China vs US: Who is winning the AI race, in four charts

The US dominates AI spending and computing power, but China is ahead in research and closing the gap on advanced models.

A worker reacts near a line of humanoid robots displayed at the World AI Conference in Shanghai, Friday, July 17, 2026 [Ng Han Guan/AP Photo]

United States President Donald Trump and China’s President Xi Jinping are set to meet in Washington, DC on Thursday for a summit that is expected to cover trade, artificial intelligence (AI), Taiwan and the US-Israel war on Iran.

In advance of the talks, US Treasury Secretary Scott Bessent said Washington had proposed an AI “ notification mechanism ” with China – effectively a hotline to alert each other when AI incidents threaten national security.

The meeting comes as leading AI companies warn about the technology’s risks, while both superpowers push to accelerate their AI race. In this visual explainer, Al Jazeera compares the US and China’s AI strength in computing power, AI models, spending and research.

Who has more computing power?

Training and running AI requires enormous amounts of computing power. The more powerful chips and computers a country has, the faster it can build and improve AI models, making computing power one of the biggest advantages in the AI race.

Computing power is measured in FLOP/s (floating-point operations per second) – or how many calculations a computer can perform in one second.

Adding up the computing power of all the AI chips a country has gives a sense of its total AI capacity. By that measure, the US is well ahead, accounting for nearly three-quarters of the world’s total AI computing power, while China holds just more than 14 percent, according to Epoch AI, an AI research institute.

The US lead comes largely from its access to the most advanced chips. US company Nvidia accounts for more than 60 percent of global AI computing capacity among major chip designers, while China’s Huawei holds a much smaller, but growing share, according to Stanford University’s 2026 AI Index.

Those chips also need somewhere to run. The US has more than 5,400 data centres – the large buildings which house thousands of computers – about 10 times as many as any other country. It also leads in dedicated AI data centres, with 84, more than the next eight countries combined.

Whose AI models are people using?

Frontier models are the most advanced AI systems, such as OpenAI’s GPT, Anthropic’s Claude and DeepSeek’s models in China. They are trained using vast amounts of data and computing power to write, reason and code.

The US has led the development of these models, but Chinese models are gaining ground.

As of March 2026, US and Chinese AI models were closely matched on Arena, a leaderboard where users compare two anonymous models and vote for the better answer. US companies Anthropic, xAI, Google and OpenAI ranked near the top, alongside China’s Alibaba and DeepSeek.

On OpenRouter, a platform that ranks AI models by real-world use, Chinese models dominate the leaderboard. Models from Chinese companies DeepSeek, Z.ai and Tencent hold the top three spots on OpenRouter by tokens processed, the units of text that AI models read and generate.

This is partly because Chinese models are cheaper to use. Many are also open-weight, meaning developers can download, adapt and run them themselves.

Chinese models can now compete with US ones in many real-world tasks, with Chinese AI models “months, not years, behind US frontier models”, according to a July analysis by the Centre for Strategic and International Studies ( CSIS ).

In May, the US government’s Center for AI Standards and Innovation (CAISI) estimated that DeepSeek V4 Pro, released in April, was about eight months behind leading US models.

Who is spending more?

The US is spending far more than China on the infrastructure needed to build AI, giving its tech companies a significant advantage in the race for computing power.

Goldman Sachs estimates that US hyperscalers, the giant tech companies that run large-scale cloud computing and data centres, will spend about $764bn in 2026. That includes Amazon, Microsoft, Google, Meta and Oracle. By comparison, China’s Alibaba, Tencent, Baidu and ByteDance are expected to spend $102bn.

But China’s spending is increasing faster. TrendForce , a market research firm, expects China’s four hyperscalers’ capital expenditure to rise by more than 80 percent in 2026, compared with a projected growth of 76 percent for US hyperscalers, according to the news agency Reuters .

The spending gap reflects the scale of the US tech sector, whose companies have larger revenues to reinvest in chips, data centres and talent.

Who is leading in AI research and talent?

China is producing more AI research and training more technical talent, but the US remains a magnet for the world’s top AI researchers.

China accounted for more than 27 percent of global AI publications in 2024, including journal articles, conference papers, working papers and preprints compared with 12 percent from the US, according to the Center for Security and Emerging Technology. The figures only cover publications with an English-language title or abstract.

China also trains more top researchers. According to a study by MacroPolo, 47 percent of the world’s top 20 percent of AI researchers completed their undergraduate studies in China in 2022, up from 29 percent in 2019. The study also found that 72 percent of the top AI researchers educated in China were working in the US.

美国在人工智能领域的支出和计算能力方面占据主导地位，但中国在研究方面领先，并在先进模型方面正在缩小差距。

2026年7月17日，星期五，在上海举行的世界人工智能大会上，一名工作人员在一排人形机器人附近做出反应。[Ng Han Guan/AP Photo]

美国总统唐纳德·特朗普和中国国家主席习近平将于周四在华盛顿特区举行峰会，预计峰会将涵盖贸易、人工智能、台湾以及美以对伊朗的战争等议题。

在会谈之前，美国财政部长斯科特·贝森特表示，华盛顿已向中国提议建立人工智能“通报机制”——实际上是一条热线，以便在人工智能事件威胁国家安全时互相发出警报。

此次会议召开之际，领先的人工智能公司纷纷警告该技术的风险，而美中两国都在加速推进人工智能竞赛。半岛电视台在这份图文并茂的解读中，对比了中美两国在计算能力、人工智能模型、投入和研究方面的实力。

谁的计算能力更强？

训练和运行人工智能需要庞大的计算能力。一个国家拥有的芯片和计算机性能越强大，就能越快地构建和改进人工智能模型，这使得计算能力成为人工智能竞赛中最大的优势之一。

计算能力以 FLOP/s（每秒浮点运算次数）来衡量——也就是计算机在一秒钟内可以执行多少次计算。

将一个国家所有人工智能芯片的计算能力相加，就能大致了解该国的人工智能总能力。据人工智能研究机构Epoch AI的数据显示，按此标准衡量，美国遥遥领先，占据了全球近四分之三的人工智能计算能力，而中国仅占14%多一点。

美国领先地位主要源于其对最先进芯片的掌控。斯坦福大学2026年人工智能指数显示，美国公司英伟达在全球主要芯片设计商中占据了超过60%的人工智能计算能力，而中国华为的份额虽然小得多，但正在不断增长。

这些芯片也需要运行的地方。美国拥有超过5400个数据中心——容纳数千台计算机的大型建筑——大约是其他任何国家的10倍。美国在专用人工智能数据中心方面也处于领先地位，拥有84个，比排名第二和第三的八个国家加起来还要多。

人们都在使用谁的AI模型？

前沿模型是目前最先进的人工智能系统，例如 OpenAI 的 GPT、Anthropic 的 Claude 以及中国 DeepSeek 的模型。它们利用海量数据和强大的计算能力进行训练，从而能够编写、推理和编码。

美国引领了这些车型的发展，但中国车型正在迎头赶上。

截至2026年3月，美国和中国的AI模型在Arena排行榜上势均力敌。Arena是一个用户匿名比较两个模型并投票选出更优答案的平台。美国公司Anthropic、xAI、谷歌和OpenAI与中国的阿里巴巴和DeepSeek并驾齐驱，排名靠前。

在OpenRouter这个根据实际应用情况对AI模型进行排名的平台上，中国模型占据了排行榜的主导地位。来自中国公司DeepSeek、Z.ai和腾讯的模型在OpenRouter的“处理词元数”（AI模型读取和生成的文本单元）排名中位列前三。

部分原因是国产模型使用成本更低。许多国产模型也是开源的，这意味着开发者可以自行下载、修改和运行。

根据战略与国际研究中心 (CSIS) 7 月份的一项分析，中国模型现在可以在许多现实世界的任务中与美国模型竞争，中国的 AI 模型“落后美国前沿模型数月，而不是数年”。

今年 5 月，美国政府人工智能标准与创新中心 (CAISI) 估计，4 月份发布的 DeepSeek V4 Pro 比美国领先的型号落后了大约 8 个月。

谁的消费更多？

美国在人工智能基础设施建设方面的投入远远超过中国，这使得美国科技公司在计算能力竞赛中拥有显著优势。

高盛估计，美国超大规模数据中心运营商（即运营大规模云计算和数据中心的科技巨头）到2026年将支出约7640亿美元。这些运营商包括亚马逊、微软、谷歌、Meta和甲骨文。相比之下，中国的阿里巴巴、腾讯、百度和字节跳动预计将支出1020亿美元。

但中国的支出增长速度更快。据路透社报道，市场研究公司TrendForce预计，到2026年，中国四大超大规模数据中心运营商的资本支出将增长超过80%，而美国超大规模数据中心运营商的预计增长率为76%。

支出差距反映了美国科技行业的规模，这些公司拥有更大的收入，可以再投资于芯片、数据中心和人才。

谁在人工智能研究和人才方面处于领先地位？

中国在人工智能研究和技术人才培养方面投入了更多资源，但美国仍然是世界顶尖人工智能研究人员的聚集地。

据安全与新兴技术中心（Center for Security and Emerging Technology）的数据显示，到2024年，中国在全球人工智能出版物（包括期刊文章、会议论文、工作论文和预印本）中所占比例超过27%，而美国仅占12%。该数据仅涵盖标题或摘要为英文的出版物。

中国也培养了更多顶尖研究人员。MacroPolo 的一项研究显示，2022 年全球排名前 20% 的人工智能研究人员中，47% 的人本科毕业于中国，高于 2019 年的 29%。该研究还发现，在中国接受教育的顶尖人工智能研究人员中，72% 的人目前在美国工作。