Currently, the bottleneck for AI is "electricity"—China's advantage and the U.S.'s anxiety

The "second wave" of generative AI is beginning to differentiate the competitiveness of countries and companies not by the intelligence of models but by the strength of the power grid. U.S. AI and data center officials who inspected China's transmission and power generation infrastructure stated, "The U.S. power grid is clearly inferior,"—a point highlighted by a Fortune report on August 14. The article's main point is that in the U.S., the power connections for new data centers are delayed, and the power grid cannot keep up with the increase in model training and inference, while China can flexibly enhance its power transmission and location capabilities. Additionally, several industry papers have reported that "China's supply capacity is several steps ahead."Tom's Hardware

In reality, recent data center investments are on a scale beyond imagination. McKinsey estimates that global investment in new and expanded data centers from 2025 to 2030 will be around $6.7 trillion. To accommodate AI, which consumes an extraordinary amount of electricity for both training and inference, simultaneous strengthening of transmission and distribution facilities, speeding up system connections, and securing long-lead-time equipment like transformers are necessary. However, in the U.S., all of these are becoming bottlenecks.Tom's Hardware

U.S.: It's not just power plants that are lacking

The most severe issue currently is the supply shortage of transformers, particularly large power and distribution transformers. While demand is increasing due to a surge in energy consumption, manufacturing lead times are long, and there is uncertainty about future procurement. Wood Mackenzie pointed out that by 2025, the shortage rate of power and distribution transformers in the U.S. could reach 30% and 10%, respectively. Supply constraints are directly linked to delays in connecting new data centers.Reuters

Pressure on pricing is also spreading. Reports from AP and industry media point out that the power demand from AI data centers is driving up local grid costs, leading to higher electricity bills for ordinary households. States are beginning to explore imposing special charges or mandating self-procurement, arguing that data centers should bear a fair share of infrastructure costs. In Oregon, legislation has already been enacted to mandate regulatory authorities to develop new pricing designs.AP NewsTom's HardwareDaily Journal of Commerce

Even the question, "How much power do data centers consume?" has divided opinions. Deloitte estimates that global data center consumption will account for only about 2% of total electricity by 2025, while other estimates suggest that the share of power consumption in the U.S. could reach 6.7% to 12% by 2028, considering the rapidly increasing AI demand. The variation in numbers itself illustrates the difficulty in forecasting supply and demand in the AI era.DeloitteBloom Energy

China: "Comprehensive Strength" in System, Permits, and Location

China's strength lies not in a single element but in the comprehensive power of "power generation (generation) × transmission (including UHV long-distance trunk lines) × system integration at demand sites × speed of permits." They have a system in place to widely accommodate large-scale renewable energy through high-voltage direct current or ultra-high-voltage AC trunk lines and quickly connect to high-load hubs. AI data centers are like new factories with outputs of 100–300 MW, and they have established means to quickly achieve power connections through public-private cooperation. This contrasts with the prolonged queue for system connections in the U.S., which is a concern for those who have seen the situation firsthand. This leads to a sense of crisis that "China might take the lead in the AI infrastructure race."Tom's HardwareGoldman Sachs

How did social media react: Pros and cons, economy, and health impacts

This issue has also heated up on social media. In Reddit's tech and future-oriented communities, voices are calling for "urgent institutional reform," citing U.S. regulatory and permitting processes as the biggest delay factors. On the other hand, opinions are emerging that call for strengthening location regulations and pollution control, stating, "The backup generators and NOx emissions from massive data centers impose external diseconomies on nearby residents." Here are some representative points (summary).Reddit

• "From a national security perspective, AI infrastructure should include self-sufficient power generation. The boundary between civilian and military use is becoming blurred" (summary).Reddit

• "In nearby residential areas, nighttime noise and diesel smell during backup power generation are issues. Community consensus is being overlooked" (summary).San Francisco Chronicle

• "Electricity bills are rising, but local employment is not increasing significantly. It's unclear who bears the external costs" (summary).AP News

The "temperature" on social media is not simply "China is great/U.S. is bad." The modernization of the grid is necessary, but it is returning to classic public policy issues of how to design the burden of costs, environmental impact, and regional acceptability.

Corporate strategies: Self-sufficient power, demand-side resources, and location selection

Major cloud companies are beginning to move towards new strategies that incorporate connection delays and rising costs. Firstly, there is the use of on-site or near-site self-sufficient power. The trend of securing "island operation" capabilities, which are not affected by grid congestion or outages, by combining gas engines, fuel cells, and storage batteries, is increasing. Reviews in academic journals also predict that "data centers with self-generation" will become common in the future.engineering.org.cn

Secondly, there is the utilization of demand-side resources (DR) and flexibility. Reducing computational load during peak times and running batch processes during nighttime or surplus renewable energy times is effective. "Traffic shifting," which moves inference tasks geographically based on price, is also effective. Goldman Sachs points out that smart demand management can contribute to alleviating the "AI energy crisis" and that the payback on capital investment can improve depending on institutional design.Goldman Sachs

Thirdly, there is the optimization of location. Distributing to areas with grid capacity, good transformer supply outlook, and easy access to renewable energy sources. However, if locations are dispersed, network latency and operational complexity increase, making "edge/region design" like CDN even more important.

Policy options: Reforming permits and improving "grid UX"

As policies for national, state, and local governments, the immediate priorities are (1) standardizing and streamlining the grid connection process, (2) long-term planning and transparency of transmission and distribution investments, (3) enhancing price signals (location-based pricing, expansion by time of day), (4) strengthening domestic supply networks for long-lead-time equipment (transformers, switchgear, etc.), and (5) location rules incorporating environmental externalities. Several states have already begun reviewing "who bears the infrastructure costs," but balancing resident burden suppression and investment incentives requires a delicate balance of the system.AP NewsDaily Journal of Commerce

"AI doesn't just consume electricity"

Data center investments are also impacting the macroeconomy. There is analysis that by 2025, data center capital investment will begin to visibly contribute to U.S. growth rates, but there is also a risk that the strain on related infrastructure such as electricity, communications, water resources, and human resources will lead to other cost increases. The expression "data centers consume the economy" may sound exaggerated, but at least the urgency of infrastructure investment and institutional design is a reality.Fortune

Implications for Japan: Three things to decide now

1. Speeding up and making grid connections transparent: The key is to visualize constraints on the grid side early and establish a system to quickly handle large demand connections by combining renewable energy, constant backup, and storage.

2. Edge distribution × demand flexibility: Pricing design that utilizes surplus electricity at night and in spring and autumn as "AI valley time" and optimization of scheduling on the application side will determine competitiveness.

3. Technology-neutral self-sufficient power policy: While balancing decarbonization goals, do not exclude diverse options such as fuel cells, small modular reactors (SMR), and geothermal.

In short, the success of AI is determined as much by "the intelligence of the model" as by "the skillfulness of electricity." What China demonstrates is the fact that by comprehensively increasing supply capacity and grid flexibility, the ramp-up of AI investment can be accelerated. For the U.S. and Japan, improving the "mundane UX" of permits, connections, and pricing design will be the next source of competitiveness.

Powered by Froala Editor