This dominance reflects a strong first-mover advantage: U.S. companies released powerful models earlier, built recognizable brands, and set user expectations. Yet the report warns against complacency. User loyalty in the AI market is thin, switching costs are low, and leadership depends less on history than on staying ahead in performance.

The DeepSeek Shock

That vulnerability became clear in January 2025 with the launch of DeepSeek R1, a Chinese-developed model. In just two months, traffic to China-based LLM platforms jumped by more than 460 percent. China’s global market share leapt from about 3 percent to 13 percent almost overnight, largely driven by DeepSeek. Importantly, this surge did not simply redistribute users within China’s AI ecosystem; it expanded China’s overall global footprint. Although China’s share later declined to around 6 percent as U.S. platforms continued growing, the episode showed how quickly the market can shift when users see a credible alternative.

Where Chinese Models Are Gaining Ground

Geography matters. Before 2025, Chinese LLM use was largely limited to China and Hong Kong. After DeepSeek’s release, adoption increased everywhere, but the biggest gains appeared in developing countries and regions with closer economic or political ties to Beijing, including Russia, parts of Africa, the Middle East, and South America. Wealthier, U.S.-aligned countries in Europe, East Asia, and Australia saw much smaller increases. Statistical analysis shows that after January 2025, lower-income countries were significantly more likely to adopt Chinese models, even when political systems and past usage are taken into account.

Why Price, Language, and Diplomacy Don’t Explain It

The report challenges several popular explanations for China’s gains. Price differences, while dramatic on paper, appear mostly irrelevant. Chinese models are often far cheaper than U.S. rivals in paid API pricing, but nearly all platforms offer free tiers, and about 98 percent of users rely on them. As a result, most people never encounter the price gap. Language support also fails to explain adoption. While U.S. models once had a clear lead, Chinese developers have rapidly expanded multilingual capabilities, with some models now claiming support for more than 100 languages.

China’s aggressive AI diplomacy, hundreds of embassy announcements promoting AI cooperation worldwide, also seems to have limited influence on everyday users. While such efforts may matter for government partnerships or large corporate deals, most LLM use comes from individuals making independent, online choices.

Performance Is What Really Matters

In the end, the evidence points to a simple conclusion: performance drives adoption. Before 2025, Chinese models were not good enough to compete globally. DeepSeek changed that by narrowing the performance gap just enough to make switching worthwhile. And switching is easy. There are no proprietary formats, no retraining costs, and little friction in moving between AI tools. Users follow quality and convenience, not flags or diplomatic signals.

For the United States, the message is clear. Global leadership in AI remains strong, but it is not secure. As competitors improve and users become more willing to experiment, dominance will depend on continuously delivering better models, not on past success or geopolitical influence. In the fast-moving world of AI, even a commanding lead can erode in months.