Shen Yan: three warnings of the U.S.’s AI predicament
PKU professor argues that speculative capital, power constraints, and labour disruption in the U.S. serve as warnings for China's AI push.
The rapid rise of artificial intelligence (AI) in the United States is accompanied by a series of pressing challenges that, according to Shen Yan, a professor of economics at the National School of Development (NSD), fall largely into three categories. First, speculative investments in AI are inflating a potential bubble, as the market shifts from blind faith in technology to a more critical evaluation of return on investment and scalable profitability. Second, AI-driven data centres are putting immense pressure on the U.S. power grid. Third, AI’s impact on lower-skilled jobs is heightening the risk of structural unemployment, with ongoing layoffs becoming increasingly commonplace.
Accordingly, Shen highlights three key lessons for China in the global AI race. The first is to be wary of capital bubbles and ensure AI investments are grounded in the real economy. Second, China should invest in training programmes that promote human-machine collaboration and plan early for social policies that support flexible employment. Lastly, China needs to balance AI growth with energy infrastructure development, ensuring that technological progress does not outstrip the country’s energy infrastructure capacity.
Shen published this article in the “Peking University NSD Think Tank Series” column on Economic View of the China News Service on 22 January, 2026. It is also available on the NSD’s official WeChat blog.
Shen has kindly authorised its translation.
—Yuxuan Jia
沈艳：美国AI困局的三重警示
Shen Yan: Three Warnings of the U.S.’s AI Predicament
Recent in-depth reviews and analyses of The Wall Street Journal’s Tech News Briefing suggest that AI development in the United States is moving out of its capital-fuelled frenzy and into a phase of fiercer competition and structural adjustment. Technological iteration and capital spending are still rising at an exponential pace, and the competitive landscape is becoming more diverse. But the economic gains from AI have yet to be widely realised. Even so, its effects on productivity, the labour market, and energy infrastructure are already emerging, with the potential to drive bigger social and economic change.
Capital spending remains strong, but the market is growing more wary of whether returns will materialise, and whether a bubble is forming
A recent Teneo survey of more than 350 CEOs at listed companies found that more than two-thirds plan to increase AI spending over the next year. Yet this willingness to double down has not eased rising investor scepticism. In the summer of 2025, for example, an MIT report found that as many as 95 per cent of corporate AI pilot projects had failed. Then, in the autumn, a web of circular capital transactions among tech giants, including OpenAI, Nvidia, AMD, Oracle, and Microsoft, deepened concern that capital was becoming too concentrated and increasingly detached from real output.
The market is, in fact, moving beyond simple faith in technology and towards a harsher assessment of return on investment and the prospects for scalable profitability.
The competitive landscape is shifting sharply, with control of core infrastructure emerging as the key strategic high ground
The landscape of technological leadership is undergoing a dramatic shake-up. OpenAI’s lead once looked unassailable, but it has recently come under pressure from rivals such as Google’s Gemini. At the same time, in underlying infrastructure such as GPUs, giants including Google are trying to loosen Nvidia’s grip. Enterprise customers are moving away from dependence on a single provider and towards combinations of models from different vendors. The frontier-model market is entering a more contested, “warring nations” phase.
Energy, meanwhile, has overtaken capital and talent as the tightest constraint on the expansion of AI computing power. Analyses by the International Energy Agency and the Electric Power Research Institute point to a sharp rise in electricity demand from AI in the United States, turning energy supply into a major challenge for the national energy system. In 2024, data centres accounted for 4 per cent of total U.S. electricity consumption. A typical AI-focused hyperscale data centre consumes as much power each year as 100,000 American households. Some large AI data centre projects now under construction are expected to consume 20 times as much electricity as a typical existing facility.
The IEA projects that by 2030, total U.S. data centre electricity demand will reach 133 per cent of its 2024 level, placing severe strain on an ageing grid. China’s relative advantage in power capacity stands in sharp contrast to current U.S. grid conditions, adding to American anxiety over competition in AI. The future of the U.S. AI industry will therefore revolve around investment in new energy, including nuclear (such as Microsoft’s plan to restart the Three Mile Island nuclear plant) and geothermal power, and policy support effectiveness.
AI’s impact on the labour market is also harder to ignore, and the risk of structural unemployment is rising
The impact of AI on employment has entered a more concrete phase of job displacement. Amazon CEO Andy Jassy has publicly said that AI will reduce its corporate workforce, a sign that companies are shifting from “stockpiling” talent to optimising it. This carries two dangerous structural risks.
Entry-level white-collar positions are the first to face the threat of unemployment. A sharp decline in openings for recent graduates and young people entering the workforce could cut off the traditional starting point for a generation’s career progression, creating lasting mismatches in human capital.
Second, ongoing layoffs may become the new normal. At the micro level, AI-driven productivity gains could turn workforce reduction from a periodic adjustment into a steady, drip-feed process of optimisation.
From a macroeconomic perspective, while AI can create new jobs, these positions require highly specialised skills and will not fully offset the loss of large numbers of traditional roles. The result could be a labour market stuck in prolonged instability. If ordinary workers face declining incomes and weaker job security, then the goods and services produced through AI-enabled productivity gains may run into a demand shortfall, which would in turn drag on growth.
Insights into U.S. AI development trends and their economic impact
This leaves the United States at a critical crossroads in 2026. The technological race and the build-out of infrastructure (energy and semiconductors) will continue at speed. But the ability of the U.S. economy and society to absorb and adapt to these changes will face a severe test. The main pressures will fall in three areas.
The first is energy and digital infrastructure. If AI is to deliver expected profits, the United States must elevate the upgrading of the national power grid, the promotion of energy innovation, and the optimisation of data centre planning as matters of national strategic competitiveness.
The second is labour transition and the social safety net. The U.S. needs a large-scale and efficient reskilling system, as well as new forms of social protection and income distribution suited to an era of persistent employment volatility.
The third is systemic risk. Vigilance is required against sharp market swings caused by the gap between investment expectations and real returns, while in-depth research is needed into the potential dampening effects of automation on aggregate demand.
Even if Washington raises these issues to the level of national strategy, implementation will remain difficult. Take energy and digital infrastructure as examples.
The first difficulty is whether a distant solution can solve an immediate problem. Expansion of the ageing U.S. grid is nowhere near keeping pace with demand. In some regions, new projects must wait up to seven years for grid connection, and operators have already warned that “There is simply no new capacity to meet new loads”. However much the government tries to speed up approvals, the physical construction cycle of power systems remains badly out of step with the exponential growth of AI computing demand.
Another difficulty lies in resources and social equity. The vast electricity and water consumption of this build-out poses a twin challenge to local ecosystems and carbon-reduction goals. At the same time, the costs of the required infrastructure are being passed disproportionately to households and small businesses through higher electricity bills, potentially exacerbating regional inequality.
The United States also faces institutional friction stemming from intense political and social division. The federal government’s use of funding restrictions to pressure states into loosening regulation could trigger constitutional litigation and sharpen tensions between Washington and the states. Meanwhile, efforts to weaken AI ethics, safety governance, and environmental standards in pursuit of development have already drawn broad criticism and may leave the U.S. in a weaker position when global governance rules are set.
In sum, the United States has shown strong strategic intent and considerable capacity to mobilise resources, but deep-seated contradictions such as ageing infrastructure, political and social polarisation, and structural economic imbalance remain fundamental obstacles in this long contest.
The above analysis of trends in U.S. artificial intelligence development indicates that the pace of technological change may far exceed the capacity of social systems to adapt. That offers China three warnings as it develops its own AI sector.
The first is to guard against capital bubbles and technological myths. That means strengthening scrutiny of the commercial viability and integration of AI investments into the real economy, and emphasising that AI should be used to raise productivity across all sectors.
The second is to confront employment disruption and skills transition head-on. China needs to plan early for vocational training built around human-machine collaboration, and for social policies that support flexible forms of employment. AI development must remain people-centred, rather than being pursued at the cost of large-scale joblessness.
The third is to move beyond a narrow view of energy. Computing power, green energy, especially nuclear power, and grid upgrading need to be planned together as an integrated system of national strategic infrastructure to ensure the sustainable momentum of the technological revolution and keep social costs under control. That, in turn, requires a systematic approach to AI development, one that balances incentives for innovation, protection of people’s livelihoods, and national security.
 |
|