By Xie Zefeng, Edited by Yang Xuran
As AI elevates to the dimension of great power competition, erecting trade barriers in the field of advanced chips has become an indisputable fact.
Since October 2022, the U.S. Department of Commerce's Bureau of Industry and Security (BIS) launched the first round of "Advanced Computing" export controls, directly placing high-end chips such as Nvidia's A100 and H100, as well as AMD's MI300, on the control list, thereby kicking off a protracted chip blockade war.
Since then, control rules have undergone multiple rounds of upgrades. From A800 and H800 to H20 and then H200, the performance of Nvidia's "China-specific" chips has been repeatedly downgraded, yet it has never been able to change their ill-fated destiny.
Against the backdrop of restricted high-end chips and generational gaps in single-card performance, chips, especially overseas high-end ones, have long fallen out of the scope that domestic AI companies can truly control independently.
Shifting focus away from the chip issue, a deeper structural change is emerging.
As channels for acquiring overseas chips are continuously compressed and domestic substitution is increasingly rolled out, the biggest variable determining the commercial costs of domestic AI large models has quietly shifted from "whether chips can be bought" to "whether computing centers can be sustained."
On this basis, infrastructure investments in power, cooling, and server rooms, traditionally viewed as "supporting," are becoming the real heavy burdens on the desks of AI companies and have also been the core of capital expenditures for AI giants for some time.
This article is an in-depth value piece from the content team of "Juchao WAVE." Welcome to follow us across multiple platforms.
Blockade
When AI chips become a tool for technological strangulation, they evolve into a crucial bargaining chip in geopolitical strategic competition.
As early as October 7, 2022, the U.S. Department of Commerce's Bureau of Industry and Security issued the "Advanced Computing Rule" sanctions. It placed AI chips with computing power greater than 4800 (INT8 exceeding 600 TOPS or FP16 exceeding 300 TFLOPS) and a bandwidth cap exceeding 600 GB/s on the control list, prohibiting their export to China.
This threshold directly "banned" Nvidia's A100 and H100, as well as AMD's MI300 series, leaving Chinese companies unable to access the world's most powerful AI chips.
Nvidia promptly made adjustments, launching the "downgraded" A800 and H800. These two chips complied with the export controls at the time by reducing interconnect bandwidth (NVLink speed), becoming the main products in China's AI computing market. DeepSeek V3/R1 was trained and inferred based on H800 GPU clusters.
The good times did not last long. In October 2023, the U.S. further tightened control rules, and the A800 and H800 were also included in the banned sales list.
To meet export standards, the Nvidia H20 was quickly introduced. This is a "simplified version of the H100 based on the Hopper architecture." By significantly cutting performance to fall just within the control red line, it became the backbone of Nvidia's sales in the Chinese market at that time.
Many domestic companies adopted the H20 to deploy DeepSeek-V3/R1 inference services, and a situation of "extreme scarcity of cards" even emerged.
Entering the second half of 2025, control policies saw new changes. In April, the sales of "China-specific" chips such as the H20 and MI308 to China were suspended. Some media even reported that Nvidia, AMD, and others needed to hand over 15% of their revenue from chip sales to China to the U.S. government in exchange for export licenses.
Just three months later, on July 15, the U.S. Department of Commerce resumed the export of H20 and MI308 to China. Almost at the same moment, the Cyberspace Administration of China summoned Nvidia regarding the "vulnerability and backdoor security risks" of the H20 chip, sending a strong signal of high vigilance toward the security of overseas chips.
At the end of that year, the U.S. further allowed Nvidia to sell H200 chips (an upgraded version of the H100) to China, albeit with extremely harsh transshipment and review conditions.
With the popularization of domestic substitution, coupled with information security concerns and review conditions, actual procurement of H20/H200 was almost zero, failing to achieve substantial "exports."
Although the single-card performance of top domestic chips still lags behind the H200, chips like Huawei's Ascend 910B and Cambricon's Siyuan 590 significantly outperform the H20; domestic solutions (such as the Ascend 950PR) are priced at only about a quarter of the H20 (estimated), giving them a clear advantage in the process of domestic substitution.
By the first quarter of 2026, Nvidia's revenue contribution from China had dropped to less than 6%, a year-on-year halving; its market share in China's AI accelerator card market fell to 37%, surpassed by domestic chips.
Currently, the U.S. chip ban on China has become normalized, with strict prevention and blocking of various pathways.
Nvidia's most advanced GB200/B200 and GB300/B300 currently far exceed the native TPP (the restriction metric introduced by BIS), and are directly banned from sale to China.
In terms of pathways, shifting from "geographical control" to "entity penetration," the U.S. Department of Commerce no longer just looks at the destination of chip exports but focuses on reviewing the identity of the end user. Even if transactions occur in third countries like the UAE or Singapore, or if Chinese companies rent computing power in the Middle East by registering subsidiaries overseas, these "detour" behaviors will be defined as violations.
To this day, the path to introducing high-end chips is basically welded shut. Against this backdrop, the cost of overseas chips is basically no longer within the consideration scope of AI companies.
Infrastructure
On August 6, deep in the Ulanqab grasslands of Inner Mongolia, the Ulanqab Xinghe Base, a "carrier" of China's computing-power and electricity synergy invested by Envision Group, was officially put into operation.
This project is the world's largest single intelligent computing center, with a planned capacity of over 2GW and a green electricity share of up to 67% across the entire area. By utilizing the local abundant wind energy resources and dedicated lines to achieve direct green power supply, it has leaped to become the AI infrastructure complex with the highest proportion of green electricity.
On June 17 this year, Envision launched the Mission Gobi initiative in Europe, announcing that by 2030, it will build up to 5GW of green AI computing centers in global Gobi deserts. The Ulanqab Xinghe Base is the first project among them.
In human history, energy utilization has evolved from oil, gas, and coal to electricity, and then to computing power. Jensen Huang categorizes energy as the base of the five-layer AI cake.
Computing centers are undisputed power-guzzling behemoths, forming a clear "competitive relationship" with traditional major power consumers.
According to estimates, in 2025, U.S. computing power electricity consumption was about 180-220 TWh, accounting for approximately 4.3%-5.2% of the total U.S. electricity consumption, making it the sixth-largest power user in the country. Last summer, California suffered from high-temperature heatwaves, overwhelming the power grid. Tesla's AI data center was forced to limit power, resulting in a 30% drop in model training efficiency; wind power fluctuations increased in Texas, and Microsoft's data center could only maintain half of its energy supply for three consecutive days.
On the path to AGI, China and the U.S. face different constraints—the U.S. suffers from energy and power grid infrastructure bottlenecks, while China still lacks high-end chips.
In 2023, China proposed the goal of building a national integrated computing power network, requiring that by the end of 2025, the proportion of green electricity in data centers exceeds 80%, the overall rack utilization rate is no less than 60%, and the average Power Usage Effectiveness (PUE) drops below 1.5.
As the saying goes, "brute force overcomes all." When high-end chips are restricted, the power consumption and infrastructure investment of computing centers become another more decisive cost variable.
According to Jinko Technology's previous investment plan, the total investment for the 1GW computing center in Zhongwei, Ningxia, is as high as 24.5 billion RMB. If fully equipped with high-end Nvidia chips across the stack, the investment intensity would rise to 40-47 billion USD.
Computing costs have become the largest commercialization cost expenditure for AI companies. Guolian Minsheng Securities estimates that, including Doubao's free C-end services, ByteDance's daily cost is as high as 132 million to 240 million RMB. Taking the upper limit, the estimated annual loss approaches 100 billion RMB.
"Burning over 200 million RMB in computing costs every day, yet daily revenue is less than 1 million RMB." This is the dilemma of the current largest national-level AI application, and other AI companies are facing similar challenges.
To solve these problems, ByteDance currently plans to invest 70 billion RMB in Ulanqab to build an AI large model training computing center, which is currently under phased construction; at the same time, it is laying out the Volcengine Intelligent Computing Center in Horinger, with a planned investment of 12 billion RMB, planned as a GW-level green computing power park. The first phase of the project has entered substantive construction.
With the surge in downstream demand, AI computing infrastructure service providers have kicked off a new round of price hikes. The rental prices for AI high-density cabinets have been raised by 10%-20% overall, and liquid-cooled high-power cabinets have seen even higher increases. Computing infrastructure is entering a period of rapid expansion.
Migration
The future of AI is not only a competition of computing power and algorithms but also an energy race.
Tian Lei, Director of the Energy Strategy Center at the Energy Research Institute of the National Development and Reform Commission, once stated this.
In 2025, the total electricity consumption of China's computing centers reached as high as 170 billion kWh, accounting for 1.6% of the total electricity consumption of the whole society. The computing power electricity consumption of the eight major hub nodes of the national integrated computing power network has grown at a rate of nearly 40% over the past three years, far exceeding the average growth rate of total electricity consumption of the whole society.
During the "15th Five-Year Plan" period, the national computing power electricity consumption is expected to increase by over 100 billion kWh annually, reaching 800 billion kWh by 2030, accounting for approximately 6% of the total electricity consumption of the whole society.
Chip procurement is a one-time capital expenditure (CapEx). Although the amount is large, it can be amortized over the next few years; power consumption, however, is a continuous operational expenditure (OpEx). As long as the computing center is turned on, electricity bills will continue to be generated.
According to calculations, generating a 5-second high-definition video with Doubao consumes enough electricity to fully charge 10 smartphones. A 700W AI chip often consumes more than 10 times the energy of a traditional CPU, meaning that the power of just the chips in a 10,000-card cluster can reach up to 7MW.
Adding supporting systems such as cooling, temperature control, and operations, running at full load 24 hours a day consumes at least 200,000 kWh of electricity per day. Over a year, this is equivalent to the annual residential electricity consumption scale of a small county town.
If it is a large 1GW computing center running at full load all year round, the annual power consumption is about 8.7 billion kWh. Adding other auxiliary systems, the total annual power consumption approaches 10 billion kWh.
Even calculated at the cheap Xinjiang computing power electricity price of 0.35 RMB/kWh (theoretical average), the annual electricity bill is about 3.5 billion RMB; adding GPU depreciation and maintenance operating expenses, the annual operating cost will be quite high.
Similar to overseas tech giants, domestic internet giants are also increasing their capital expenditures.
In 2025, ByteDance's full-year capital expenditure exceeded 150 billion RMB, of which 90 billion RMB was used for AI computing power procurement; in 2026, this figure may exceed 200 billion RMB.
Alibaba Group CEO Eddie Wu announced on February 24 last year that over the next three years, more than 380 billion RMB will be invested in building cloud and AI infrastructure. The total amount exceeds the sum of the past decade, also setting a record for the largest investment by a Chinese private enterprise in the relevant field.
In the first half of this year, Tencent's cumulative expenditure reached 84.72 billion RMB, already exceeding the full-year level of 2025. This was mainly used for the training and upgrading of the Hunyuan (Hy) model, inference support for WorkBuddy/CodeBuddy, the WeChat AI ecosystem, and the improvement of cloud computing infrastructure.
As AI infrastructure investment continues to increase, the profits of major companies are inevitably affected. Previously, there were reports that ByteDance's net profit dropped by over 70% in 2025, which was later "debunked" by Li Liang, Vice President of Douyin Group; however, AI infrastructure investment has indeed pulled down ByteDance's profit growth to a certain extent.
Alibaba's net profit (adjusted EBITDA) in the fourth quarter of fiscal year 2026 plummeted by 84%; Tencent, consistently steady, has also seen its cash flow and profits affected by large-scale investments.
When the performance of domestic chips comprehensively surpasses the H20 and the market share breaks through 60% or even higher, the marginal difficulty of acquiring chips is decreasing, while the cost of AI infrastructure is rising significantly.
The cost migration of internet giants and large model companies has actually begun; it's just that these expenditures are underwater and invisible to AI users.