Six months after DeepSeek’s breakthrough, China speeds on with AI
深度探索突破六个月后，中国加速推进人工智能

The mecca for China’s boom in artificial intelligence is Liangzhu, a leafy suburb of Hangzhou, the tech-heavy capital of Zhejiang province. The Communist Party has long touted Liangzhu’s famous archaeological remains, dating back to 3300BC, as proof of the age of Chinese civilisation. Now Liangzhu, with its myriad AI startups, represents the future. Investors from all over China flock there to meet growing numbers of founders, app engineers and other AI developers and dreamers. It is six months since a barely known AI startup, DeepSeek, caused a huge stir by releasing an impressive open-source model trained for a sliver of the cost of fancier Western ones. Its founder studied at Zhejiang University, a tech mothership not far from Liangzhu. The area is at the heart of an AI ecosystem which China hopes will soon rival America’s.
中国人工智能繁荣的圣地是位于浙江省省会杭州的一个绿树成荫的郊区——良渚。中国共产党长期以来将良渚著名的考古遗址（可追溯至公元前 3300 年）视为中国文明悠久历史的证明。如今，良渚凭借其众多的人工智能初创企业，代表着未来。来自全国各地的投资者纷纷涌向这里，寻找越来越多的创始人、应用工程师以及其他人工智能开发者和梦想家。六个月前，一家鲜为人知的人工智能初创公司 DeepSeek 因发布一款以远低于西方同类产品成本训练的令人印象深刻的开源模型而引起了巨大轰动。其创始人在离良渚不远的浙江大学学习。该地区正处于一个人工智能生态系统的核心，中国希望这个生态系统能很快与美国的相媲美。

The signs are promising. Last month three Chinese labs introduced stellar large-language models that are reckoned to be among the world’s best. AI technologies have “broken through the critical threshold of usefulness”, says one early-stage investor who frequents Liangzhu. He predicts a surge in how AI can be applied. “Once the water boils,” he says, “many people want to build a steam engine.”
迹象令人鼓舞。上个月，三家中国实验室推出了被认为是世界上最优秀的大型语言模型。一个经常光顾良渚的早期投资者表示，人工智能技术已经“突破了实用性的关键门槛”。他预测人工智能的应用将激增。“一旦水开了，”他说，“很多人就想造蒸汽机。”

Sam Hu, at a sunny café in Liangzhu’s main plaza, is one such person. After stints at Tencent, a tech giant, and two ride-hailing firms, last year Mr Hu struck out on his own to develop an AI agent that helps managers make decisions. The moment was right. These days, he explains, “the cost of trial and error is lower.”
在良渚主广场的一家阳光明媚的咖啡馆里，胡先生就是这样一个人。在腾讯这家科技巨头和两家打车公司工作后，胡先生去年选择独立发展，开发一种帮助管理者做决策的人工智能代理。时机正好。他解释说，如今“试错的成本更低”。

The prize for Mr Hu and his peers is enormous. Morgan Stanley, a bank, predicts that China’s AI industry will grow from $3.2bn last year to $140bn by 2030; that figure jumps to $1.4trn when AI-related sectors such as infrastructure and component suppliers are included. In June last year some 8% of Greater China firms surveyed by Gartner, a consultancy, were using generative AI. Less than a year later the figure had leapt to 43%.
胡先生和他的同行们面临着巨大的机会。摩根士丹利预测，中国的人工智能产业将从去年的 32 亿美元增长到 2030 年的 1400 亿美元；如果将基础设施和组件供应商等与人工智能相关的行业纳入计算，这一数字将跃升至 1.4 万亿美元。去年 6 月，咨询公司 Gartner 调查的中国大中华区企业中约有 8%在使用生成式人工智能，而不到一年后，这一数字跃升至 43%。

The question now is how to keep the industry steaming along. Some of the early frenzy around DeepSeek has passed, and plenty of users still grumble about how models can “hallucinate”. But DeepSeek’s breakthrough has helped shift China’s approach to AI in profound ways. It has lowered costs and moved the emphasis away from cutting-edge development and cut-throat competition among developers of AI models towards explorations of how AI can be applied across business, industry, the public sector and society itself. The importance of AI for China—and indeed the way it can outdo America, its promoters increasingly argue—is through its adoption, adaptation and diffusion, that is, spreading the use of AI more broadly.
现在的问题是如何让行业持续发展。围绕 DeepSeek 的早期热潮已经过去，许多用户仍然抱怨模型可能会“幻觉”。但 DeepSeek 的突破在深刻改变中国对人工智能的看法方面发挥了重要作用。它降低了成本，并将重点从尖端开发和开发者之间的激烈竞争转向探索人工智能如何在商业、工业、公共部门和社会本身中应用。越来越多的推广者认为，人工智能对中国的重要性——以及它如何超越美国——在于其采用、适应和扩散，即更广泛地传播人工智能的使用。

It helps greatly that China’s leadership believes the same thing, and is offering state backing. President Xi Jinping’s handshake with DeepSeek’s founder, Liang Wenfeng, broadcast earlier this year into the country’s living rooms, helped transform how ordinary Chinese view AI. Grandparents were suddenly keen to try out chatbots. The early frenzy was all a bit over the top, says Louis Dong, who developed AI courses for schoolchildren after parents bombarded the education company he works for with requests. Now, Mr Dong says, the focus is on identifying applications for specific industries and putting AI capabilities to pragmatic use. Rapid adoption is key, many experts argue, and DeepSeek’s breakthrough, along with advances in other models, is enabling it.
中国领导层的共识和国家支持极大地推动了这一进程。今年早些时候，习近平主席与 DeepSeek 创始人梁文峰的握手被直播到全国各地的客厅，改变了普通中国人对人工智能的看法。祖父母们突然对尝试聊天机器人产生了浓厚的兴趣。开发了针对学童的人工智能课程的刘东表示，早期的热潮有些过于夸张，因为家长们向他所在的教育公司提出了大量请求。现在，刘东表示，重点是识别特定行业的应用，并将人工智能能力付诸实际使用。许多专家认为，快速采用是关键，而 DeepSeek 的突破以及其他模型的进展正在推动这一进程。

Companies appear to be adopting a more can-do attitude by experimenting with the technology. There is still room for improvement. Accenture, a consultancy, found that 46% of Chinese firms have broadly integrated generative AI, but only 9% saw real benefit in productivity or profit growth. It is early days, AI executives insist. Jiang Xinghua, chief technology officer at Yizhi Intelligence, which develops AI customer-service agents for chemists and furniture stores and AI livestreamers for beauty brands, says ever more clients want to adopt AI-enabled products, and are even tolerant of them making minor mistakes.
公司似乎正在通过实验技术来采取更积极的态度，但仍有改进的空间。咨询公司埃森哲发现，46%的中国企业已广泛整合生成式人工智能，但只有 9%的企业在生产力或利润增长方面看到了实际收益。人工智能高管们坚称，这仍然是早期阶段。易智智能的首席技术官姜星华表示，越来越多的客户希望采用人工智能产品，甚至对其出现小错误持宽容态度。

A big question is how much AI products can improve. A lot, say China’s techno-optimists. Chinese efforts to train new models have been complicated by America’s ban on exports of its most advanced AI chips. In April President Donald Trump’s administration banned shipments to China of Nvidia’s H20 chip, which is deliberately hobbled to make it less effective for training (though it is well suited to “inference”—the process of actually running AI models for customers). In July America reversed course. Yet the political uncertainty and risk remains.
一个大问题是人工智能产品能提高多少。中国的技术乐观主义者表示，提升幅度很大。中国在训练新模型的努力受到美国禁止出口其最先进人工智能芯片的影响。四月，特朗普总统的政府禁止向中国发运英伟达的 H20 芯片，该芯片故意被限制以降低其训练效果（尽管它非常适合“推理”——即实际为客户运行人工智能模型的过程）。七月，美国改变了立场。然而，政治不确定性和风险依然存在。

Strikingly, however, China’s AI development seems to sail on through the turbulence. At China’s biggest AI conference to date, in Shanghai last month, Yan Junjie, the boss of MiniMax, a leading Chinese model-maker, declared that “the inference costs for the best models have fallen by an order of magnitude.” He predicts further sharp falls in the next year or two, driven by innovation and fierce domestic competition.
然而，令人瞩目的是，中国的人工智能发展似乎在动荡中稳步前行。在上个月于上海举行的中国最大人工智能会议上，领先的中国模型制造商 MiniMax 的老板闫俊杰宣称，“最佳模型的推理成本已经下降了一个数量级。”他预测，未来一两年内，随着创新和激烈的国内竞争，成本将进一步大幅下降。
Mr Jiang of Yizhi Intelligence says that although Chinese AI development is frustrated by being denied the best foreign models (such as America’s GPT and Gemini), domestic open-source models are improving rapidly.
易智智能的姜先生表示，尽管中国的人工智能发展受到无法获得最佳外国模型（如美国的 GPT 和 Gemini）的限制，但国内的开源模型正在迅速改善。

Meanwhile, the state is playing a big role, not only by supporting AI development, but also in creating demand. In April, when the leadership convened a study session on AI, Mr Xi talked about making the most of China’s juguo tizhi youshi—its systematic, state-led advantage in mobilising the whole country.
与此同时，国家在其中发挥了重要作用，不仅支持人工智能发展，还创造了需求。今年四月，当领导层召开人工智能学习会议时，习近平提到要充分利用中国的“举国体制优势”——即国家主导下动员全社会的系统性优势。

Lower levels of government, used to reading cues from the top and keen to find new areas of economic growth, have embraced the message, sometimes rather keenly. When Rokid, a Hangzhou startup making AI-equipped glasses, submitted an application to the district-level government for a subsidy, 3m yuan ($420,000) appeared in its account within eight minutes.
地方政府习惯于从上级获取信息，并渴望寻找新的经济增长领域，因此积极响应这一信息，有时甚至表现得相当热切。当杭州初创公司 Rokid 申请区级政府的补贴时，300 万元（42 万美元）在八分钟内便出现在其账户中。
Across China, cities are rushing to offer subsidies for everything from housing to computing power to tailoring open-source models to suit business needs. The state is also an important customer. State museums offer Rokid glasses for tours, while state power workers use them to identify faults along transmission lines.
在中国，各城市纷纷推出补贴政策，涵盖从住房到计算能力，再到调整开源模型以满足商业需求的各个方面。国家也是一个重要的客户。国家博物馆为游客提供 Rokid 眼镜，而国家电力工作人员则利用这些眼镜识别输电线路上的故障。

Meanwhile, provincial and municipal governments have rushed to use DeepSeek to improve hotlines, analyse data and interpret policies for residents.
与此同时，省市政府迅速利用 DeepSeek 来改善热线服务、分析数据和解读居民政策。
Shen Yang, a professor at Tsinghua University, says he has consulted for twice as many government units across China this year compared with last year. They all want to know how to use AI. One example he cites is Beijing’s Haidian district, well-known for its tech universities and businesses, which has deployed AI software to help teachers prepare lessons. Another is Jiangxi province, which wants to integrate AI into its rare-earth and ceramics industries.
清华大学教授沈阳表示，今年他为全国各地的政府单位提供咨询的数量是去年的两倍。他们都想知道如何使用人工智能。他提到的一个例子是北京的海淀区，以其科技大学和企业而闻名，已部署人工智能软件帮助教师备课。另一个例子是江西省，想将人工智能整合到稀土和陶瓷产业中。

Plenty of risks come with the state and investors rushing in. Analysts at Jefferies, an investment bank, identify enthusiasm for humanoid robots as a bubble in the making. Humanoid robots are part of an “embodied AI” push written into the central government’s work report this year. Local governments are falling over themselves to back them. Yet it may be years before the humanoids can be put to profitable use—if they ever can.
随着国家和投资者的涌入，风险也随之而来。投资银行杰富瑞的分析师指出，对类人机器人热情的高涨可能是一个正在形成的泡沫。类人机器人是今年中央政府工作报告中提到的“具身人工智能”推动的一部分。地方政府纷纷争相支持这一领域。然而，类人机器人能否在盈利上得到应用，可能还需要数年时间——如果它们能够做到的话。

Other applications of AI might carry greater risks than excess capacity. In April Tsinghua researchers asked if the more than 300 Chinese hospitals already using DeepSeek were moving “too fast, too soon” and were in danger of making diagnoses based on false outputs.
人工智能的其他应用可能带来的风险超过了过剩产能。今年四月，清华大学的研究人员询问，已经在使用 DeepSeek 的 300 多家中国医院是否“过快、过早”地推进，是否有可能基于错误的输出做出诊断。
Last month another group of researchers identified “an urgent need” to ensure that AI was being safely and responsibly deployed at hospitals. The central leadership is growing more aware of the possible pitfalls. Last month Mr Xi named AI, along with computing power and electric vehicles, as areas shown excessive attention by local governments. “Do all provinces have to develop industries in these directions?” The answer is surely “no”. But for now, their involvement is another reason for thinking that China’s AI juggernaut will keep rolling. ■
上个月，另一组研究人员指出“迫切需要”确保人工智能在医院的安全和负责任的部署。中央领导层对可能的陷阱越来越关注。上个月，习近平主席将人工智能与计算能力和电动车辆列为地方政府过度关注的领域。“所有省份都必须在这些方向上发展产业吗？”答案肯定是“不是”。但目前，他们的参与是另一个让人认为中国的人工智能巨轮将继续前行的理由。■

Subscribers can sign up to Drum Tower, our new weekly newsletter, to understand what the world makes of China—and what China makes of the world.
订阅者可以注册我们的新周刊《鼓楼》，以了解世界对中国的看法，以及中国对世界的看法。