In March 2026, the Political Bureau of the Central Committee of the CPC for the first time included the “Computing Power Network” into the national “Six Networks” system — ranking alongside the water network, power grid, and transportation network. What does this mean? It means that computing power has been upgraded from a “technical tool” to a “national infrastructure”. Just as electricity reshaped industrial civilization more than a century ago, new-type computing power is now reshaping intelligent civilization. And the depth and breadth of this revolution far exceed most people’s imagination.
01 A set of figures to see how explosive the computing power boom is
Let’s start with a few numbers.
As of March 2026, China’s total intelligent computing power scale has reached 1882 EFlops (exaFLOPS, or 10^18 floating-point operations per second). The number of standard racks in active computing centers has reached 14.45 million. More than 70 major computing power channels have been completed. The national integrated computing power network monitoring and scheduling platform has connected approximately 70% of the country’s intelligent computing resources.
The daily average Token call volume has exceeded 140 trillion, an increase of over 1000 times compared to the beginning of 2024.
1000 times. You read that right.
What does this mean? It means AI is no longer just a toy in the lab, but has penetrated into every capillary of the economy and society like water and electricity. The explosive growth of intelligent agents has made computing power a “fundamental factor of production” in the intelligent era — this is not a metaphor, but the exact wording from official policy documents.
But a problem has also emerged: amid the boom of computing power, a sharp contradiction has surfaced — the end of computing power is electricity. Without power, computing power is nothing but piles of overheating silicon chips.
02 Computing-Power and Power Coordination: When AI Starts to “Compete for Electricity”
“The end of AI is computing power, and the end of computing power is electricity.” This statement is evolving from industry consensus into policy action.
The 2026 Government Work Report for the first time included “computing-power and power coordination” in the new infrastructure projects. This marks the first time that “computing-power and power coordination” has been written into the national-level new infrastructure category. Looking back at the policy timeline: the computing-power and power coordination mechanism was first proposed in 2023, pilot implementations were clearly launched in 2024, and it was directly elevated to a new infrastructure project in 2026 — a three-leap jump in three years.
Why the urgency? Because the data is staggering.
First, electricity cost has become the largest expense for computing centers. Research reports from Huachuang Securities show that power costs account for as high as 56.7% of data center operating expenses, ranking first. In Shenzhen, an intelligent computing center with a scale of over 6000 PFLOPS sees electricity bills make up more than 70% of its operating costs.
Second, global computing power electricity consumption has reached a non-negligible scale. Data from the International Energy Agency (IEA) shows that global data center electricity consumption in 2025 was close to 500 TWh, accounting for about 1.6% of total global electricity consumption. And this figure is still growing rapidly.
Third, the eastern computing power hubs are already “running out of electricity”. Research from Guosheng Securities points out that there are power supply gaps in the Beijing-Tianjin-Hebei region, the Jiangsu-Zhejiang-Shanghai region, and the Guangdong-Hong Kong-Macao Greater Bay Area. The power gaps in Jiangsu, Zhejiang, and Guangdong have reached 245.8 billion, 234.9 billion, and 249.5 billion kWh respectively. Local green power supplies can no longer meet the expansion needs of computing facilities.
This is exactly the underlying logic of “East Data, West Computing” — it’s not that the western region is more suitable for placing servers, but that the west has cheaper and more abundant green power. The green intelligent computing center project in Haidong, Qinghai, has a total investment of 3.5 billion yuan, a total computing power of 20,000P, and a PUE (Power Usage Effectiveness) value of only 1.19. This figure means that almost every kilowatt-hour of electricity is used for computing, with very little waste.
Policies have set hard targets: new data centers at national hub nodes must have a green power proportion of over 80%. A responsible official from the National Data Administration further clarified: ensure that the green power application proportion of new computing facilities at hub nodes reaches over 80%.
This is not a suggestion, it’s a hard constraint.
03 Breakthrough of Domestic Chips: Leap from “Usable” to “Excellent”
The core of computing power lies in chips. Without domestically independent and controllable chips, no matter how many computing centers you build, you are just making wedding clothes for others.
On May 26, 2026, a landmark event took place: the China Information Technology Security Evaluation Center and the National Secrecy Science and Technology Evaluation Center jointly issued the “Announcement of Security and Reliability Evaluation Results”, creating a separate category for “artificial intelligence training and inference chips” for the first time. Nine domestic AI chips from seven Chinese enterprises were all rated Security Level I.
What does this mean? It means that China’s domestic AI computing power infrastructure has officially been incorporated into the national information technology application innovation (ITAI) security certification system. The ITAI project has fully extended from traditional fields of CPUs, operating systems, and databases to core AI hardware. This certification result will become the de facto access catalog for government, enterprise, and critical sector entities when purchasing AI chips.
Market data is equally encouraging. In 2025, China’s AI server market delivered approximately 4 million AI GPUs in total, with domestic chips accounting for 41% of the share. Morgan Stanley predicts that by 2030, China’s AI chip market will reach a size of 67 billion USD, and domestic chips are expected to meet around 76% of market demand.
Specifically, several forces are worth noting:
Huawei Ascend — the flagship force. Both Ascend 310 (inference) and Ascend 910 (training) were selected into the first batch of ITAI Level I certifications. Shipments in 2025 were approximately 812,000 units, and the revenue from the AI processor business is expected to exceed 12 billion USD in 2026. The new-generation Ascend 950PR matches the performance of NVIDIA H200, with leading enterprises including ByteDance, Tencent, and Alibaba increasing their procurement.
The “Four Little Dragons of GPUs” — Enflame, Biren, MetaX, and Iluvatar CoreX. Enflame Technology’s STAR Market IPO application was accepted in January 2026, making it the first enterprise accepted by the STAR Market in 2026. It has independently developed and iterated four generations of architectures and five chips, deployed intelligent computing centers in Qingyang (Gansu), Wuxi (Jiangsu), Yichang (Hubei) and other locations, and deeply participated in the “East Data, West Computing” project.
Vimicro Technologies — XPU Multi-core Heterogeneous Computing. The Starlight Smart 5 chip, with 8 units deployed together, can support the full-performance operation of the 671-billion-parameter DeepSeek large model. It launched its STAR Market listing counseling in August 2025, with a valuation exceeding 20 billion yuan.
Alibaba Pingtouge — a new force in cloud ecosystem computing power. The two chips Zhenwu M530 and M890 both passed the Level I certification, and are deployed and verified on a large scale through Alibaba Cloud.
From “usable” to “excellent”, from “optional” to “mandatory” — domestic AI chips are undergoing a qualitative transformation. This is not the breakthrough of a single enterprise, but the systematic rise of an entire industrial cluster.
04 Processing-in-Memory: A Paradigm Shift to Break the “Memory Wall”
If domestic chip substitution is “overtaking on a different track”, then processing-in-memory (PIM) is “rebuilding the road”.
To understand this issue, we must first grasp a fundamental contradiction: data movement is “eating away” at computing efficiency.
Since John von Neumann proposed the stored-program computer architecture in 1945, the computing unit and storage unit have always been separated. Data is frequently moved between the processor and memory. It’s like a factory where the raw material warehouse is far away from the production line: to make every part, workers have to carry materials from the warehouse to the production line, and then carry the finished products back to the warehouse.
This doesn’t matter when parts are small. But when the parameters of large models grow from billions to hundreds of billions, the energy and time consumed by data movement become a fatal bottleneck. Jensen Huang, CEO of NVIDIA, once admitted: “GPUs spend 70% of their time waiting for data.”
This is the so-called “memory wall” and “power wall”.
The core logic of PIM is very straightforward: embed computing units into the memory array, so that calculations can be completed where the data is stored. Build the office directly inside the warehouse, so raw materials are right at hand, ready to use at any time.
In 2026, this technology achieved a landmark breakthrough. At ISSCC 2026, a joint team from Tsinghua University, Huawei, and ByteDance published a paper: a hybrid in-memory computing chip based on the 28nm process, which improves the core computing efficiency of recommendation systems by 1-2 orders of magnitude — QPS increased by 66 times, and QPS/W increased by 181 times.
Note that this is a 28nm process. You don’t need the most advanced manufacturing process to achieve order-of-magnitude efficiency gains. What this means for China’s semiconductor industry is self-evident.
Chinese enterprises have formed a rich technological ecosystem in this track. Based on SRAM PIM, MemryX launched Hongtu H30, China’s first PIM intelligent driving chip, with a computing power of 256 TOPS and power consumption of only 35W. The WTM2101 chip from CIX Technology has shipped over 10 million units, applied in smart wearable devices from brands including Huawei and Xiaomi. XSemi is the only domestic enterprise that has achieved ReRAM mass production, and ByteDance has previously invested in it. NanoCore pioneered the 3D-CIM (3D Processing-in-Memory) architecture, achieving a computing power density increase of over 4 times and power consumption reduction of over 10 times.
It is predicted that the global PIM chip market will exceed 12 billion USD in 2025, with China accounting for 30% of the share.
05 Computing Power Network: Allocate Computing Power Just Like a Power Grid
With chips and computing centers in place, what’s the next step? Connect them all together.
In 2026, the Outline of the 15th Five-Year Plan proposed “building new-type infrastructure in a moderately proactive manner”, making clear arrangements for the construction of the national integrated computing power network. The Political Bureau of the Central Committee for the first time included the computing power network into the national “Six Networks” system — ranking alongside the water network, power grid, transportation network, oil and gas network, and logistics network.
What is the core goal of the computing power network? Connect scattered data centers, supercomputing centers, and intelligent computing centers across the country, and allocate them uniformly for on-demand use just like a power grid.
From an architectural perspective, this network has three layers:
The “Skeleton” — Communication Networks. Relying on the most widely covered communication networks and data center resources across the country, the three major telecom operators have become the primary “network weavers” of the computing power network. China Mobile has built the largest operator-level intelligent computing center in the country, China Telecom continues to deploy ten-thousand-card clusters, and China Unicom is promoting large-scale computing power investments in multiple locations.
The “Heart” — Chips and Computing Capability. This is the core source of computing power supply.
The “Blood Vessels” — Optical Communication System. Cross-regional computing power scheduling requires the support of an optical communication system with ultra-low latency and ultra-large bandwidth.
There have already been successful cases in practice. China Telecom’s “Xirang” computing power interconnection and scheduling platform has a total self-owned and accessed intelligent computing scale of over 91 EFlops, ranking first in China’s computing power internet scheduling. It provides differentiated scheduling services prioritizing cost, latency, and green low-carbon performance, accessing self-owned computing power as well as multi-party resources from internet enterprises, private clouds, and idle social computing power. In Shenzhen, the smart city computing power coordination and scheduling platform based on “Xirang” has achieved unified management and scheduling across service providers, regions, and architectures.
Since its launch in April 2024, the National Supercomputing Internet Platform has connected 14 provinces and municipalities, and more than 30 national-level supercomputing and intelligent computing centers, capable of providing nearly 70 types of standardized computing services, with over 7,300 online application service products and more than 1.2 million registered users.
Not long ago, 60,000 domestic AI acceleration cards were put into use at the core node of the National Supercomputing Internet in Zhengzhou, marking the completion of the largest scientific intelligent computing cluster in the country. Work that used to take quantum material experts a full day to complete can now be done in just one hour using the supercomputing internet intelligent agent.
06 Green Computing Power: Not Just Fast, But Also “Green”
The “new” of new-type computing power lies not only in fast computing, but also in “green” computing.
In 2026, “computing-power and power coordination” was written into the Outline of the 15th Five-Year Plan, and the Government Work Report for the first time included it in the new infrastructure projects. The National Data Administration made it clear: ensure that the green power application proportion of new computing facilities at hub nodes reaches over 80%.
China’s green power development provides the confidence to achieve this goal. Data from the National Energy Administration shows that in 2025, China’s newly added installed capacity of wind power and solar power exceeded 430 million kW, and the total grid-connected installed capacity of wind power and solar power reached 1.84 billion kW, accounting for 47.3% of the total, historically surpassing thermal power.
But the challenges remain severe. The computing power industry is reshaping the structure of power demand — as of June 2025, China’s total computing power of computing equipment has reached 962 EFlops, accounting for about 21% of the global total, with a year-on-year growth rate of 73%. The intelligent computing power scale reached 782 EFlops, a year-on-year increase of 96%. It is estimated that China’s total computing power will reach 767 EFlops in 2026, of which intelligent computing power will account for 73%.
This means that the demand for electricity from computing power is still growing exponentially. The eastern region is already “running out of electricity”, and while the western region has advantages in green power, it requires long-distance transmission.
The industry is exploring the new “DC direct supply” (DC/DC) model to reduce AC-DC conversion losses and improve energy efficiency. Zhong Baoshen, Chairman of LONGi Green Energy, suggested: promote the large-scale implementation of “green power direct connection”, establish a cross-regional benefit feedback mechanism, reasonably verify transmission fees, and improve the physical traceability technology system. The essence of these suggestions is — to make green power, like tap water, flow directly through pipelines from the “power plants” in the west to the “computing centers” in the east, reducing losses and costs in intermediate links.
07 The “Four New” Features of New-Type Computing Power
After sorting through the above content, we can clearly see the four core features of “new-type computing power”:
First, New Architecture — Processing-in-Memory. Break the 80-year dominance of the von Neumann architecture by moving computing into the memory. Instead of running faster on the old road, build a completely new road. Achieving