夜雨聆风The Economist / Shanghai
China's AI investments lag far behind America's. Its models do not
中国AI投资远逊美国，模型却并不落后
美国科技巨头在人工智能上豪掷千金，中国企业却出奇节俭。尽管如此，中国模型的性能却接近美国前沿水平。本文分析这种反差背后的成本、技术、制裁与战略因素。
· 01
HOW MUCH money should companies lavish on artificial intelligence? American technology giants think the answer is: a lot. Their capital spending on data centres may surpass $740bn this year, according to Bloomberg Intelligence, a research firm. Investors, who also used to think that more is more, are having second thoughts. When on July 26th it emerged that Nvidia was in talks to underwrite $250bn of a gargantuan $500bn-plus data centre, stuffed full of its chips and run by OpenAI, a leading model-maker, its share price tumbled. Days earlier investors blanched when Alphabet, Google's parent company, said it would boost its AI spending to $205bn this year. On August 5th SpaceX's stock took a hit after the rocketry firm announced huge AI investments (see Business section).
企业在人工智能上该砸多少钱？美国科技巨头认为答案是：越多越好。据研究机构彭博智库估算，它们今年在数据中心的资本支出可能超过7400亿美元。过去信奉"多多益善"的投资者如今也动摇了。7月26日，英伟达被曝正洽谈为一座由OpenAI运营、塞满其芯片、规模超过5000亿美元的巨型数据中心承保2500亿美元，消息一出其股价应声下跌。几天前，当谷歌母公司Alphabet宣布将今年AI支出提高至2050亿美元时，投资者也脸色发白。8月5日，火箭公司SpaceX宣布巨额AI投资后，其股票同样遭受重创（详见商业版块）。
重点词汇与表达 · 5 个
· 02
The American AI splurge looks especially profligate compared with Chinese parsimony. In 2026 China's technology titans are forecast to invest less than a tenth as much in data centres as their American counterparts (see chart). All this frugality notwithstanding, their AI models currently appear only fractionally less powerful. K3, an advanced model launched last month by Moonshot AI, a Beijing-based startup, is 95% as clever (on widely used benchmarks) as Fable 5, a frontier model from Anthropic, another top American lab. On August 3rd Alibaba, China's answer to Amazon, released a model which reportedly scored among the world's best by some measures.
与中国企业的节俭相比，美国的AI开支显得尤为挥霍。据预测，2026年中国科技巨头在数据中心的投入将不足美国同行的十分之一（见图）。尽管如此节俭，它们的AI模型目前看来只比美国模型稍逊一筹。总部位于北京的初创公司月之暗面上月推出的先进模型K3，在常用基准测试中达到了Anthropic旗下前沿模型Fable 5的95%水平。8月3日，被誉为"中国亚马逊"的阿里巴巴发布了一款模型，据称在某些指标上跻身世界前列。
重点词汇与表达 · 5 个
· 03
The cost of land (for data centres), some equipment (to fill these with AI servers) and labour (from construction workers to AI whizz-kids) is lower in China than in America. Chinese firms are also accused of training their models using the outputs of expensive American ones. Such "distillation", as the process is known, lets China's AI developers free-ride on some of the American spending.
在中国，数据中心用地、部分设备（用于装入AI服务器）以及劳动力（从建筑工人到AI奇才）的成本都低于美国。中国企业还被指利用昂贵的美国模型输出来训练自己的模型。这种被称为"蒸馏"的技术，让中国AI开发者得以在美国投入上"搭便车"。
重点词汇与表达 · 3 个
· 04
Moreover, a portion of Chinese AI bill may not be captured in the headline capital-expenditure figures. Part of the sums that Chinese AI firms, especially smaller ones, might have spent on Nvidia's best chips may instead have gone towards squeezing better results from inferior semiconductors. DeepSeek, which gave American AI darlings and investors a jump scare by launching a surprisingly powerful and efficient model in early 2025, is probably ploughing a fair bit into its techniques that reduce demand for computing power.
此外，中国AI支出的部分金额可能并未体现在 headline 资本支出数据中。中国AI公司——尤其是较小的公司——本可能用于购买英伟达顶级芯片的资金，或许被转而用于从性能较差的半导体中压榨出更好的结果。2025年初，DeepSeek推出了一款令人惊讶的强大而高效的模型，让美国AI宠儿和投资者吓了一跳，它很可能在大量投入能降低算力需求的技术研发。
重点词汇与表达 · 4 个
· 05
Yet such factors alone cannot explain why China is investing so much less than America for what seem to be comparable AI results. So what does?
然而，仅凭这些因素还不足以解释：为何在AI成果看似可与美国媲美的同时，中国的投资却远少于美国。那么真正的原因是什么呢？
重点词汇与表达 · 1 个
· 06
One obvious answer for lower investments would be a shortage of capital. Chinese firms are, though, flush with it. The bigger check on companies' AI outlays has been a dearth of things on which to spend money. American restrictions on technology exports to China mean that Chinese firms cannot buy the best—and priciest—AI chips, which are designed by Nvidia, an American firm, and manufactured by TSMC, a Taiwanese contract manufacturer (or foundry) with close links to America.
投资额较低的一个明显答案可能是资本短缺。但中国公司其实并不缺钱。真正限制企业AI支出的，是"无物可买"。美国对华技术出口限制意味着中国企业无法购买最顶尖、最昂贵的AI芯片——这些芯片由美国公司英伟达设计，由与其关系密切的台湾代工制造商台积电生产。
重点词汇与表达 · 4 个
· 07
Unofficially, regulators in Beijing have gone so far as to discourage the use of even second-rate Nvidia chips, which Chinese companies can continue to procure for now. At times they have blocked the import of these, in order both to sanction-proof China's AI infrastructure and to promote homemade alternatives.
据非官方消息，北京监管机构甚至不鼓励企业使用目前仍可采购的二流英伟达芯片。它们有时还会阻止这些芯片进口，既为了让中国AI基础设施免受制裁冲击，也为了推动国产替代方案。
重点词汇与表达 · 2 个
· 08
Consider Alibaba. Last year it said that it would invest $53bn in AI between 2026 and 2028. The planned spending is chump change relative to that of Amazon or Alphabet. Even so, it may be out of reach. Alibaba, which boasts one of China's most advanced chip-design programmes, used to outsource manufacturing to TSMC. American sanctions have forced it to find domestic suppliers instead. The only Chinese purveyors capable of producing comparably powerful chips are Huawei, a technology behemoth blacklisted by America since 2019, and SMIC, a state-controlled foundry, which manufactures Huawei's designs. However, American restrictions extend to sophisticated chipmaking gear, forcing Huawei and SMIC to come up with costly workarounds that limit how many sufficiently powerful chips the duo can churn out. Moreover, SMIC cannot offset this inefficiency by ramping up capital spending because the same sanctions also constrain how much less sophisticated chipmaking kit it can get its hands on.
以阿里巴巴为例。去年它宣布将在2026至2028年间向AI投入530亿美元。与亚马逊或Alphabet相比，这笔计划支出只是九牛一毛。即便如此，它可能仍难以实现。阿里巴巴拥有中国最先进的芯片设计项目之一，过去一直将制造外包给台积电。美国制裁迫使它转向国内供应商。而能够生产同等强大芯片的中国供应商只有两家：自2019年起被美国列入黑名单的科技巨头华为，以及制造华为设计的国有代工厂中芯国际。然而，美国限制还延伸至先进的芯片制造设备，迫使华为和中芯国际采用代价高昂的变通方案，限制了它们能大量产出的高性能芯片数量。此外，中芯国际无法通过增加资本开支来弥补这种低效率，因为同样的制裁也限制了它能获得的低端芯片制造设备数量。
重点词汇与表达 · 8 个
· 09
Moreover, the Communist Party's AI ambitions are more about diffusing the technology through the economy than creating ever cleverer systems that require ever more powerful semiconductors in ever more data centres. Dozens of American AI companies are thought to be trying to develop artificial general intelligence, which could outdo humans in most intellectual tasks. In China, by contrast, fewer than ten companies are known to be working on such AGI, notes Xu Qi, the head of the Communist Party committee at the Shanghai AI Association.
此外，中共的AI雄心更多是关于将技术扩散到整个经济，而非在越来越多数据中心中打造需要越来越强半导体的越来越聪明的系统。据认为，数十家美国AI公司正试图开发能在大多数智力任务上超越人类的人工通用智能（AGI）。相比之下，据上海人工智能协会党委书记徐琪指出，中国已知从事此类AGI研发的公司不足十家。
重点词汇与表达 · 3 个
· 10
Another signal to keep AI spending in check comes from investors familiar with all these constraints. Chinese ones have long been as wary of overspending on AI as those in America are apparently now becoming. Whereas America's tech giants had until recently been rewarded with higher share prices for aggressive AI-spending plans, Chinese ones were likelier to be punished for AI profligacy.
另一个抑制AI支出的信号来自熟悉上述种种约束的投资者。中国投资者长期以来对AI过度支出保持警惕，而美国投资者似乎现在才开始变得谨慎。直到最近，美国科技巨头因激进的AI支出计划而获得股价上涨奖励，而中国公司则更有可能因AI挥霍而受到惩罚。
重点词汇与表达 · 3 个
· 11
Investors may currently be entertaining the idea that America has too many data centres. But China, for all its admirable restraint, may not have enough to take full advantage of its AI stars' innovations—or for those stars to cash in. ByteDance, creator of the world's top-ranked video-making AI (and owner of TikTok), suffers from intense computing bottlenecks that mean some clips take ten hours to process. Certain services from Zhipu AI, another Beijing lab, and Alibaba's cloud-computing unit have to be strictly metered and sell out in minutes. K3 has a long waiting list of hopeful users. Diffusion of practical AI is less hungry for computing power than the quest for superintelligence. An overly restrictive diet can still stunt growth.
投资者目前或许正倾向于认为美国的数据中心已经过剩。但中国尽管克制可圈可点，数据中心数量却可能不足以让其AI明星企业的创新充分发挥——也不足以让这些企业变现。字节跳动旗下拥有全球排名第一的视频生成AI（同时拥有TikTok），却受困于严重的算力瓶颈，导致某些片段需要十小时才能处理完成。智谱AI（另一家北京实验室）和阿里巴巴云计算部门的某些服务必须严格限量，几分钟内即售罄。K3也积累了大量期待用户的等待名单。实用AI的扩散对算力的渴求低于对超级智能的追求。但过度节制的"饮食"仍会阻碍成长。
重点词汇与表达 · 5 个
SELF-CHECK READING · 不看译文，连续读完全文
HOW MUCH money should companies lavish on artificial intelligence? American technology giants think the answer is: a lot. Their capital spending on data centres may surpass $740bn this year, according to Bloomberg Intelligence, a research firm. Investors, who also used to think that more is more, are having second thoughts. When on July 26th it emerged that Nvidia was in talks to underwrite $250bn of a gargantuan $500bn-plus data centre, stuffed full of its chips and run by OpenAI, a leading model-maker, its share price tumbled. Days earlier investors blanched when Alphabet, Google's parent company, said it would boost its AI spending to $205bn this year. On August 5th SpaceX's stock took a hit after the rocketry firm announced huge AI investments (see Business section).
The American AI splurge looks especially profligate compared with Chinese parsimony. In 2026 China's technology titans are forecast to invest less than a tenth as much in data centres as their American counterparts (see chart). All this frugality notwithstanding, their AI models currently appear only fractionally less powerful. K3, an advanced model launched last month by Moonshot AI, a Beijing-based startup, is 95% as clever (on widely used benchmarks) as Fable 5, a frontier model from Anthropic, another top American lab. On August 3rd Alibaba, China's answer to Amazon, released a model which reportedly scored among the world's best by some measures.
The cost of land (for data centres), some equipment (to fill these with AI servers) and labour (from construction workers to AI whizz-kids) is lower in China than in America. Chinese firms are also accused of training their models using the outputs of expensive American ones. Such "distillation", as the process is known, lets China's AI developers free-ride on some of the American spending.
Moreover, a portion of Chinese AI bill may not be captured in the headline capital-expenditure figures. Part of the sums that Chinese AI firms, especially smaller ones, might have spent on Nvidia's best chips may instead have gone towards squeezing better results from inferior semiconductors. DeepSeek, which gave American AI darlings and investors a jump scare by launching a surprisingly powerful and efficient model in early 2025, is probably ploughing a fair bit into its techniques that reduce demand for computing power.
Yet such factors alone cannot explain why China is investing so much less than America for what seem to be comparable AI results. So what does?
One obvious answer for lower investments would be a shortage of capital. Chinese firms are, though, flush with it. The bigger check on companies' AI outlays has been a dearth of things on which to spend money. American restrictions on technology exports to China mean that Chinese firms cannot buy the best—and priciest—AI chips, which are designed by Nvidia, an American firm, and manufactured by TSMC, a Taiwanese contract manufacturer (or foundry) with close links to America.
Unofficially, regulators in Beijing have gone so far as to discourage the use of even second-rate Nvidia chips, which Chinese companies can continue to procure for now. At times they have blocked the import of these, in order both to sanction-proof China's AI infrastructure and to promote homemade alternatives.
Consider Alibaba. Last year it said that it would invest $53bn in AI between 2026 and 2028. The planned spending is chump change relative to that of Amazon or Alphabet. Even so, it may be out of reach. Alibaba, which boasts one of China's most advanced chip-design programmes, used to outsource manufacturing to TSMC. American sanctions have forced it to find domestic suppliers instead. The only Chinese purveyors capable of producing comparably powerful chips are Huawei, a technology behemoth blacklisted by America since 2019, and SMIC, a state-controlled foundry, which manufactures Huawei's designs. However, American restrictions extend to sophisticated chipmaking gear, forcing Huawei and SMIC to come up with costly workarounds that limit how many sufficiently powerful chips the duo can churn out. Moreover, SMIC cannot offset this inefficiency by ramping up capital spending because the same sanctions also constrain how much less sophisticated chipmaking kit it can get its hands on.
Moreover, the Communist Party's AI ambitions are more about diffusing the technology through the economy than creating ever cleverer systems that require ever more powerful semiconductors in ever more data centres. Dozens of American AI companies are thought to be trying to develop artificial general intelligence, which could outdo humans in most intellectual tasks. In China, by contrast, fewer than ten companies are known to be working on such AGI, notes Xu Qi, the head of the Communist Party committee at the Shanghai AI Association.
Another signal to keep AI spending in check comes from investors familiar with all these constraints. Chinese ones have long been as wary of overspending on AI as those in America are apparently now becoming. Whereas America's tech giants had until recently been rewarded with higher share prices for aggressive AI-spending plans, Chinese ones were likelier to be punished for AI profligacy.
Investors may currently be entertaining the idea that America has too many data centres. But China, for all its admirable restraint, may not have enough to take full advantage of its AI stars' innovations—or for those stars to cash in. ByteDance, creator of the world's top-ranked video-making AI (and owner of TikTok), suffers from intense computing bottlenecks that mean some clips take ten hours to process. Certain services from Zhipu AI, another Beijing lab, and Alibaba's cloud-computing unit have to be strictly metered and sell out in minutes. K3 has a long waiting list of hopeful users. Diffusion of practical AI is less hungry for computing power than the quest for superintelligence. An overly restrictive diet can still stunt growth.
读完了。以上为完整双语精读内容，可复制到公众号编辑器直接使用。
