Huawei Cloud Avoids Token Price War, Focuses on Productivity with New AI Infrastructure

"I don’t really care about the total supply of tokens or the total revenue." At the Huawei Cloud INSPIRE Innovators Conference on June 5, Zhou Yuefeng, Director of Huawei and CEO of Huawei Cloud, delivered his first media interview since taking office, clearly and unequivocally conveying Huawei Cloud’s current strategic focus.

This is a rare statement in today's Chinese AI cloud market.

Over the past six months, cloud providers such as Alibaba Cloud and Volcano Engine have consistently emphasized the narrative of AI cloud, using daily token invocation volume and MaaS revenue scale as new growth metrics. Even large model providers like Moonshot AI, DeepSeek, and Zhipu have continuously lowered their inference pricing, making model invocation volume and scale the industry’s key focus.

Huawei Cloud has chosen a different approach to enter this crowded battlefield. Huawei Cloud has launched its most concentrated batch of AI-focused new products since last year: the AICS Lingqu AI Computing Cluster, AMS Agentic Memory Storage, CCE Volcano Next Unified Intelligence and Computing Scheduling Engine, AgentSphere Secure Autonomous Runtime Foundation, as well as ModelArts Next and the enterprise-grade agent platform AgentArts (open-source version: openJiuwen), bundled together under the new paradigm of "Agentic Infra."

Zhou Yuefeng defined Huawei Cloud’s KPI not by the number of tokens, but by whether each token truly enhanced productivity. During this window of limited domestic computing power supply and ongoing business model transformation, Huawei Cloud has stepped away from the race to become the second-largest AI cloud provider.

At the meeting, Zhou Yuefeng unusually addressed the differences with Alibaba Cloud and Volcano Engine directly. He said that Huawei Cloud is different from other cloud providers for three reasons.

First, the computing power approach is different. Huawei Cloud uses a fully domestic computing hardware and software stack, including an integrated suite of self-developed technologies such as Ascend, Kunpeng, CANN, and Euler. This path is more challenging because Huawei cannot rely on external computing resources; it must turn domestication into an industrial-scale solution.

Therefore, Huawei Cloud must build a second computing plane, offering an alternative ecosystem choice beyond the globally dominant computing path formed by NVIDIA and major public clouds. Huawei Cloud neither can nor intends to compete with competitors on computing scale using a patchwork of hardware from various vendors. Zhou Yuefeng said, "I have no interest in competing with other cloud companies to see who ranks second, third, or lower in revenue or scale—it’s meaningless."

Second, their business focus differs. Internet-based cloud providers naturally rely on C-end traffic and developer ecosystems, whereas Huawei Cloud heavily prioritizes the government, enterprise, and critical national infrastructure sectors. For instance, Huawei’s hybrid cloud has ranked first in market share for government, finance, and central and state-owned enterprises for multiple consecutive years, serving over 5,500 customers worldwide.

Zhou Yuefeng stated that the pace of model and compute iteration is so fast that systems may already be outdated by the time they are deployed. He therefore recommends that government and enterprise customers avoid building their own ten-thousand-GPU clusters, and instead adopt a hybrid approach: keeping local data while leveraging remote public cloud AI compute and model services, combined with technologies such as confidential inference, confidential training, and confidential computing, to achieve a balance between data sovereignty and compute sharing. Essentially, this approach delivers the public cloud’s iteration advantages to customers who cannot fully migrate to the public cloud.

Third, the ecosystem strategy is different. Huawei Cloud has thoroughly open-sourced its technologies, including Ascend CANN, Euler OS, CCE Volcano scheduling, and ModelArts toolchain. The open-source version of the agent platform, openJiuwen, shares over 90% code similarity with its commercial counterpart.

The meeting also brought together more than 20 leading model providers—including Zhipu, DeepSeek, MiniMax, Kimi, Step星辰, Baidu, Meituan LongCat, and iFlytek Spark—to launch the "Hundred Models, Thousand Forms, Cloud Collaboration for Mutual Success" initiative.

When domestic computing power remains limited in capability and supply, only by expanding the ecosystem and increasing model options can the second computing plane be firmly established.

If the computing power strategy determines what Huawei Cloud won't compete in, Agentic Infra determines what it aims to compete in.

Zhou Yuefeng offered a perspective on the evolution of the AI industry: four years ago, building AI meant buying compute cards; three years ago, it meant training large models; this year, it’s about using agents. Compute and models are receding into the background, while agents are stepping into the spotlight.

The focus of AI cloud competition has shifted from token throughput to whether agents can truly operate effectively within enterprises.

Huawei Cloud’s product portfolio has also been reorganized based on this assessment. The "four essentials" of Agentic Infra—efficient Token factory, continuous learning, integrated intelligence and scheduling, and secure autonomy—each address critical engineering challenges enterprises face when deploying agents.

AICS Lingqu reduces token latency for a 100,000-GPU cluster to under 10 milliseconds; AMS achieves petabyte-scale memory space via NPU direct passthrough to CMS, resolving long-term memory bottlenecks for agents; CCE Volcano Next improves resource utilization by over 30% through shared pools for training and inference; AgentSphere enables sub-100-millisecond startup and tens of thousands of batch creations per minute using lightweight sandboxes.

ModelArts Next has reimagined the MaaS experience, offering model routing with three strategies—cost-optimized, performance-optimized, and balanced. It has integrated over 15 SOTA models, achieving over 95% scheduling accuracy and reducing average invocation costs by 20%.

However, Huawei Cloud's true differentiating bet is its industry-specific zones. At this event, Huawei Cloud launched four new "Industry AI Dream Factories" at once: Smart Healthcare, Embodied Intelligence, Intelligent Manufacturing, and Scientific Computing.

The Smart Healthcare zone, in collaboration with Shanghai Ruijin Hospital, has developed the RuiPath large model, with over 20 tertiary, municipal, and county-level hospitals—including Handan, Ruian, Qianxinan, and Wu’an—collectively onboarded. This marks the first large-scale delivery of pathology diagnosis, a capability heavily reliant on expert experience, in the form of "cloud services" to county-level hospitals.

The Embodied Intelligence section introduces CloudRobo, the world’s first end-to-end embodied intelligence development platform, designed to meet the full-stack tooling needs of over 300 embodied intelligence startups in China.

Zhou Yuefeng stated that healthcare and finance are China’s most mature and data-rich industries in digitalization. "If AI cannot succeed in these sectors, it will be even harder in others." In these fields, the measure of AI’s value should not be daily active users or token counts, but rather the proportion of financial risk mitigation, improvements in credit efficiency, and the probability of accurate diagnoses for patients in remote areas.

Connecting these dots reveals Huawei Cloud’s strategic outline: building on a fully domestic compute infrastructure and open-source ecosystem, covering government and enterprise markets with hybrid cloud and confidential computing, and shifting the competitive focus from “selling tokens” to “selling productivity” through agentic infrastructure and industry-specific zones.

This path is much slower and harder to present with impressive year-over-year metrics than chasing MaaS revenue, but it avoids the fiercest price wars in today’s AI cloud market, instead betting on an undervalued market—the position of those who control the underlying infrastructure once agents truly enter industry.

In the AI cloud赛道, Huawei Cloud can only take a different approach. Zhou Yuefeng summarized: "I cannot build a multinational silicon-based black soil." While other cloud providers compete over who offers better token value, Huawei Cloud is focused on whether this domestic computing infrastructure can truly meet China’s industrial AI needs in the future. (Author: Zhang Shuai; Editor: Yang Lin)