How many Tokens do you consume each month? Token is the smallest unit of text processed by AI language models. Since AI services usually charge based on the number of Tokens processed, Token has also become a common unit of measurement for model usage, cost, and context length.
With the large-scale application of AI, the "Token economy" has become a buzzword. China is already a major consumer of Tokens. By March this year, the national daily call volume exceeded 140 trillion, increasing by more than 40% in three months. It's easy to imagine a scene from this number: a city's call volume growing rapidly, once-idle computing power centers running at full capacity, companies integrating intelligent agents into customer service, R&D, sales, and office processes, and employees massively using AI to assist in their work... These things are indeed happening. Tokens, like GDP, can be counted, ranked, and written into year-end reports. But this picture doesn't answer a question: What does this mean for the economy and society? What can ordinary citizens gain from it?
We can understand this from the perspective of supply and demand. Encompass Tokens in the two accounts of the "supply side" and the "demand side." The former covers production capacity, calls, costs, investment, corporate income, and local political achievements; the latter relates to wages, employment, leisure, social security, public services, and household purchasing power.
The peculiarity of the Token economy in China lies in its remarkable proficiency in expanding the first account, and the habit of directly equating the achievements of the first account with those of the second. Thus, supply is called demand, intermediate inputs are called consumption, and corporate cost reduction is called prosperity. Precisely because AI is truly productive, it is even more worth questioning who controls the interpretation of this value and who reaps its benefits.
Token, a New Kind of "Electricity Consumption"
Token can indeed be measured as a technical indicator like electricity consumption, but an increase in industrial electricity consumption doesn't tell you whether the current flows to more efficient industries or inefficient excess capacity. Similarly, a Token call might generate a report that genuinely saves you several hours, or it might just be repeatedly revising a piece of text that ultimately wasn't used.
Usage is an input, tasks are the process, income and profit are results at the enterprise level, and productivity is an even higher-level economic outcome. These layers are at completely different distances from citizens' lives. At least there is one path for the improvement of corporate profits: distributing income to employees, who then spend it. But if Tokens continuously improve a food delivery rider's order-taking efficiency and volume without increasing their actual income, this path is cut off.
This is not an abstract concern; the 140 trillion daily call volume itself deserves to be deconstructed. ByteDance's Volcano Engine disclosed in April that the daily Token usage of the Doubao large model has exceeded 120 trillion. That is to say, the vast majority of the national total comes from this single platform, and what drives this number up is mainly AI video generation: generating and iterating a single video often consumes tens of millions of Tokens. This landscape is not at all the same as "comprehensive penetration across all industries." It is more like one company, one type of scenario, one content industry that is exploding. In the face of this structure, it is hard to justify using the daily call volume as an indicator of the overall economic health.
For enterprises, buying Tokens is of course demand on their own books. But placed within the entire national economy, it is likely just an intermediate input from beginning to end. Only when it lowers product prices, increases wages, adds leisure time, improves public services, or creates new products that people are willing and able to buy, do Tokens truly enter the second account.
This is the divide between the supply side and the demand side. The former is a typical productivism, which is probably the default narrative when many people talk about the "Token economy": prioritizing production capacity, fixed assets, technological capabilities, and industrial scale over household income, social security, and final consumption. This prioritization is not accidental. The sense of security for residents is not easily quantified, and service quality is hard to write into year-end reports; but clusters of ten thousand GPUs, Token output per kilowatt-hour, daily call volume, number of industrial parks, and investment scale can all be quantified, planned, financed, subsidized, ranked, and assessed. Chips, servers, power grids, liquid cooling, data centers—all have clear procurement entities suitable for industrial funds, bank credit, state-owned enterprise investment, and local project involvement.
Conversely, increasing unemployment insurance, providing pensions for migrant workers, and improving grassroots public services require long-term, regular fiscal commitments (China's current unemployment-related expenditures are less than 0.1% of GDP, while the OECD average is 1.0%). They also directly touch upon the division of responsibilities between central and local governments. The former is political achievement; the latter is trouble.
Thus, a technological revolution is first read as new fixed assets, new local production capacity, and new administrative visibility. This interpretation is not completed by a single authority. Local governments need investment and achievements, state-owned data companies need business, computing power centers need to run at full utilization, industrial parks need investment attraction, cloud vendors and model companies need early customers—and state-owned enterprises and government departments just happen to be able to act as those earliest buyers.
The Token economy did not grow from the market itself, nor was it created by a single command. It emerged from the push of various parties, but for different reasons: some want strategic security, some want asset utilization, some want revenue, some want financing, some want assessment results. The interface that stitches together these different calculations is the Token.
Movie “Blade Runner 2049”
Stimulating Demand by Expanding Production Capacity
Beijing's Yizhuang provides an almost complete sample. The locality proposed to achieve an intelligent economy industrial scale of 400 billion yuan by 2030: building four 10,000-GPU-class token factories, increasing computing power scale to over 100,000 P, raising token output per kilowatt-hour to over six times the current level, building token distribution and consumption infrastructure, nurturing more than 50 intelligent agent R&D and collaboration platforms, and attracting over 5,000 OPC entities.
This set of policy tools follows the same line as other industrial policies in China. It doesn't just focus on infrastructure but covers and subsidizes the entire industrial chain: a 30% subsidy on computing power leasing fees, 100 million yuan in annual data vouchers, up to 50 million yuan for token aggregation and unified settlement platforms, up to 10 million yuan annually for intelligent business reconstruction and delivery platforms, subsidizing all the way to enterprises consuming Tokens in actual scenarios. The government supports factories producing Tokens, platforms distributing Tokens, and enterprises using Tokens.
Subsidies have their reasonable part. New technologies often face bottlenecks in the early promotion stage, such as insufficient infrastructure, enterprises' reluctance to be the first to adopt, and suppliers unable to find application scenarios. These are typical coordination failures. Computing power vouchers and Token vouchers can lower the cost of first adoption, forcing enterprises to open up data and processes, even nurturing services that didn't originally exist.
But it's also worth noting a "semantic conversion." The relevant chapter is titled "Comprehensively Expand Intelligent Economy Consumption Models," but its main content is to encourage enterprises in healthcare, commercial aerospace, automotive manufacturing, etc., to open up high-throughput scenarios, driving large-scale, high-frequency consumption of Tokens; for enterprises that carry out applications in actual scenarios, providing up to 5 million yuan in funding support based on 50% of Token consumption costs.4
In other words, in a chapter titled "Expanding Consumption," the government subsidized half of the intermediate input consumption costs for enterprises. Enterprises purchasing Tokens is purchasing means of production, belonging to intermediate input, which is not equal to final consumption by residents; policies occurring at the purchasing end do not mean they are demand-side policies. The government lowers Token costs for enterprises, enterprises increase calls, platforms gain revenue, and call volume in turn becomes a reason to expand production capacity—money circulates in a loop among the government, platforms, and enterprises. As long as this loop does not extend a branch leading to wages, social security, public services, reliable price reductions, or resident transfer payments, the so-called demand side remains merely the supply side buying its own products.
This is the distinction I most want to clarify in this entire article. It's not that we are not stimulating demand, but that we are too accustomed to stimulating demand by expanding production capacity. The past four trillion was like this, photovoltaics were like this, and when it comes to Tokens, we extend the same habit into the intelligent era.
Movie “Ghost in the Shell”
The Limitations of Supply-Side Discourse
Supply-side discourse usually has a nice straight line: stronger models, cheaper Tokens, so more usage; more usage, so higher productivity; higher productivity, so increased corporate income and profits, followed by increased laborer income, and thus money enters the demand side. Each link alone may hold true, but years of supply-side experience tell us that between every two links lie sectoral interests, industry competition, and distribution patterns.
First, the most direct layer: Token does not equal profit; it is first and foremost an uncontrollable bill. By mid-2026, it was precisely this bill that made many American companies hesitate. KPMG's quarterly survey showed that about half of the surveyed executives scaled back intelligent agent deployment, citing costs exceeding benefits. Intelligent agents and tools like Harness integrate general models into a company's data, tools, permissions, memory, and workflows, turning a single Q&A into continuously executable tasks. Models are no longer just chat windows occasionally opened by employees but may become the carrier for all tasks, appearing efficient on the surface at least.5
But value cannot be judged solely by consumption. From the frenzy of Token Maxxing (referring to the prevalent practice around 2025 in large tech companies: using AI as intensively as possible, measuring employee innovation by their Token usage, treating consumption itself as a progress indicator. As AI pricing rises linearly with usage, this practice directly led to the 2026 bill backlash.) to now, the myth that "Token equals efficiency" has basically been shattered. Few still doubt AI's ability to improve efficiency; the real questions are: What tasks have been completed through large-scale AI use? How much time was saved? How much revenue increased? And how much verification, rework, error, and compliance cost generated? If only the first three are included in statistics, leaving the last four for employees to silently digest, enterprises get a nice but one-sided evaluation.
Thus, there is a fundamental ladder: calls do not equal task completion, task completion does not equal revenue increase, revenue increase does not equal profit improvement, and profit improvement does not equal a synchronous rise in macro productivity and social welfare. Each step up requires new evidence. This is not a threshold set specifically for AI; any intermediate input must pass the same test of financial common sense.
Even if an AI project does improve enterprise efficiency, the matter is not finished. Call volume, platform revenue, corporate profit, and downstream customer value are four different things. Looking at Zhipu's 2025 annual report, both API calls and revenue are growing rapidly: annual revenue grew 132%, MaaS platform annual recurring revenue increased 60-fold within twelve months, while adjusted net loss still widened by 29% during the same period.
And this is not trading volume for price cuts; Zhipu raised API prices by 83% in Q1 2026, while call volume grew 400%.6 Rising pricing power, usage, and revenue simultaneously, yet profits still haven't caught up. These numbers don't prove that the model business has no value—high-growth companies can of course trade losses for R&D and market—but they do prove that we cannot use a first-level indicator to masquerade as the next-level result. Photovoltaics have walked the same path: losses did not narrow with scale expansion; instead, they grew larger: Tongwei's net loss in 2024 was 7 billion yuan, and the forecast for 2025 is a loss of 9 to 10 billion yuan.7
Then comes the distribution problem. If AI creates 100 yuan of incremental value, chip companies, clouds, model providers, application platforms, using enterprises, laborers, and consumers will all come to share this cake. From a supply-side logic, this 100 yuan will likely be mostly taken by upstream enterprises in the chain. Technology hasn't abolished distribution politics; it just re-entered the stage wearing an AI coat.
Macro structures make this distribution problem more evident. The IMF estimates that in 2024, China's gross capital formation accounted for about 40% of GDP,8 while Qiushi states that the household consumption rate in the same year was about 39.9%.9 The two numbers have different calibers and cannot be directly compared precisely, but they point to the same structure: investment scale is enormous, while the economic resources occupied by the household sector are relatively limited. This is basic knowledge about the Chinese economy.
The World Bank estimates the fiscal impulse for 2025 at 1.6% of GDP, of which the part directly targeting households is only about 0.5 percentage points, with the rest mainly public investment. The bank also notes that the link between growth and employment is weakening.10 These phenomena cannot be attributed to AI, but they constitute the initial conditions when AI enters the Chinese economy. When enterprises generally face weak demand and price competition, the easiest thing to use AI for first is not inventing a new world of consumption, but reducing costs, cutting hiring, accelerating supply, and improving export competitiveness.
For an individual enterprise, this is very rational; but summed up across all enterprises, it may become a typical fallacy of composition: suppressing wages and youth employment, weakening resident expectations and consumption, intensifying price competition, prompting enterprises to rely even more on automation to further reduce costs. I don't intend to assert the final outcome of AI here, but this cycle has already repeated in many industries. If productivity gains mainly stay with fixed assets, platform rents, and upstream enterprise profits, efficiency improvements at the micro level may not translate into income growth at the macro level.
Movie “Modern Times”
One Thousand Henry Fords, and the Missing Half
In discussions about the Token economy, the issue is no longer just how many Tokens are consumed, but that the entire production model is being reshaped. In an interview with Huxiu, Professor Sun Tianshu from Cheung Kong Graduate School of Business said:
China's bigger opportunity in this wave of AI is not in replicating a new consumer internet, but in AI To B: the industrial restructuring of intelligent agents across thousands of industries; each vertical industry can have its own "industrial brain" with the help of AI, and each enterprise needs to establish its own "intelligent agent system"; completing this transformation requires a group of "Henry Fords" who understand the tacit knowledge, business architecture, and organizational changes of the industry and possess AI-native thinking—they don't just add an AI function to old processes (+AI), but redesign production methods, business models, and industrial order from first principles (AI+), just as Mr. Ford, in the electrical era 100 years ago at Highland Park in Detroit, designed the "assembly line model" centered around electricity, reconfiguring productive forces and relations of production, creating the "electricity-native" new era of the automotive industry.
The metaphor of "Henry Ford" has insight, but also a blind spot. The insight lies on the supply side: stuffing a chat window into an old process usually only brings partial improvement; what enterprises really need to do is redesign tasks, human-machine division of labor, data permissions, model routing, verification standards, and accountability. This point is correct.
The blind spot is that the assembly line was never just about efficiency improvement through technology. Materials from the Ford Museum show that the assembly line experiment from 1913 to 1914 compressed the assembly time of a Model T from 12.5 hours to 93 minutes. The cost appeared simultaneously. According to Raff and Summers' compilation of Ford archives, the annual turnover rate at the Highland Park plant in 1913 was about 370%. To maintain an average workforce of 13,623, over 50,000 hirings were needed throughout the year, with a daily absenteeism rate of 10%. Ford therefore changed the old system of "nine-hour workday, daily wage of $2.34" to the new plan of "eight-hour workday, daily wage of $5".12
These two authors explicitly ruled out labor shortage and attracting high-skilled workers as explanations—the long queues of job seekers outside the factory were already a common sight in Detroit, and the direction of technological change was precisely deskilling. This had almost nothing to do with the later romanticized "letting workers afford the cars they build." It was first and foremost self-help by a factory that couldn't retain people.
But its results were real. The turnover rate dropped to 54% the following year, to 16% in 1915, and absenteeism dropped from 10% to 2.5%; meanwhile, the nominal price of the Model T dropped from $550 in 1914 to $440 in 1916, while the company's net profit still rose. Workers gained stable employment relations and higher incomes, and also housing and savings. But this second half needs a supplement: savings and housing were not a natural spillover but conditions for eligibility. The Five-Dollar Day was only for males over 22 who had been employed for six months and had to pass inspections by about 150 investigators from Ford's "Sociological Department" checking living habits. Alcoholism, gambling, unclean dwellings, or irregular savings could lead to disqualification; in the first six months of the plan, only 69% of the workforce qualified. This was a byproduct of the high-intensity assembly line and also Ford's demand-side aspect—but it was designed, conditional, and regulated, not something that fell out of the assembly line.
What political economy calls macro Fordism is therefore two things. One is standardization, task decomposition, deskilling, mass production, and strict labor discipline. It completely changed industrial production and is exactly what today's Token supply-side imagination needs. The other is not the entrepreneur's goodwill nor a wage increase policy, but an accumulation regime mutually supported by mass production, relatively stable wages, mass consumption, welfare systems, and long-term employment.13
Token discourse easily produces the first kind of "Ford," encouraging process reconfiguration, cost reduction, capacity expansion, AI architects, and super individuals. But it hasn't yet proven that the second kind of "Ford" will appear: after productivity increases, will wages rise? Will working hours decrease? Will social security and public services improve? Can household purchasing power accommodate the new supply? What China might need is not a thousand Fords, but the missing half of Fordism.
The person pointing out this blind spot was actually at the same conference. Professor Lynn Wu from the Wharton School, based on two decades of robot adoption data in Canada, pointed out at the symposium hosted by Sun Tianshu's research department that high-skill and low-skill positions are relatively safe, with medium-skill workers bearing the greatest impact; a longer-term effect is the fracture of the promotion path from low-skill to high-skill, which will also reshape social mobility.14
Returning to the Chinese economy along this question, many see the Token and the AI economy behind it as the new engine after real estate, even a new absorber of capital. This inference has explanatory power but must be restrained. The two are indeed similar in institutional form: land, credit, and industrial parks can be replaced by electricity, computing vouchers, and state-owned enterprise clouds; square meters and investment amounts can be replaced by PFlops, Token volume, and API revenue; local investment competition can shift from real estate to chip, model, and computing power projects.
But the similarity stops there. Real estate once absorbed a large amount of low- and medium-skill labor, linked backward to steel, cement, home appliances, decoration, and forward to land finance, credit, and household wealth expectations; it did create employment, resident income, and a wealth effect.15 AI infrastructure is highly capital-intensive, and AI itself is labor-saving technology. It can absorb capital and electricity but does not automatically generate proportionate total wages, local employment, and household demand. Token factories can expand production capacity, but they cannot create their own buyers.16
Movie “Modern Times”
From Output Back to Life
The consumption of Tokens can only tell us how fast this economic machine is running, but not who thereby gains jobs, income, time, and a better life. To judge the Token economy, in the end, we probably only need to ask three questions: Does Token growth come from market payment, public subsidies, or internal organizational circulation? Do AI projects create new revenue and new services, or are they mainly used to reduce manpower and lower costs? Do productivity gains go into household income, public services, and consumption, or do they remain in corporate profits, platform rents, and fixed asset investment?
These three questions currently have no answers. The 140 trillion figure is more like a huge, bright number, like a factory operating nonstop at night. We can see the lights, electricity meters, and assembly lines, but we don't yet know if people outside the factory can afford what it produces.
The biggest risk of the Token economy is not that Tokens themselves have no value, but that the real productive power of AI is once again placed into a system only adept at expanding supply. It will create more Fords of the intelligent era, more precise assembly lines, and more quantifiable output. And the other half of Fordism will not automatically fall from the assembly line; in history, it has always been the result of negotiations involving labor relations, fiscal commitments, and political pressure, not a byproduct of capacity expansion.
Footnotes and References
1. At a press conference of the State Council Information Office, National Data Bureau Director Liu Liehong stated that by March 2026, China's daily token (Token) call volume had exceeded 140 trillion, growing over 1000 times compared to the 100 billion at the beginning of 2024, and over 40% compared to the 100 trillion at the end of 2025. Xinhua News Agency, "China's Daily Token Call Volume Exceeds 140 Trillion," March 24, 2026. National Bureau of Statistics Deputy Director Mao Shengyong reiterated the same figure at the SCIO press conference on April 16.
2. ByteDance's Volcano Engine disclosed on April 2, 2026, that the Doubao large model's daily Token usage had exceeded 120 trillion. The main driver of consumption is multimodal video generation scenarios; generating and iterating a single AI video can consume tens of millions of Tokens. See The Paper's report in April 2026 and Volcano Engine's public disclosure.
3. World Bank statistics show that only 47% of China's urban workforce is covered by unemployment insurance. In 2023, the average monthly unemployment benefit was 1,814 yuan, less than 20% of the average urban wage; national unemployment-related expenditures were less than 0.1% of GDP, while the OECD average in 2021 was 1.0%. World Bank, China Economic Update, June 2025, p. 35.
4. "Beijing Economic and Technological Development Zone Measures on Supporting Token-Driven Intelligent Economy High-Quality Development (Trial)" (Jing Ji Guan Fa [2026] No. 9, referred to as the "Ten Token Measures"), finalized on August 3, 2026, released on August 5, effective until December 31, 2030, applicable to the 225 square kilometer area of Yizhuang New City.
5. KPMG "Q2 2026 Global AI Pulse" shows that about 49% of surveyed executives scaled back intelligent agent deployment, citing costs exceeding benefits, while 79% still ranked AI as an investment priority. Multiple independent evidence groups: McKinsey "Enterprise AI FinOps Survey" (May 2026) stated 93% of respondents exceeded AI budgets; PwC "2026 Global CEO Survey" said only 12% of CEOs reported AI benefits in both cost and revenue, 56% reported neither; Gartner (March 2026) estimated intelligent agents require 5 to 30 times more Tokens per task than standard chatbots. Uber exhausted its annual AI coding tool budget in April 2026; its COO admitted at a May 25 meeting the inability to connect this expenditure to consumer-facing product improvements; OpenAI CEO Sam Altman told CNBC in June 2026 that skepticism about AI expenditure returns is currently the fairest criticism of AI. The trigger point was an Axios investigation report in late May 2026 on "AI sticker shock."
6. Zhipu (Knowledge Atlas Technology) listed on the Hong Kong Stock Exchange on January 8, 2026, and released its first post-listing annual report on March 31. 2025 revenue was 7.24 billion yuan, a year-on-year increase of 131.9%; MaaS platform annual recurring revenue (ARR) was about 17 billion yuan, a 60-fold increase within twelve months; MaaS platform gross margin rose from 3.3% to 18.9%. During the same period, net loss was 4.718 billion yuan, a year-on-year increase of 59.5%; after excluding non-cash and one-time items such as share-based payments (558 million yuan), changes in fair value of financial instruments (937 million yuan), and listing expenses, adjusted net loss was 3.182 billion yuan, a year-on-year increase of 29.1%. R&D expenditure was 3.18 billion yuan, a year-on-year increase of 44.9%, 4.4 times the annual revenue. This article uses the adjusted figures.
7. Wang Bohua, honorary chairman of the China Photovoltaic Industry Association, pointed out that the scale of losses caused by industry fluctuations in 2024 far exceeded those of the previous three industry fluctuations. Annual sequences for individual companies show losses expanding with scale: Tongwei's net loss in 2024 was 7.039 billion yuan, forecast 2025 loss 9-10 billion yuan; Trina Solar's loss in 2024 was 3.443 billion yuan, forecast 2025 loss 6.5-7.5 billion yuan, having lost for six consecutive quarters; JA Solar's loss in 2024 was 4.656 billion yuan, forecast 2025 loss 4.5-4.8 billion yuan. As of the end of 2024, main industrial chain capacities for silicon materials, wafers, cells, and modules reached 1,447 GW, 1,160 GW, 1,193 GW, and 1,428 GW respectively. Even so, from January to February 2025, 31 projects across the entire PV industrial chain were announced, signed, or started construction, with a total scale of nearly 200 GW and total investment of nearly 90 billion yuan.
8. IMF World Economic Outlook database, "total investment" item, China 2024 approximately 40.4%; National Bureau of Statistics expenditure method calculation shows capital formation rate around 40.5%. This value changes slightly with each WEO revision, so an approximate number is used here.
9. Qiushi.com commentator, "[Theoretical Power of China] Why Increase the Household Consumption Rate?", December 5, 2025. The same article points out that China's household consumption rate still lags behind developed countries by 10 to 30 percentage points, with the service consumption share being relatively low. Explanation: The household consumption rate refers to household final consumption expenditure as a percentage of GDP; the final consumption rate also includes government consumption. National Bureau of Statistics data shows the final consumption rate in 2024 was 56.6%, stable above 50% since 2013; the capital formation rate stable above 40%. Therefore, 39.9% and 40% are not additive; the difference mainly consists of government consumption and net exports.
10. World Bank, China Economic Update, June 2025: Unlocking Consumption, pp. 1–2, 11. The bank estimates the 2025 fiscal impulse at 1.6% of GDP, of which the part directly targeting households is about 0.5% of GDP, with the rest mainly public investment. Special chapter Jobs in Transition, p. 21. From 2015 to 2019, GDP grew annually by 6.7%, urban employment grew annually by 2.7%, with a net increase of 55.5 million urban jobs over five years, and real per capita resident income grew annually by 6.7%; from 2020 to 2024, GDP grew annually by 4.9%, urban employment grew annually by only 0.9%, with a net increase of 21 million urban jobs over five years, and real per capita resident income grew annually by 4.8%.
11. The Henry Ford, "Ford's Five Dollar Day Revolution." The museum records that the assembly line experiment from 1913 to 1914 compressed Model T assembly time from 12.5 hours to 93 minutes, with a turnover rate of 370% during the same period; another exhibit description gives the figure as "turnover rate of 380% at the end of 1913" and states that to net add 100 workers, 963 had to be hired.
12. Daniel M. G. Raff & Lawrence H. Summers, "Did Henry Ford Pay Efficiency Wages?", NBER Working Paper No. 2101, December 1986 (later published in Journal of Labor Economics, 1987)
13. "Macro Fordism" and "accumulation regime" are concepts from the French Regulation School, referring to the overall arrangement of mass production, relatively stable wages, mass consumption, welfare systems, and long-term employment mutually supporting each other. Main literature: Michel Aglietta, A Theory of Capitalist Regulation: The US Experience (1979); Alain Lipietz, Mirages and Miracles (1987); Robert Boyer, The Regulation School: A Critical Introduction (1990). Can also be traced back to Antonio Gramsci, "Americanism and Fordism" (Prison Notebooks, 1934).
14. Minutes of the scholar symposium hosted by the Cheung Kong Graduate School of Business AI Intelligent Industry Research Department, April 2026. Lynn Wu is a tenured professor at the Wharton School, University of Pennsylvania, Operations, Information and Decisions, whose judgment is based on two decades of robot adoption data in Canada.
15-16. The World Bank points out that since 2021, the downturn in real estate has led to a continuous decline in construction employment, with public infrastructure investment only partially offsetting the reduced labor demand from real estate. World Bank, China Economic Update, June 2025, pp. 22–23.pp. 28–29. Nearly a quarter of jobs in China are in the "routine manual" category, higher than other upper-middle-income and high-income economies; the share of "routine cognitive" jobs is comparable to developed economies; while the share of "non-routine cognitive" jobs that can be complemented rather than replaced by AI is lower than in developed economies, potentially limiting China's space to benefit from AI (citing Arias et al., 2025). Other empirical evidence shows that Chinese manufacturing enterprises (especially high-tech ones) after adopting AI increased demand for high-skilled labor and decreased demand for low-skilled labor (Xie et al., 2021); robot adoption from 2010 to 2016 negatively impacted employment and wages for on-the-job salaried workers, especially low-skilled and older workers (Giuntella et al., 2025).
This article comes from the WeChat public account "YouthologyYouthology" (ID: openyouthology001), author: Guyu, editor: Yang Shao
