The global AI computing power race in 2026 has long moved beyond the single dimension of chips and algorithms. Power supply and energy infrastructure capabilities are becoming the core variables determining the efficiency of computing power deployment and the upper limit of industrial expansion.
Recently, PJM, the largest regional grid operator in the US, submitted a new power supply framework to the Federal Energy Regulatory Commission (FERC), completely rewriting the power supply rules for US data centers: merely obtaining grid access qualification for large-scale computing projects no longer equates to having equal rights to reliable power supply. To achieve stable and uninterrupted power supply, data centers must independently configure, sign contracts for, and lock in sufficient new power capacity; otherwise, during periods of tight grid load, they will be the first to be instructed to reduce load and operate during off-peak hours.
Simply put, the US computing power industry has officially entered a new stage of "bring your own power to secure reliable supply."
Behind this rule change lies the most authentic structural difference in computing infrastructure between China and the US: the US does not lack advanced chips, capital, or power generation technology, but uniquely lacks reliable megawatt-level power delivered on schedule; China does not lack power installed capacity, infrastructure production capacity, or cross-regional dispatch capabilities, with the core constraint stuck on high-end AI chips and the software and hardware ecosystem. The industrial chains of the two major economies are highly complementary, yet constrained by multiple factors, making it difficult to form strategic dependence, ultimately leading to two completely different paths to break through the computing power bottleneck.
PJM Reshapes Grid Rules, Completely Separating "Grid Access Qualification" from "Power Supply Reliability"
This new PJM policy is a fundamental institutional adjustment for the US to adapt to the explosive growth of AI computing power, far more complex than the online rumor of "forcing data centers to build their own gas power plants." Its core logic is to break the inherent model of "equal power supply upon grid access," separating grid access and reliable power supply into two independent products, making new large-scale computing loads bear the primary responsibility for the power gap they bring.
According to the proposal submitted by PJM, new large-scale data centers will have only three power supply options in the future:
First, BYONC (Bring Your Own New Capacity) self-built new capacity. Projects can build their own gas power stations, expand nuclear power capacity, configure energy storage, or sign long-term rigid contracts to lock in grid-certified incremental reliable power resources, fundamentally making up for the system capacity gap;
Second, grid backup procurement as a safety net. If regional power capacity is insufficient, they can participate in the PJM reliability backup procurement mechanism, locking in power supply contracts for up to 15 years, with related costs shared among the project, local regulators, and power entities;
Third, Interim Reliability Assurance Service (IRAS). Projects without supporting new capacity and not participating in backup procurement can still connect to the grid normally, but only enjoy lower-priority interim power supply. When the grid enters peak load or power emergency states, such computing projects must reduce load ahead of residential, traditional commercial, and industrial users to ensure basic livelihood power consumption.
More critically, starting from the 2029/2030 annual capacity auction, computing loads without new power support will no longer be included in the unified grid reliability guarantee system. The grid may allow projects to connect and land, but will no longer unconditionally provide a safety net for their power supply, completely ending the era of computing enterprises "occupying public grid capacity for free."
Currently, this proposal is in the public comment and pending approval stage. PJM has requested the Federal Energy Regulatory Commission (FERC) to issue a ruling within 60 days, which may comprehensively reshape the development logic of computing clusters in the eastern US.
94% of New Power Consumption Comes from Computing, "Paper Overcapacity vs. Actual Shortage" in the US Grid
A set of core data disclosed by PJM intuitively demonstrates the supply-demand imbalance between computing power and electricity in the US: from 2024 to 2030, the total new power demand in its jurisdiction is about 32GW, of which 30GW comes from the new peak load of data centers, with the computing industry contributing 94% of the power increment.
It needs to be clarified that this 30GW is not merely the IT load of GPUs and servers, but the complete system-level peak power demand. The industry has a clear distinction in load caliber: IT load is only the power consumption of the computing equipment itself, superimposed with the losses of supporting facilities such as cooling, power distribution, fans, and water pumps (PUE factor), plus the auxiliary power consumption of the park, ultimately forming the peak load on the grid side. For a 100MW IT computing cluster, under a PUE of 1.15, the actual power load of the park can reach over 115MW, which is also the core caliber for PJM to count incremental loads.
While demand is exploding, the carrying capacity of the US grid has long failed to keep up with the expansion speed of computing power. PJM's long-term forecast for 2026 shows that although the regional summer peak load has been slightly adjusted downward, the average annual load growth rate over ten years still reaches 3.6%, far exceeding the 0.3% predicted in 2021, and the power supply-demand gap continues to widen.
The current US power market presents an extreme structural contradiction: paper projects are piled up like mountains, while actual available capacity is severely insufficient. In the PJM interconnection study queue, 715 energy projects with a total nameplate capacity of 201.5GW are hoarded, but there is still a reliable capacity gap of 6831MW in the 2028/2029 annual capacity auction, with only 525MW of new effective capacity finally landed, directly pushing power prices to the upper limit.
Data from the Lawrence Berkeley National Laboratory illustrates the problem even better: by the end of 2025, over 2060GW of power generation and energy storage projects nationwide are queued for grid connection, with 549GW already contracted but unable to be put into production for a long time. The median cycle from application to grid connection for new energy projects exceeds 5 years, transmission line construction cycles often take a decade, and the delivery cycle for core power equipment exceeds two years.
In contrast, large-scale AI data centers take only 1-2 years from site selection to production. The construction speed of computing power has completely outpaced the speed of US power approval, construction, and grid connection. The US dilemma has never been a lack of power generation technology, but a lack of an efficient and on-time delivery power infrastructure system, and a lack of "reliable megawatts" that match the rhythm of computing power.
From Buying Power to Generating Power, US Computing Enterprises Transform into Energy Developers
The new PJM policy forces the industry to completely reconstruct its development logic. The past model of "acquiring land first, building data centers, and then applying for power" is completely eliminated. Power First has become the sole criterion for US computing projects. Land, policies, and tax incentives are no longer core prerequisites; deliverable power sources, transmission channels, gas supporting facilities, and approval cycles are the core yardsticks for whether a project can be landed.
The identity of the computing industry has also undergone an essential transformation: from a mere power consumer to an energy investor, capacity developer, and long-term power service provider. Various mature "computing + energy" binding models have emerged in the industry.
The old coal power base in Homer City, Pennsylvania, is being transformed into a 4.4GW large-scale AI energy park. Relying on existing power land and pipeline network resources, it is equipped with gas power generation to ensure stable power supply for computing power, and is expected to be gradually put into operation in 2027; Microsoft has locked in an 835MW nuclear power restart project, locking in incremental nuclear power capacity through long-term power purchase agreements to provide a safety net for high-end computing clusters; Meta has innovated the "user-funded, grid-built" model, specifically supporting a 2.2GW gas combined cycle power plant for the exclusive supply of its own data centers.
Meanwhile, modular gas turbines, distributed energy storage, and fuel cells have become transitional rigid demands, supporting the phased production of computing projects and making up for the construction gap of large-scale energy power stations.
The industry has also formed a clear consensus: not all green power and power purchase agreements can meet grid reliability requirements. The nameplate installed capacity of wind and solar power cannot be equated with stable available capacity under extreme weather; purchasing power from existing power plants can only lock in power costs and cannot create new reliable capacity for the grid, naturally failing to meet PJM's BYONC rules. The core competition in US computing infrastructure has already become a scramble for incremental reliable power resources.
Eastern Data, Western Computing + Computing-Power Synergy: China's Differentiated Path to Break Through Computing Infrastructure Bottlenecks
Facing the core bottleneck of AI computing power development, China has embarked on a completely different solution from the US. If the core proposition for the US is to solve the timeliness and reliability of power supply, the core proposition for China is to rely on extreme infrastructure capabilities to offset the supply constraints of high-end chips.
China has not copied the model of "data centers building their own power plants," but has achieved the optimal national allocation of computing power, electricity, land, and network resources through top-level coordinated "Eastern Data, Western Computing" and computing-power synergy. Since its launch in 2022, eight national computing hubs and ten major data center clusters have built the core skeleton of national computing power.
As of July 2026, the national intelligent computing scale has reached 2.4 million PFLOPS, with the eight major hubs bearing over 80% of the national intelligent computing power, and 1.4 million PFLOPS of computing power connected to the national integrated computing network pilot platform. The core logic of this system is to let computing power follow electricity and tasks follow latency: migrating low-latency-sensitive tasks such as large model training, offline rendering, and batch data analysis to western distributed wind, solar, and hydropower energy bases; keeping high-precision, low-latency businesses such as financial transactions, industrial control, autonomous driving, and real-time inference in eastern cities and edge nodes to achieve precise resource matching.
Relying on the massive 4.04 billion kW of national power generation installed capacity and the 340 million kW of ultra-high voltage transmission capacity for "West-to-East Power Transmission," China's computing infrastructure has formed advantages in scale, low cost, and low carbon. Policies clearly require that the proportion of green power application in new hub computing projects exceeds 80%, continuously reducing computing energy consumption and carbon emissions through source-grid-load-storage synergy, direct green power supply, and waste heat utilization.
It needs to be viewed objectively that China's computing infrastructure still has shortcomings. Currently, the rack deployment rate of the ten major clusters has reached 62.72%, but rack utilization rate does not equal GPU effective computing power utilization rate, and issues of hardware idleness and computing power mismatch still exist. The most core constraint remains the shortcomings in high-end training chips, advanced processes, high-bandwidth memory, and supporting software and hardware ecosystems, which is also the core direction of special research and development by multiple domestic departments.
US-China Computing Infrastructure: Complementary but Hard to Depend on Each Other, Each Making Up for Core Shortcomings
Reviewing the two paths of AI computing infrastructure in China and the US, they are inevitable choices under the industrial endowments, institutional systems, and resource constraints of the two countries, forming distinct corresponding differences.
Relying on leading chip design, EDA tools, large model technology, and capital markets, the US occupies the top advantage in AI technology, but is constrained by a fragmented grid, lengthy approval processes, and lagging infrastructure delivery capabilities, trapped in the bottleneck of "lacking reliable power"; relying on market-oriented capacity markets and rules binding rights and responsibilities, it forces the computing industry to build its own energy system, breaking through with extreme breakthroughs in single-point technology + project self-provided energy.
Relying on globally leading power equipment manufacturing, ultra-high voltage transmission, large-scale engineering construction capabilities, and top-level coordinated dispatch systems, China has solved the problem of large-scale landing of computing infrastructure, but is constrained by the high-end semiconductor industry chain, trapped in the bottleneck of "lacking top computing hardware"; relying on national resource coordination and computing-power synergy dispatch, it makes up for single-point hardware shortcomings with systemic advantages.
From the perspective of industrial logic, the two are originally perfectly complementary: the US exports high-end chips, model technology, and capital, while China exports power infrastructure, engineering capabilities, and large-scale computing power carriers, which can form the world's optimal AI industrial ecosystem.
However, multiple barriers such as geopolitical policies, security reviews, and export controls have completely blocked the possibility of deep strategic dependence. The US implements case-by-case licensing reviews for high-end GPUs, high-bandwidth memory, and semiconductor equipment, and foreign investment security mechanisms restrict capital cooperation in the AI and semiconductor fields between China and the US; CFIUS reviews, cross-border data supervision, and computing infrastructure security controls have basically invalidated the deep co-construction between China and the US in the fields of core computing power, energy, and data.
Ultimately, the global AI computing power industry has formed a pattern of dual-track parallelism and independent closed loops: the US accelerates the construction of an autonomous and controllable energy-computing system, adapting power infrastructure to chip technology advantages; China accelerates breakthroughs in the semiconductor software and hardware ecosystem, offsetting hardware shortcomings with the scale of computing infrastructure.
New Industry Trend: Computing Power Development Enters the Era of Integrated Design of "Power + Chip + Task"
The new PJM policy and the divergence of paths between China and the US mark that global computing infrastructure has completely bid farewell to the extensive stage of "valuing data centers over energy" and entered a new stage of refined and coordinated development, also pointing out a clear direction for the subsequent development of the industry.
For computing projects in the US PJM region, the core is to adhere to power-first prerequisites: locking in incremental capacity, transmission schedules, and equipment scheduling before project landing, distinguishing between ordinary power purchases and incremental capacity contracts; splitting rigid inference loads and flexible training loads, reducing power supply pressure through peak shaving and migration; combining modular transitional power sources and long-term baseload power sources to build a layered and reliable energy architecture, eliminating false claims of nameplate capacity.
For China's computing industry, the core is to promote efficient resource synergy: continuously deepening the division of labor in computing power between the east and west, eliminating blind westward migration and invalid construction; shifting assessment standards from a single PUE to comprehensive indicators such as GPU utilization rate, computing power energy consumption per unit, and green power matching rate; promoting joint optimization of chips and power, algorithms and hardware, reducing reliance on high-end chips through software and hardware adaptation; improving the national computing power trading and dispatch mechanism, solving cross-regional settlement, task scheduling, and data security issues, avoiding resource waste of "having data centers, having power, but no effective computing power."
Conclusion
The ultimate competition in AI computing power is no longer a contest of single hardware, but a comprehensive game of computing infrastructure systems, energy guarantee capabilities, and resource dispatch efficiency.
PJM's rule of "bringing your own power in exchange for reliable supply" is essentially the inevitable result of the computing industry bearing external costs for its own high-speed growth under a market-oriented mechanism; China's model of "computing-power synergy and national coordination" is the optimal solution for optimal resource allocation in the era of large-scale infrastructure.
The US is solving the problem of "having power available," while China is breaking through the bottleneck of "having chips available." The two major economies have their own strengths and weaknesses and could originally be complementary, but are moving towards independent closed loops under the boundaries of industrial security. In the future, the global AI computing power landscape will maintain a long-term dual-power divergence pattern of "US extreme single-point technology and China extreme system scale."
Author: Qu Haifeng
A senior domestic expert in the field of digital infrastructure and computing power energy, and deputy director of the CDCC Data Center Expert Technical Group. With 27 years of deep cultivation in the industry, he is a core reviewer and editorial board member of the "Code for Design of Data Centers" and "Standard for Data Center Projects," leading the promotion of standardization and low-carbon upgrading of the domestic computing industry.
He formerly served as an Infrastructure Researcher (P10) at Alibaba Cloud, and is a core initiator of the Western Cloud Base in Zhongwei, Ningxia, and a pioneer of the Alibaba Cloud Zhangbei Super Cloud Base, creating an early benchmark computing hub for "Eastern Data, Western Computing" in China.
He has long been deeply engaged in the construction, operation, and energy planning of the entire computing power industry chain, focusing on AI computing infrastructure and the construction of an energy-autonomous computing system. He is a highly influential senior expert in the fields of industry carbon neutrality and new computing infrastructure.
Disclaimer: The article only represents the personal views of the author and is for readers' reference and learning only. If there is any infringement, please contact us, and it will be deleted within 24 hours.