EN / 中文

AI Agents Reshape Infrastructure: openEuler’s Xiong Wei on OS Evolution and Token Economy

by zhongguodianzibao·November 19, 2026

Artificial intelligence is triggering a new round of infrastructure transformation. Recently, during the Agentic Infra Operating System Conference, Xiong Wei, Chairman of the OpenAtom openEuler Committee, accepted interviews from media outlets including China Electronic News. In his view, the changes brought by artificial intelligence to the basic software industry are not limited to the technical level; they will also profoundly impact the industrial ecosystem, market participants, and the coupling process between hardware and software industries.

"Operating Systems Have Not Undergone Disruptive Changes"

Since the beginning of this year, with the accelerated deployment of agents, concepts such as AI Infra, Agentic Infra, and Agent OS have emerged in large numbers. In the era of artificial intelligence, does it require reinventing operating systems?

Xiong Wei believes that at least at the current stage, operating systems have not undergone disruptive changes but continue to evolve based on existing capabilities. He specifically pointed out that many people tend to strongly bind or link Agents with operating systems. In reality, above the hardware layer, there are still operating systems, AI software stacks, etc. "There are at least one or two layers between an Agent and the operating system."

However, Xiong Wei also stated that as a watershed product in human history, the value of intelligent agents still needs to be understood. Like its initial Chinese translation "proxy," if the main function of large models is to accept Tokens and output thoughts, then intelligent agents act as "proxies" connecting the brain and limbs, transforming the content output by large models into specific, executable actions. With the enhancement of generalization capabilities of large models, the combination of the two creates conditions for the scaled implementation of artificial intelligence.

Xiong Wei gave an example of sending emails: In the past, sending 300 different emails required a large amount of repetitive operations; even with programs, parameter settings were needed. Large models can generate a large number of instructions at once and call intelligent agents to execute them. "You will find that the human world seen through the eyes of a large model is like super slow-motion action," he said. "The entire software stack from top to bottom will undergo drastic changes because the speed of large models is thousands or tens of thousands of times faster than human speed."

Taking sandbox technology required for agent operation as an example, currently, industry attention to sandboxes focuses more on security. However, in Xiong Wei's view, the sandbox itself is not a new technology; the underlying virtualization technology has long existed. The real challenge facing sandboxes in the era of intelligent agents is how to achieve further lightweighting while retaining original virtual machine security mechanisms to adapt to the speed at which large models issue instructions. Similar issues exist in scheduling systems. Traditional scheduling mechanisms struggle to meet the demand for high-frequency instruction issuance by large models, easily leading to a situation where "hands cannot keep up with the brain."

"In the Future, Token Vendors May Be Everywhere"

The artificial intelligence industrial ecosystem is expanding rapidly. Will the basic software ecosystem also become increasingly complex? Xiong Wei offered a different judgment: In the AI era, the expansion of the commercial ecosystem and the simplification of the software development ecosystem may occur simultaneously.

In the traditional cloud computing era, cloud vendors needed to provide rich software services for various business needs; the more service types, the stronger customer stickiness often became. However, with the rise of the Token economy, although the forms of upper-layer applications are becoming richer, the underlying infrastructure requirements are converging to some extent, namely serving Token production. The complexity of the software stack is likely to decrease significantly accordingly.

This change has already shown signs in the practice of the openEuler open-source community. According to Xiong Wei, openEuler has currently built key technical solutions for sandboxes, KV Cache, memory systems, and Agent security, covering multiple critical links in the current stage of agent operation.

Xiong Wei further stated that the simplification of the software stack may bring about two completely different industrial changes: First, the number of vendors may converge; Second, as the technical threshold for Token production lowers, more vendors who previously lacked cloud computing business experience will have the opportunity to build their own production systems. "In the future, Token vendors may be everywhere."

At the same time, being able to produce Tokens does not mean being able to produce them at low cost. In Xiong Wei's view, industrial competition under the Token economy has two important dimensions: One is improving the intelligence level of the large model itself, and the other is reducing the cost of supporting large model operation. This puts forward new requirements for the direction of basic software innovation.

Taking the openEuler KV Cache solution as an example, Xiong Wei introduced that by reducing redundant calculations during large model inference, every 1% improvement in cache hit rate could reduce Token production costs by 30%. For vendors with similar Token production capabilities, this will directly affect the price competitiveness of their products.

"Build the Shell First, Then Do the Interior Decoration"

Beyond changes in business software, Xiong Wei stated that as a root community, another core problem openEuler aims to solve is at the hardware layer, with supernodes being a major direction he particularly favors.

In Xiong Wei's view, supernodes have become the mainstream form of infrastructure in the Agentic era but will not completely replace traditional servers. In the future, supernodes may primarily carry critical, high-value businesses, while general businesses will still be carried by standard servers. This means that operating systems need to face the coexistence of traditional hardware and supernodes in the long term, requiring compatibility with different hardware architectures and support for building new software ecosystems.

How to fully unleash the performance advantages of supernodes while ensuring software adaptation is another challenge facing the industry. "Supporting hardware is not so difficult, but truly using it well is hard," Xiong Wei said. Although traditional software can run on supernodes via intermediate software isolation layers, it may not achieve the expected cost-performance ratio. Software needs to consider the characteristics of the supernode architecture from the design phase to fully release the performance of new hardware, which he calls a "supernode-native" software ecosystem.

Xiong Wei shared that openEuler aims to bridge hardware gaps to ensure southbound compatibility on one hand, and support new software industrial ecosystems on the other. However, compared to the gradually clarifying hardware evolution direction, changes on the northbound application side remain highly uncertain. Xiong Wei believes that at the current stage, the industry should not rush to define "big frameworks, big stories, or big concepts," but should start from actual business needs, gradually building and perfecting the software ecosystem by solving specific technical problems. "When you solve problems to a certain degree and look back, you will find that your building, the shell, is ready, and what remains is just interior decoration for businesses and scenarios."

It is reported that openEuler solutions have already demonstrated application potential in some high-value business scenarios. During this year's "618" and "Double 11" shopping festivals, the community collaborated with JD.com to support the large-scale application of supernodes, resulting in a 15% reduction in latency for related businesses.

Author | Chen Cun

Editor | Wu Lilin

Art Editor | Ma Liya Supervisor | Lian Xiaodong