The "computing power war" among chip giants has extended to the power sector.

From November 5 to 10, the 9th China International Import Expo was held at the National Exhibition and Convention Center (Shanghai).

This edition of the CIIE attracted 4,142 enterprises from 148 countries and regions, with the exhibition area for enterprise displays exceeding 370,000 square meters for the first time. The Technical Equipment Exhibition Zone covers all links of the industrial chain including semiconductor materials, equipment, design and finished products, with artificial intelligence, green intelligent computing and intelligent terminals as key display directions.

On the exhibition booths, various AI products and solutions ranging from AI PCs, intelligent robots to data center infrastructure became the core display highlights. For example, Qualcomm demonstrated its intelligent terminal ecosystem covering AI smartphones, PCs and robots; Siemens introduced industrial AI and humanoid robots into flexible production scenarios; Schneider Electric brought AI data center solutions for ten-thousand-card clusters, covering modular deployment, high-density power supply and liquid cooling technologies.

As AI technology accelerates its penetration into industrial applications, the integration of artificial intelligence with sectors including industrial manufacturing, energy and healthcare has further deepened, and enterprises have begun to pay more attention to the practical application and commercial value of AI technologies.

In Intel's exhibition zone, green computing power cabins, enterprise-grade token all-in-one machines and agent PCs became the focus. Wang Zhicong, Chairman of Intel China, proposed that the company hopes to explore the transformation path from green power to AI Token through the combination of computing power and new energy.

"In the past, customers might be more concerned about what AI can achieve, but now more and more customers are starting to ask about the cost of deploying a set of AI systems, whether it can run stably, and how long it will take to recoup the investment," a participant from a technology enterprise told Tencent Technology. As AI enters more actual business scenarios, enterprises' focus on technology is gradually shifting from capability demonstration to application effect and input-output ratio.

For AI to be truly implemented, it relies not only on technology, but also on solving problems in power supply, computing power cost and actual benefits. From terminal equipment to industrial production, and then to computing power infrastructure, global technology enterprises are launching a new round of layout centered on large-scale AI applications.

01 When Power Supply and Cost Become the Core

In the past few years, competition in the AI industry has mainly focused on model performance and computing power scale.

Large technology companies have continued to invest in GPUs and build data centers to support increasingly large-scale model training and inference tasks.

However, as AI accelerates into the practical application stage, the growing demand for computing power has also exposed new infrastructure problems.

According to the 2026 updated forecast from the International Energy Agency, global data center power consumption will reach about 485 terawatt-hours in 2025, and may hit 950 terawatt-hours by 2030, nearly doubling. Among them, the power demand of AI data centers is growing at a faster rate.

Training large models requires concentrated investment of a large amount of computing resources, while inference services need to run continuously according to user requests. With the development of applications such as AI programming, content generation and enterprise agents, the scale, usage frequency and deployment locations of inference workloads have become more diverse.

For enterprises, the cost of using AI is no longer only the expense of purchasing computing power or invoking models, but also needs to take response speed, data security, system maintenance and resource utilization into consideration.

This makes the cost of computing power a more complex issue.

For example, building data centers in regions rich in new energy resources may bring certain advantages in electricity prices, but network latency, energy storage cost and computing power utilization also need to be considered.

On the contrary, deploying inference devices inside enterprises may reduce part of the cloud invocation costs, but enterprises need to bear the costs of hardware depreciation, operation and maintenance, and model upgrading.

Therefore, the core problem faced by the AI industry is expanding from how to obtain more computing power to how to make computing power operate more economically and stably.

02 The Competition Logic Is Changing

Focusing on the above problems, global chip enterprises, cloud service providers and data center operators are exploring different technical routes and business models.

First, as computing power becomes increasingly powerful, the power supply configuration must keep pace.

In 2026, enterprises including NVIDIA, Google and Microsoft further promoted the 800V high-voltage DC power supply architecture, and formulated relevant technical specifications together with industry partners through the Open Compute Project (OCP).

At the CIIE, Intel also demonstrated the integrated green power energy storage and computing power solution adopting this power supply architecture, and introduced technologies such as solid-state transformers (SST) in cooperation with partners to explore power supply modes for high-density AI computing facilities.

Compared with traditional low-voltage power distribution, the 800V DC architecture can reduce transmission current under the same power, lower the demand for copper cables and power distribution space, and reduce power conversion links with appropriate system design.

This means that AI chip manufacturers no longer only need to consider the computing capability of processors, but also the overall efficiency from the data center power supply system to the computer racks.

Second, AI computing power is no longer only concentrated in large data centers, but also starts to enter enterprise and personal devices.

Large cloud data centers still undertake high-intensity model training, complex inference and large-scale concurrent tasks. However, with the improvement of open-source model capabilities, part of the computing tasks have met the conditions to run on local enterprise and terminal devices.

For example, workloads such as internal enterprise document processing, code assistance and speech recognition do not necessarily need to invoke remote large models.

But there is no absolute advantage or disadvantage between different deployment methods.

The cloud end is suitable for tasks with large scale, obvious fluctuation or requiring continuous update of cutting-edge models; local computing is more suitable for scenarios with stable demand, sensitive data or requirements for offline operation.

In the future, AI infrastructure will probably form a pattern where large cloud ends, regional edge nodes, enterprise local ends and personal terminals coexist.

Third, chip competition not only focuses on performance, but also on software and application experience.

As AI agents gradually shift from chat tools to task execution systems, processors need to cope with more complex workloads.

For example, an agent may need to complete speech recognition, file retrieval, model inference and cross-application operations at the same time. These tasks have different requirements for computing performance, memory bandwidth, power consumption and system stability.

Therefore, the task allocation among CPUs, GPUs and NPUs, as well as the adaptation of underlying software to models and applications, are becoming important factors affecting user experience.

This also explains why chip enterprises including Intel, Qualcomm, AMD and Apple continue to invest around terminal AI computing and developer ecosystems.

In the next stage of competition, apart from checking which chip has higher AI computing power indicators, it will also depend on who can convert computing capability into practical application experience more effectively.

03 "New Computing Power Business"

As AI computing demand extends to different scenarios, major chip manufacturers have also started to re-layout, looking for new opportunities at different levels from computing power centers, enterprise local deployment to personal terminals.

Taking Intel as an example, at this CIIE, Intel demonstrated three tiers of AI computing solutions: the green computing power cabin for large data centers and enterprise parks, the token all-in-one machine for small and medium-sized enterprises, and the agent PC for individual users.

On the exhibition booth, Wang Zhicong, Chairman of Intel China, focused on introducing the integrated green power energy storage and computing power solution.

"We have observed that there is a large amount of surplus new energy in western China, and the actual utilization rate of some of this new energy is not very high," Wang Zhicong said. Intel hopes to deploy modular computing power facilities to new energy power generation areas, carry out AI computing on site using green power, and then provide Token services to the outside through the network.

He summarized this idea as "From Green Power to Green Token".

According to on-site experts, this solution tries to solve the contradiction between new energy power generation fluctuation and stable data center power supply through long-term and short-term energy storage, 800V high-voltage DC power supply and modular computing power cabin design.

This layout reflects that competition in AI infrastructure is extending from chip performance to energy efficiency and system deployment capability. However, rich green power resources do not necessarily mean lower inference costs, and energy storage investment, network latency and computing power utilization are still the key factors determining its economy.

In the enterprise market, Intel hopes to promote part of AI inference tasks to shift from the cloud to the local end. The token all-in-one machine demonstrated this time can run open-source large models in ordinary offices, supporting workloads such as AI programming, document processing and video generation.

Taking short drama production companies as an example, Wang Zhicong believes that some small and medium-sized enterprises can also meet their daily AI computing needs through local devices.

Behind this is enterprises' re-evaluation of AI usage cost and data security. For scenarios with high usage frequency and relatively fixed tasks, local deployment may have certain advantages, but enterprises still need to bear the costs of hardware depreciation, model maintenance and upgrading.

In the personal terminal market, Intel is promoting the agent PC ecosystem through cooperation with Chinese internet enterprises.

Wang Zhicong put forward two key directions on site:

The first is end-cloud collaborative deployment, and the second is deep integration of multi-agents and specialized models.

Focusing on this idea, Intel stated that it is cooperating with enterprises including Tencent, ByteDance, Alibaba and Baidu to integrate local model operation and heterogeneous computing capabilities of CPUs, GPUs and NPUs into existing agent applications.

Compared with simply emphasizing AI computing power indicators, this model pays more attention to how chip capabilities can be integrated into actual work processes. However, whether agents can generate new PC replacement demand still depends on the actual application experience, rather than how many AI functions are integrated.

As the underlying support for the above layout, Intel also demonstrated PC and server processors based on Intel 18A advanced process, including Panther Lake, Wildcat Lake and Clearwater Forest.

This also reflects Intel's current two business paths: on the one hand, it continues to promote advanced processes and chip architectures, and on the other hand, it participates in the broader AI computing market through ecological cooperation with energy, power, server and internet enterprises.

For chip manufacturers, the expansion of AI computing demand to different scenarios means new market opportunities, and also brings more complex competition. In the future, the ability to provide more efficient and economical computing solutions and promote their implementation in actual scenarios will become the key to the next stage of competition.

This article is from "Tencent Technology", written by Li Hailun, edited by Xu Qingyang, and published by 36Kr with authorization.