The forcibly acquired 100,000-card computing power projects are turning into hot potatoes for some.

In early July, China's first full-domestic 100,000-card AI supercomputing cluster was completed and put into operation in Zhengzhou, where computing power supply cannot meet the surging demand; thousands of miles away in Inner Mongolia, in an already operational intelligent computing park, the annual comprehensive utilization rate of computing power is less than 50%.

"It's not that there isn't enough computing power, but that computing power is deployed in the wrong place," a person from a local state-owned computing power platform said.

The problem this statement points to is exactly what the *Action Plan for Promoting the Collaborative Development of Large, Medium, and Small Enterprises in the Platform Economy (2026–2028)*, jointly issued on June 18 by seven departments including the Ministry of Industry and Information Technology, the Central Cyberspace Administration of China, the National Development and Reform Commission, the Ministry of Science and Technology, the Ministry of Commerce, the State Administration for Market Regulation, and the National Data Administration, aims to solve: that is, to promote the construction of a national integrated computing power monitoring, scheduling and service platform, and guide all types of intelligent computing clusters to uniformly connect to the national-level scheduling system.

A person close to the relevant department of the National Data Administration told the Economic Observer reporter that the overall implementation path spans a three-year cycle, with large-scale nationwide grid connection to be achieved by 2028.

These three years may be the most challenging period for all regions to balance their computing power accounts.

01

Scrambling for Position

The focus of this round of fierce competition across regions is 100,000-card-level intelligent computing clusters.

In early July, in Qingyang, Gansu, the Smart Blueprint Western Intelligent Computing Center with a total investment of 5.1 billion yuan broke ground. Its first phase features 8,860 AI cabinets, all designed for a 100,000-card scale, and is scheduled for delivery in May 2027.

In the same month, Jinan Urban Construction Investment pushed forward the "Quancheng Computing Power One Network" project, which is planned according to the 100,000-card cluster standard, with construction starting in the second half of 2026 and phased operation in 2027.

In Dianjun District, Yichang, Hubei Huayun Computing Energy, a special SPV company established in the early years with state-owned capital incubation, won the bid for the EPC project of its Huayun Intelligent Computing Hub Center at a value of about 740 million yuan. The project is scheduled to enter the electromechanical construction phase in the third quarter of 2026 and is positioned as a regional medium-sized computing power supporting node.

Meanwhile, China's first full-domestic 100,000-card AI supercomputing cluster has been completed in Zhengzhou.

This is only a part of the publicly available information.

According to calculations by China International Engineering Consulting Co., Ltd. on July 11, the direct investment in national computing power infrastructure in 2026 will exceed 400 billion yuan, and the investment across the entire industrial chain during the "15th Five-Year Plan" period will reach about 6 trillion yuan. The computing power network has been included in the list of six core infrastructure projects for the "15th Five-Year Plan".

Urban construction investment platforms, state-owned capital platforms, and third-party computing power enterprises across regions are all scrambling for positions. Government special bonds, digital industry funds, policy-based long-term loans, and EPC (Engineering, Procurement, and Construction) advance payments — all types of capital are pouring into the same direction.

"Everyone is rushing to secure land, electricity, and energy consumption indicators," the aforementioned person from a local state-owned computing power platform told the Economic Observer. "We have to occupy the position first, no matter what."

Computing power projects with a scale of less than 100,000 cards can hardly squeeze into this game anymore.

In Inner Mongolia, in an already operational intelligent computing park, most projects are market-oriented thousand-card-level general-purpose computing power projects, which were built on a large scale relying on green power. However, due to delays and scattered customers, the annual comprehensive utilization rate of computing power is less than 50%. The newly reserved machine room for 100,000-card domestic intelligent computing in this park currently has a deployment rate of less than 20%.

But what comes after securing the position?

Zhang Shouye said, "The hardware depreciation period for AI servers is only five years. Even if the machine room does not run computing power, the annual costs of operation and maintenance, air conditioning, and basic circuits still need to be paid."

The construction period for a 100,000-card cluster from groundbreaking to operation is 8 to 12 months, but the supporting power facilities often take 18 to 24 months to complete.

"The machine room is waiting for power supply," the aforementioned person from the local state-owned computing power platform described this time mismatch.

02

Calculating the Accounts

The concentrated start of projects is only the first step. Whether the invested capital can generate returns is another matter.

In the park where Zhang Shouye works, the comprehensive utilization rate of computing power is about 30%, and the most severely underutilized part is the domestic heterogeneous computing power pool, with an average annual utilization rate of less than 20%. The reason is that this computing power pool adopts a mixed deployment scheme of multiple domestic chips, and the heterogeneous architectures of different manufacturers cannot be uniformly managed at the scheduling layer. Customer model migration requires adapting to drivers and operator libraries one by one, which severely reduces the actual available computing power. A high-end GPU (Graphics Processing Unit) is depreciated over five years, and being idle for one year will directly affect the book profit by tens of millions of yuan.

The situation is even more challenging at the district and county levels. A computing power hub project in a certain district and county in Hubei has a total investment of several hundred million yuan, with the district-level state-owned capital contributing only tens of millions of yuan, and more than 500 million yuan funded by the EPC general contractor through advance payments, which will be paid in installments over five years. The comprehensive payback period of the project exceeds 10 years, nearly three years longer than that of computing power projects in eastern provincial capitals. "70% of the computing power orders rely on cross-provincial sales, but the cross-provincial settlement mechanism is not fully established, and the payment collection period is two to three months," a staff member of the project operator said. "The cash flow pressure runs through the entire cycle of construction and operation."

For the Jinan municipal urban construction investment project financed by government special bonds, the static pure computing power lease payback period is 10.5 years, which is reduced to 8.3 years after considering comprehensive benefits. The hard bottom line of the calculation is: in the 6th year of project operation, the operating net cash flow must cover the principal and interest of the special bonds for that year; otherwise, the special bond application will not be approved.

The divergence in the assessment of state-owned computing power assets also reflects differences in investment logic. A provincial state-owned enterprise in the eastern region implements a dual-track system of annual business assessment and three-year strategic tenure assessment. Losses of computing power assets in the first three years are not included in negative points, and the three-year tenure focuses on qualitative indicators such as the implementation of domestic computing power and the driving effect on the regional digital industry.

Compared with the lenient tenure fault tolerance mechanism of industrial provincial state-owned enterprises, the assessment constraints of Jinan Urban Construction Investment are more rigid. Relying on the joint supervision framework of multiple departments including state-owned capital, finance, and big data, its computing power projects need to simultaneously complete tasks such as maintaining and increasing the value of state-owned capital, ensuring the cash flow for special bond repayment, and supplying inclusive computing power to local small and medium-sized enterprises. The completion of tasks is linked to the compensation and performance of the management team.

The aforementioned person from the local state-owned computing power platform summarized this: "Industrial state-owned enterprises can focus on long-term goals, but urban construction investment must carefully calculate every account, because special bonds need to be repaid."

The upstream equipment side is also doing the math. Chen Xinxin, the business leader of the aforementioned domestic AI chip manufacturer, said that for the offline training scenario of large models with hundreds of billions of parameters, under the same 100,000-card scale, it takes about 12 days for overseas chip clusters to complete training, while full-domestic chip clusters take 18 to 21 days, and the effective computing power output is only 55% to 65% of that of the overseas solution. The annual comprehensive operation and maintenance cost per card is about 4,100 yuan for domestic chips and about 2,800 yuan for overseas chips, nearly 46% higher. She said, "Many urban construction investment customers only calculate the one-time hardware input and ignore the continuous operation and maintenance expenses during the five-year depreciation period, and they do not feel the cost pressure until the cluster is put into operation."

Since the accounts are so hard to balance, why are regions still rushing to build?

More than one interviewee from the state-owned platform expressed a similar judgment to the Economic Observer: the national integrated computing power scheduling required by the new policy means that computing power can be allocated across regions in three years. At that time, whoever holds sufficient computing power assets will be qualified to access this network and obtain scheduling orders.

The aforementioned person from the local state-owned computing power platform said, "What we are scrambling for now is the admission ticket. By the time the scheduling platform is fully operational, it will be too late to build new ones."

But after getting the admission ticket, the accounts still need to be carefully calculated.

Chen Xinxin said that for new 100,000-card clusters starting construction after June 18, 2026, a unified scheduling interface can be prefabricated at the server factory stage, and the transformation cost of a single machine room only accounts for 3% of the total hardware investment. However, for the existing clusters that have been put into operation before, in parks that have mixed purchased multiple domestic chips and overseas servers, the private protocols of different manufacturers are not interoperable, and the transformation cost often amounts to tens of millions of yuan, with the transformation period of a single cluster generally 4 to 6 months.

In other words, the new policy can help new projects save a sum of transformation investment, but for existing parks like Zhang Shouye's, grid connection itself is a new account to calculate.

03

Waiting for Grid Connection

The policy solution is "national integrated computing power scheduling", which weaves the scattered computing power across the country into a network, allowing idle computing power in the west to receive orders from the east.

The aforementioned person close to the relevant department of the National Data Administration said that a two-year buffer transformation period will be given to the existing operational 100,000-card clusters, and full computing power grid connection will not be forced to complete at one time. Idle computing power resources are allowed to be managed in batches.

But the process of "weaving the network" is more complex than expected.

Zhang Shouye's park has a lot of idle computing power, but he is well aware that this is not a price issue. The proportion of green power in the park has remained above 84% all year round, and the on-grid electricity price is about 0.35 yuan per kWh, about 0.25 yuan lower than that in the east. The price is competitive, but the computing power still cannot be rented out.

The core constraint is latency. The one-way network latency from Inner Mongolia to eastern cities is 4.2 to 5.8 milliseconds. For businesses such as offline model training and batch simulation rendering, this latency is acceptable, but such demand only accounts for a small part of the total computing power demand. Online quality inspection in intelligent manufacturing requires real-time issuance of AI inference instructions; if the latency exceeds 3 milliseconds, the production line will stall. For AI painting, intelligent customer service, and real-time interaction of digital humans, users are sensitive to response speed, and the latency of western computing power will reduce product experience.

Zhou Ming, the person in charge of computing power procurement at a manufacturing enterprise in the east, said, "More than 60% of the computing power demand cannot be transferred to the west."

Zhou Ming's judgment comes from the classification of actual businesses: online real-time inference businesses can only stay in the east. Offline training tasks can theoretically be migrated, but involving industrial confidential drawings and self-developed model weight data, the compliance department has clearly drawn a red line that core business data is not allowed to flow out of the province.

Data security is another hurdle. The theoretical design of the national integrated computing power scheduling platform is to automatically match idle and low-cost western computing power, and allocate surplus eastern computing power during off-peak hours. But for enterprises, the cross-provincial flow of computing power means the cross-provincial transmission and storage of data. The compliance and legal departments of the manufacturing enterprise where Zhou Ming works clearly require that industrial drawings and production process parameters must not leave the province. The scheduling platform cannot fix the physical location of the computing power machine room, so there is a risk of data leakage during cross-provincial flow.

Latency is a physical constraint, data security is an institutional red line, and cross-provincial settlement is an institutional shortcoming. According to the aforementioned person close to the relevant department of the National Data Administration, cross-regional computing power scheduling and settlement have been demonstrated in multiple computing power hubs in the Yangtze River Delta, Inner Mongolia, and Guizhou. The current core bottlenecks are concentrated in three major mechanisms: cross-provincial tax retention, GDP (Gross Domestic Product) accounting, and energy consumption index allocation. The Ministry of Finance and the State Taxation Administration have simultaneously drafted the guiding opinions on tax sharing for cross-regional computing power collaboration, but there are still shortcomings in the complete market-oriented closed loop. The valuation standards for computing power assets of local state-owned platforms are not unified, and there is no industry benchmark for cross-regional computing power pricing.

Zhang Shouye also has experience in this. The park where he works has accessed the pilot of the national integrated scheduling platform. The platform charges a 5% scheduling service fee for matching orders, and the remaining revenue is divided between the park and the eastern demander at a ratio of 8:2. But the unified cross-provincial payment collection period is as long as two to three months. He said, "There are no unified detailed rules for cross-provincial tax and energy consumption index allocation, which makes the financial accounting period very long."

Zhang Shouye has another pressure: the state-owned capital assessment requires the park to reserve 15% of its computing power every year to supply local small and medium-sized enterprises at a preferential price.

Zhang Shouye said, "This part of the business has almost no profit, but the assessment indicators are there, so we have to do it."

On the one hand, idle computing power cannot be rented out; on the other hand, local enterprises cannot afford or access it. Zhang Shouye's situation is not unique.

For people like Zhang Shouye, the three-year transition period means two layers of waiting: waiting for the scheduling platform to truly operate, and waiting for the cross-provincial settlement rules to be fully implemented. But every year the idle computing power is vacant, it eats up tens of millions of yuan in book profits, and accessing the scheduling platform means additional transformation investment and a longer payment collection period.

Every month of delay adds more costs.

(As requested by the interviewees, Zhou Ming is a pseudonym in this article)

This article is from the WeChat official account "Economic Observer", written by WANG Yajie, and published with authorization by 36Kr.