How China Is Winning the Global AI Race

Foreign Policy·🕐 1 sa önce·👁 0 görüntülenme

Cutting-edge U.S. models are too expensive for much of the world.

Which artificial intelligence model is the most popular these days? Ask anyone in America or Europe, and you’ll probably hear about the respective merits of OpenAI’s ChatGPT, Anthropic’s Claude, or Google’s Gemini. All wrong. Over the past two weeks, the most widely used AI in the world was one that few Westerners had ever heard of: Kimi K2.6, an open-source Chinese model that topped the OpenRouter leaderboard.

This ranking highlights how Western policymakers and CEOs could be fixating on the wrong race as they focus on semiconductor benchmarks and the question of which AI model is the most advanced. China, meanwhile, is quietly building something different: an ecosystem of open-source models that are both cheap and good enough for most use cases. By the time Western capitals notice, Chinese AI models may well have become global standards and prove hard to displace—even by more advanced technology.

The mainstream Western narrative is that U.S. export controls are preventing China from getting access to cutting-edge processor chips, slowing down the development of Chinese AI firms. This is true but eclipses a key part of the story: If Beijing’s AI pitch centers on universal access and cost-effectiveness, then Chinese AI firms do not need the latest chips to win the global AI race.

Take Kimi K2.6, the open-source flagship model of Beijing-based Moonshot AI. On industry benchmarks, the model sits within touching distance of cutting-edge Western ones such as Anthropic’s Claude Opus 4.7 and OpenAI’s GPT-5.5. On price, it is in a different universe: Kimi costs roughly $4 per million output tokens, about six to eight times cheaper than Opus 4.7 and GPT-5.5. (A token is a unit measuring the input and output of AI models.) For a casual user, this cost difference may not matter much. For a company running hundreds of AI agents, it can be decisive.

Kimi is not an outlier. While the Moonshot model tops the charts for usage through an interface, another Chinese model, Alibaba’s Qwen, is fast becoming the default ecosystem for self-hosted AI models. Two data points help illustrate this point. First, by March, Qwen had captured more than 50 percent of global open-source model downloads, having overtaken its biggest Western competitor, Meta’s Llama, in late 2025. Second, Qwen has been downloaded around 1 billion times. Interest in Qwen is not confined to cost-minded firms. Last November, the Singaporean government announced that it would ditch Llama and build its sovereign AI model with Qwen instead.

With its open-source AI strategy, Beijing is revamping the logic of its Belt and Road Initiative (BRI)—but with a twist. BRI entailed Chinese firms delivering fully financed infrastructure projects in a bid to lock third countries into China’s orbit. The rollout of Chinese open-source AI models follows the same logic, but this time the infrastructure is both invisible and free. The marginal cost of AI diffusion is near zero (the main costs—servers and the electricity to power them—are borne by the hosting country), which makes the AI rollout a far cheaper investment for Beijing than building ports, railways, or power plants. What’s more, BRI created highly visible, Chinese-owned infrastructure that sometimes proved unpopular. By contrast, AI dependency is invisible to both policymakers and the population, limiting pushback.

China’s AI bet is a long-term one with a clear end goal: ensuring that Chinese AI models become (and thus shape) global standards. The underlying logic is that a default AI tool could quickly become a de facto industry standard. Once developers and firms build on China-made architectures, they use Chinese technical assumptions and standards—thus giving Beijing long-term influence. It’s part of a larger strategy; in its Standards 2035 blueprint, Beijing sets itself the goal of seeing Chinese products become global defaults in a bid to shape standards for next-generation technologies. As a precedent, China has been handing over access to its Logink shipping software for free in recent years. The platform quickly became a household name in the shipping sector after being deployed in at least 86 ports across 24 countries.

The battle to set global AI standards will mostly be waged in the global south. Western economies are locked into U.S. models, while China runs on Chinese ones. The swing states are everywhere else, and three factors make Chinese models well-placed to make headway there. First, U.S.-made AI is too expensive for broad rollout in cost-sensitive developing economies. Second, U.S. models are typically trained with Western data, making them ill-suited to grasp local contexts in global south countries. By contrast, open-source Chinese models can be downloaded and fine-tuned with country-specific data. Exhibit A is AfriqueQwen-14B, a Qwen-based AI that was adapted to 20 African languages through training on African data. The flagship Western open-source model, Llama, offers only patchy coverage of global south languages.

Third, growing global resentment against the United States is reinforcing China’s strategy. To make the most of this, Beijing does not even have to lift a finger. Washington frames global AI competition as a national security race to be won against China, with no offerings to third countries needing cost-effective, good-enough tools.

The outcomes of two major AI summits for global south economies illustrate this point. Last April in Kigali, Rwanda, the Global AI Summit on Africa and resulting Africa Declaration on AI called for AI governance built around ethics, sustainability, and responsibility—something that would probably sound alien to Silicon Valley tech executives.

At the India AI Impact Summit in February, global south policymakers doubled down. New Delhi anchored the conference’s agenda in a “People, Planet, and Progress” framework emphasizing social empowerment, development, and inclusion. As Amitabh Kant, India’s former G-20 sherpa, put it on the sidelines of the conference: “We are providing more data to OpenAI than the U.S. Data from [the] global south is helping refine models. These will sell you high-cost products. So India needs to build models on its own data.”

The obvious objection to this narrative—that global south countries could reject Chinese AI on security grounds—appears shaky. Global south governments did not reject Chinese ports, railways, or power plants on security grounds under BRI. With no cheap Western AI offerings on the horizon (U.S. AI firms are cash-strapped and thus unlikely to cut prices), there is even less reason to expect a different pattern this time. Besides, the Chinese AI offer is harder to refuse. It helps that open-source models have no single vendor; they are distributed quietly through public websites, not via state-to-state contracts that can grab headlines.

Skeptics could also flag that Beijing’s models come with Chinese Communist Party censorship baked in. Even when it is self-hosted, China’s DeepSeek sometimes refuses to answer sensitive questions. But such censorship has been squarely aimed at China-sensitive topics such as Taiwan, Tiananmen, Tibet, and Xinjiang, making it mostly irrelevant to local needs in the global south. Yes, Beijing could always threaten to pull the plug on its AI models, for instance by tightening access to updates in countries pursuing relations with Taiwan. Yet as Washington has become erratic and unpredictable under U.S. President Donald Trump, global south leaders have plenty of reason to think that relying on Chinese AI could be safer in the long term than dependency on U.S. technology.

The real AI competition may not be a hardware arms race won by access to the most cutting-edge chips. Instead, the global AI race may well be a contest to decide which models and standards become the default infrastructure in countries that remain up for grabs. China does not need to dominate the most advanced models to win the AI race. If Chinese models become the affordable, good-enough default across emerging markets, Beijing will have built durable influence for decades.

Instead of fixating on the AI horse race between ChatGPT, Claude, and Gemini, Western policymakers may want to ponder how they can make an AI offer of their own to the global south before the defaults are set. Meanwhile, developers across Asia, Africa, and Latin America will continue to fine-tune their Kimi, Qwen, and DeepSeek agents.

Agathe Demarais is a columnist at Foreign Policy, a senior policy fellow on geoeconomics and technology at the European Council on Foreign Relations, a visiting professor at the College of Europe, a former global forecasting director at the Economist Intelligence Unit, and the author of Backfire: How Sanctions Reshape the World Against U.S. Interests. Bluesky: @agathedemarais.com

Commenting is a benefit of a Foreign Policy subscription.

Already a subscriber? Log In.

Join the conversation on this and other recent Foreign Policy articles when you subscribe now.

Please follow our comment guidelines, stay on topic, and be civil, courteous, and respectful of others’ beliefs.

I agree to abide by FP’s comment guidelines. (Required)

The default username below has been generated using the first name and last initial on your FP subscriber account. Usernames may be updated at any time and must not contain inappropriate or offensive language.

I agree to abide by FP’s comment guidelines. (Required)

The flow of global capital is hitting new walls.

Kaynak: Foreign PolicyOrijinal Habere Git →

İlgili Haberler