China has set a five-year goal for the expansion of intelligent computing capacity to drive artificial intelligence (AI) that calls for a more than fourfold increase on what was in place at the end of June this year.

To reach the 2030 target while being blocked by the United States from importing the most advanced Western-made AI chips will require innovation, adaptation and the application and coordination of vast resources. Judging by the 177 per cent increase in capacity achieved in the year to June, the target may be achievable.

The additional resources are needed to cope with the growing compute requirements of ever-expanding large language models and the expected staggering proliferation in China of AI agents – digital applications that perform real-world tasks autonomously (see SCMP Plus factsheet).

The biggest current Chinese AI model has 2.8 trillion parameters. Shenzhen-based Huawei Technologies and South China Morning Post owner Alibaba, based in Hangzhou, are among the Chinese tech companies already planning for AI foundation models with 10 trillion parameters. Huawei expects models 10 times bigger than that by 2030. The bigger the model, the more compute necessary for training.

Nearly 5 million AI agents are already in active use in China, a number that will grow by more than 150 per cent annually up to 2030, market researcher the International Data Corporation (IDC) and Jinan, Shandong-based IT infrastructure manufacturer Inspur Information said in a September report. A guideline issued by the State Council, China’s cabinet, calls for 90 per cent-plus penetration of AI agents and next-generation smart computing terminals by 2030.

The proliferation of AI agents is evident in the rapid growth in usage of tokens – the currency of AI inference. Users, whether human or digital, “spend” tokens every time they ask AI to do something. Daily token calls in China have risen from 100 billion a day in January 2024 to 500 trillion a day, Huawei said last month.

Inspur and IDC forecast compound annual growth in token consumption in China of almost 3,500 per cent up to 2030. AI agents that can now work on the same task for hours will probably be handling tasks that take months by 2030, Huawei said.

AI has become a critical battleground in the rivalry between China and the United States. How to achieve higher computing efficiency with limited smart compute will become the core focus of competition in the global AI industry, IDC and Inspur say.

Central to China’s five-year plan to expand intelligent computing capacity is the construction of data centres, optimising AI models to work on domestically produced chips and the expansion and scaling up of supercomputing clusters. None of this can be fully achieved without sufficient electricity.

“Energy is the first principle of AI infrastructure and the binding constraint on how much intelligence the system can produce,” Jensen Huang, chief executive officer of US chipmaker Nvidia, wrote in a blog in March in which he likened AI to a five-layer cake. Applications form the top layer, followed by models, infrastructure and chips, with energy the bottom layer, he said.

Echoing Huang, Zhang Lei, founder and CEO of Shanghai-based green tech company Envision Group, said in May that power systems were becoming the core infrastructure of AI.

When it comes to energy, China starts its smart compute buildout with a big advantage. Two decades of rapid expansion in electricity generating capacity means there are unlikely to be power supply constraints on building AI data centres, unlike in the United States, according to the Brookings Institution, a public-policy think tank.

The National Energy Administration (NEA) projects China’s data centres will consume nearly four times as much electricity in 2030 as they did in 2025. Other estimates are even higher. Much of that demand will be met by wind, hydro and solar power.

In 2024, China’s economic planners proposed that new data centres meet 80 per cent of their power needs from renewable sources by 2030. The target became policy a year later. They’ve got a way to go. In 2023, renewable energy met only 11 per cent of the power needs of China’s data centres.

In June this year, the country’s first AI data centre designed to operate solely on electricity generated by renewable energy – in this case, wind power – began operating in Zhongwei in the Ningxia Hui autonomous region in the northwest.

Most wind and solar power in China is generated in western regions, but consumed in the east. Consequently, the country has built a world-leading network of long-distance ultra-high-voltage (UHV) transmission lines. Capacity for those lines is targeted to grow by nearly a quarter by 2030, with a further 28 UHV lines being built to add to the 39 operating at the end of 2025 (see SCMP Plus Quick Digest).

Furthermore, since 2022 the country has shifted power-intensive computing workloads to western regions under its “east data, west computing” plan. Intelligent software schedules heavy computing tasks for hours when solar and wind power production is forecast to peak.

As part of the government’s “AI Plus” initiative to spur adoption of the technology, China’s top economic planner and the NEA last year announced plans to embed AI in energy infrastructure, with a focus on areas including smart power regulation and renewable energy forecasting. Shanghai electricity grid managers recently achieved record peak-load reduction by smart coordination of computing power demand with available energy.

China already files more than half the world’s patents for AI-led smart grids. Researchers recently completed a decade-long project to develop technology that cuts the time it takes to recover from a blackout to a 10th of a second. Already in 2022, the state-owned power grid company said AI had reduced the blackout recovery time to 3 seconds compared to between six and 10 hours in the pre-AI era.

Meanwhile, there are plans to use quantum computing to prevent blackouts following a successful trial in Anhui province.

Optimising AI across electricity generation, transmission, consumption and storage could actually produce net energy savings – over 15 per cent by 2060 – even as AI’s power needs grow, Hefei University of Technology researchers wrote in the journal Nature Communications Sustainability in May.

China is pursuing technological self-reliance, and that includes developing and producing AI chips. The US has banned the export to China of the most advanced graphics processing units (GPUs) developed by market leaders Nvidia and Advanced Micro Devices in an attempt to stymie the country’s AI development.

Chinese tech giants Huawei, Alibaba, Baidu and Tencent Holdings develop chips for AI model training and inference, as do chip developers Biren Technology, Cambricon Technologies, Enflame, Hygon Information Technology and Moore Threads Technology. Chinese chip suppliers, led by Huawei and Cambricon, are set to capture nearly 80 per cent of the domestic AI server market this year, according to global tech research firm TrendForce.

Huawei recently announced it was speeding up releases of new AI chips. Its Ascend chips are used as alternatives to Nvidia GPUs. However, it faces supply constraints. “We don’t have enough capacity to even satisfy the demand in China,” rotating chairman Eric Xu Zhijun said.

The company earlier unveiled a new chip design paradigm, the Tau Scaling Law, based on its LogicFolding architecture. The development would allow it to match by 2031 the performance of the current cutting-edge processors without using advanced lithography tools whose export to China the US has also banned, Huawei said.

In September, Alibaba unveiled what it called China’s most powerful AI chip, the Zhenwu V900 processor, and touted its performance as being three times that of its predecessor.

AI chips doesn’t function on their own. They fit into a server’s circuit boards alongside storage drives, random access memory, high-speed interconnects and network interface cards. Servers can be grouped together in clusters to provide more compute with greater efficiency.

Since 2023, various players, including Baidu, Alibaba, China Mobile and ZTE, have built smart computing clusters. Employing 10,000 or more GPUs and accelerator chips working as one unit, they function as supercomputers. Alibaba Cloud, Moore Threads, Huawei and MetaX Integrated Circuits have unveiled clusters that can scale to 100,000 cards, another term for chips. Beijing-based Sugon has already built China’s first 100,000-card cluster in Zhengzhou, Henan province, using domestically made AI chips, storage and networking technology. JD Cloud has plans to build a similar-sized computing cluster using Moore Threads processors.

Bigger ones are on the way. Alibaba said a cluster built around its new processor could link up to 500,000 chips. Huawei this month unveiled a “SuperCluster” built with its Ascend neural processing units that it said can scale to more than 1 million cards thanks to an innovative system architecture.

Beijing has made the development of data centres a strategic priority, which should help drive the buildout.

China’s data centre capacity will nearly double by 2030 to 60 gigawatts (GW), with AI-ready facilities accounting for nearly half of the installed capacity, global business intelligence company Rystad Energy expects. US-based research firm SemiAnalysis estimates China’s will have data centre capacity of 24GW by the end of 2026, with a further 20GW in the pipeline and 30GW more in announced projects. By comparison, it projects US data centre capacity will reach 56GW by the end of this year.

The Ministry of Industry and Information Technology has called for cumulative investment of 3.8 trillion yuan (US$530 billion) in information infrastructure between 2026 and 2030 to meet a smart computing capacity target of 9,800 million trillion floating-point operations per second (flops). Goldman Sachs previously forecast compound annual growth of 20 per cent in Chinese data centre demand up to 2028.

Much expenditure will come from China’s hyperscalers – operators of huge, highly scalable data centres. These include state-owned telecoms, Huawei, TikTok and Doubao operator ByteDance, as well as Tencent, Alibaba and Baidu. The latter three more than doubled their capital expenditure year on year to a combined US$20 billion in the second quarter of 2026, SemiAnalysis said.

China has designated eight national computing hubs and 10 data-centre clusters and aims to form a national integrated computing power network. By the end of March a government computing interconnect platform had aggregated 316 million trillion flops of intelligent computing capacity from 155 companies and 578 resource pools, SCMP reported.

China’s most advanced AI models are still being trained using Nvidia chips, sources at major Chinese developers of large language models told the SCMP in August. Complex AI inference tasks like coding also still require Nvidia chips, industry insiders said.

AI companies have found workarounds. For instance, Beijing-based start-ups Moonshot AI and Approaching.ai use Nvidia H20 and Huawei Ascend 910B processors together for inference tasks despite the chips having different software architectures.

Huawei expects that, by next year, a lot of Chinese model training will be based on its Ascend chips, specifically the 950DT processor launched in September. “Pushing full chip self-sufficiency is the certain path forward for China,” Xu, the rotating chairman, said. “I believe it will become a reality sooner or later.”This is already happening. In June, researchers said they had completed post-training for Hangzhou-based research lab DeepSeek’s V4-Pro model using Huawei Ascend 910C chips. In January, Beijing-based firm Zhipu AI’s latest image-generation model was also trained on Huawei chips, becoming the first powerful open-source model developed on an entirely domestic training stack.

This week, DeepSeek open-sourced a suite of core tools tailored for Huawei Ascend AI chips, another step towards ending Chinese dependence on US-made processors. By Wency Chen, Howard Liu