China's power edge brings mixed AI blessings

From afar, China's artificial intelligence progress looks unstoppable. Up against Washington's chip curbs, the country boasts world-class models and a plethora of apps and chatbots. Moreover, the underlying infrastructure - data centres running on cheap and plentiful power - stands in sharp contrast to the United States, where an aging electricity grid is struggling to cope and firms are facing a years-long wait for new capacity to come online. Little wonder Jensen Huang and other executives have sounded the alarm. The Nvidia boss warned in November that "China is going to win the AI race" thanks to lower energy costs and looser regulations. A month earlier, OpenAI flagged an "electron gap" that will jeopardise America's leadership. The panic looks overblown.

Huang has a point. In 2024, the People's Republic generated over 10,000 terawatt-hours (TWh) of electricity, more than double the United States, where installed capacity was just a third of China's, per energy think tank Ember. Moreover, this electron gap, as it is called, is widening as Beijing pushes ahead to wean the country off fossil fuels: by 2030, wind and solar capacity will double and almost triple respectively from current levels, reckons Wood Mackenzie. In fact, the consultancy estimates that by then, renewables will supply 5,500 TWh of power output, or 40 percent of total generation, more than enough to meet the country's forecast 479 TWh of data-centre demand for that year.

That's a far cry from the United States, where demand for electricity has stayed flat for decades. With little reason to invest in new capacity, data centres now face an estimated 44 gigawatt power shortfall between 2025 and 2028, Morgan Stanley analysts calculate. A huge backlog of energy generation projects seeking grid connection - equivalent to nearly double the country's installed capacity as of the end of 2024 - has become a key bottleneck.

In theory, abundant electricity should make it easier for Chinese firms like Alibaba and ByteDance to deploy more domestic chips - albeit less powerful and less energy efficient than Nvidia's - to train AI models, potentially offsetting the impact of US export controls. Industrial electricity bills are also on average 30 percent cheaper than in the United States, according to Bank of America analysts.

This advantage has not translated into AI gains yet. The People's Republic has lagged the United States in building new data centres and bringing more computing power online, per Bernstein, suggesting US chip controls remain a formidable constraint. In fact, the research outfit's analysts estimate Chinese firms will spend just $147 billion on AI capital expenditures in 2027. That's less than Amazon.com's expected total capex that year, per Visible Alpha forecasts.

Moreover, critics like Huang are also missing a key point: China is investing huge sums to build new capacity in order to keep up with rising electricity demand, which Bernstein notes has outpaced GDP growth in each of the past five years. They forecast electricity consumption will reach 13,500 TWh by 2030, driven by energy-intensive industrial sectors switching from fossil fuels as well as the fast-growing popularity of products like electric vehicles. Data centres are expected to account for just 3 percent of total consumption.

Having greater renewable capacity also means little by itself when deploying it is less straightforward. In the first half of 2025, China's curtailment rate - how much power entering the network is limited due to oversupply or grid constraints - for solar rose to 6.6 percent, up from 3.9 percent a year earlier. In Tibet, solar and wind curtailment rates were as high as 34 percent and 30 percent respectively.

One factor is that most renewable resources are in far-flung areas in the country's west. Transmitting power long-distance to the east and other regions where AI workloads, not to mention competing EV and other manufacturing hubs, are concentrated, has become a huge challenge.

To address this, Beijing is ramping up investments in high-voltage transmission and energy storage and targeting a new grid system to support a west-to-east power transmission program exceeding 420 gigawatts by 2030. That should help ease curtailment rates.

Until then, officials have resorted to relocating data centres out west. The idea behind the ambitious "Eastern Data, Western Compute" plan launched in 2021 was that it is cheaper and more efficient to transmit data across the country using fibre-optic networks. That sparked a construction boom across Inner Mongolia, Guizhou and other resource-rich areas. In reality, though, transfer speeds have proven too slow, rendering many new facilities unsuitable for AI tasks that require real-time responses. That has contributed to a glut of unviable data centres with utilisation rates as low as 20 percent, Reuters reported in July, citing sources.

Improving data and transmission technologies will help, but these current pockets of surplus - in renewable power and data centres - may be part of a broader, systemic problem: rising overcapacity across the AI stack. Even though there is a shortage of high-end chips used in training, Qingyuan Lin at Bernstein reckons local supply of less-powerful processors used for inference, when models respond to user queries, will exceed demand by 2028.

In models and apps, companies like Alibaba, DeepSeek and ByteDance are in a race to the bottom on prices. Smaller challengers like MiniMax and Zhipu, both of which are readying Hong Kong initial public offerings, are bleeding red ink. The latter cites in its prospectus that China's AI market will grow to 993 billion yuan, or $142 billion, by 2030, up from just 219 billion yuan last year; yet despite the heady growth, Zhipu reported a net loss of 2.4 billion yuan for the first half of 2025, or more than 12 times its revenue for the period. Officials have also recently warned that an investment bubble may be forming in the humanoid robotics industry, as there are now more than 150 manufacturers despite the technology being unproven, regulations uncertain and demand questionable.

This type of scenario, known as "involution", is one that Beijing now wants to avoid. Destructive competition and overcapacity may not have stopped Chinese firms from dominating electric vehicles, batteries, solar panels and other industries. But they have resulted in deflationary price wars, poor returns on investment, massive misallocation of capital and worsening structural imbalances in the $20 trillion economy. All of that will have knock-on effects on innovation and growth down the road.

The U.S-China electron gap may slow progress at OpenAI and others, buying time for firms in the People's Republic to catch up in chips. But it may also set the stage for another boom-and-bust cycle that has long plagued China's industries.