This article is part of Sixth Tone Voices & Opinion’s end-of-year series reviewing how China’s AI, film, game, and VR industries evolved in 2025.

China’s AI industry began 2025 as a follower. But within the space of a year it has outgrown that status, passing milestone after milestone.

Early on, DeepSeek-R1 shook the global AI industry with its low training costs and open-source approach; soon after, the Chinese government showed support for using AI in industrial and other robots; and by year’s end, large-scale AI model companies Zhipu AI and MiniMax submitted IPO filings, potentially becoming the world’s first such startups to go public.

China was also seen as trying to catch up with Silicon Valley while it dealt with American export restrictions on the most advanced chips. But one year later, the country’s AI industry has shown that it can achieve breakthroughs on its own terms, and this new trajectory promises an exciting future.

China’s Sputnik moment

On Jan. 20, the release of DeepSeek-R1 caught everyone by surprise. A Chinese team had achieved performance levels close to that of top-tier reasoning models at a cost far below industry expectations.

In China, DeepSeek’s sudden rise pushed generative AI beyond tech circles, turning it into an everyday tool for ordinary users.

But its global impact was even more profound. It was described as China’s Sputnik moment, comparable to how the first Soviet satellite shocked the U.S. in 1957. The inclusion of DeepSeek founder Liang Wenfeng in scientific journal Nature’s annual “Nature’s 10” list further underscored the significance of this event.

DeepSeek was proof that reaching frontier capabilities does not rely solely on piling up computing power, challenging the assumption that whoever has the most advanced chips will be in the lead by default. And at a time when closed-source models from companies like OpenAI and Anthropic dominated the field, DeepSeek’s open-source approach ignited intense debate in Silicon Valley.

It has also changed the global market. According to data from multiple app-tracking platforms, more than half of DeepSeek users are located in China, India, and Indonesia. In such price-sensitive markets, open-source and cost-effective models have quickly become developers’ top choice, allowing local entrepreneurs to bypass expensive application programming interfaces (APIs) and build applications tailored to their own needs.

Surging investments, with a big backer

After DeepSeek, a wave of AI investment surged across China. It pushed the total valuation of the AI industry to a projected 1.2 trillion yuan ($171 billion) this year, an increase of 300 billion yuan compared to 2024.

Crucially, AI has the full support of the Chinese state. In March, the Chinese government’s yearly work report mentioned embodied intelligence — using AI in robots and other machinery — for the first time. An increasing number of Chinese cities are permitting self-driving car and delivery van companies to use their roads.

The Chinese state is also providing investment. For example, filings from Zhipu AI show that, ahead of its IPO, it received roughly 3 billion yuan in investment from entities linked to the governments of major cities such as Beijing, Hangzhou, Zhuhai, Chengdu, and Shanghai. This accounts for more than one-third of the 8.3 billion yuan it has raised to date.

These investments also reflect local governments’ plans to anchor AI development within their regions. Hangzhou, Zhuhai, and Chengdu have announced initiatives to partner with Zhipu AI to build localized AI industry ecosystems, aligning regional industrial resources more closely with cutting-edge AI technologies. This trend is also reshaping how success is measured for large-model companies: beyond technical breakthroughs and user growth, their ability to support industrial digital transformation is becoming an increasingly important benchmark.

The varied applications of AI

A pursuit of both the American and Chinese AI industries this year was to have models take on more advanced tasks than generating content.

Companies on both sides of the Pacific pursued AI agents — allowing users to give a model tasks that it will then carry out, such as making online purchases. Chinese tech giants such as ByteDance, Alibaba, Baidu, and Tencent are all rolling out agent-based products.

Embodied intelligence has also emerged as a popular sector for investment, accounting for over 20 billion yuan in the first 10 months of the year — one-third of the total investment in AI industries. Despite this momentum, this approach still faces major challenges, including limited application scenarios and unresolved core technological bottlenecks.

As in the U.S., Chinese companies such as Alibaba, ByteDance, and Tencent invested heavily in AI infrastructure such as data centers and computing networks. These are not simply bets on raw computing power, but strategic positioning for the “AI as a service” (AIaaS) business model: whoever controls the infrastructure can offer the best AI services and create the richest AI ecosystem.

In the meantime, while Silicon Valley pushes the limits of core technologies in pursuit of reaching AGI — artificial general intelligence, models as capable and smart as humans — China focuses more on understanding users, integrating scenarios, and building service systems.

For example, open-source models like DeepSeek, Qwen, and Doubao provide a complete toolchain — from training and fine-tuning to deployment — allowing small- and medium-sized enterprises to build AI applications at a fraction of the costs they would face when using American competitors.

Through deep customization based on open-source models, the health care, legal, and education sectors will produce truly functional, professional AI models. AI capabilities are also increasingly being built into devices like smartphones, vehicles, and home appliances. Combined with China’s strengths in manufacturing and supply chains, this trend could give rise to an entirely new generation of consumer electronics.

AI as cultural technology

Which language dominates AI training data has long been a power struggle hidden behind a technical narrative: whichever language is prioritized in training holds more influence in the AI era.

The past year saw rapid development for large multilingual models in China — a natural focus given the country’s linguistic diversity. Apart from Mandarin and English, there are also dialects and minority languages, as well as the frequent mixing of different languages.

When Chinese models entered the global market, this complexity proved an advantage. For example, Singapore’s Southeast Asian multilingual model, Qwen-SEA-LION-v4, was built on Alibaba’s open-source Qwen model, which supports an industry-leading 119 different languages. AI Singapore contributed its regional expertise, providing over 100 billion tokens covering Burmese, Indonesian, Thai, and other languages to specifically retrain and optimize the model.

This collaboration aims to solve a long-standing hurdle for AI in Southeast Asia. The region has over 1,200 languages and a high occurrence of code-switching — mixing languages in daily life. Since most global models are English-centric, they struggle to serve such markets, exacerbating what is known as the AI divide. Dominant languages such as English surge ahead, while languages with fewer resources languish.

This shift is also reshaping how Chinese AI companies go global. Unlike earlier waves driven by platform dominance or low pricing, Chinese firms in the Middle East, Southeast Asia, and Latin America are no longer just touting their model’s raw performance parameters. Instead, they are emphasizing local language support, adaptation to local education systems, and understanding of cultural nuances in content generation. AI is no longer just about translating Chinese products; it is helping Chinese companies learn how to be understood by different cultures.

Looking back at 2025, while the development of the AI industry is almost inseparable from the context of U.S.-China competition, the future of technology is not written by a single laboratory or a single nation. It is a cultural exploration shaped collectively by countless developers, enterprises, and users worldwide. This is perhaps the most significant asset the industry has accumulated over the past year.

(Header image and icons: Visuals from FStop, Shijue, and Vectorstock/VCG, reedited by Sixth Tone)