Huawei Cloud does not engage in token price wars; Zhou Yuefeng wants to give AI Cloud a new way to win.

“I don’t really care how many tokens there are in total, and I don’t really care how much total revenue there is.” At the 2026 Huawei Cloud INSPIRE Innovators Conference held on June 5, Huawei company director and Huawei Cloud CEO Zhou Yuefeng’s first media interview since taking office clearly and explicitly conveyed Huawei Cloud’s current strategic focus.

This is a rare stance in China’s current AI cloud market.

Over the past six months, cloud providers represented by Alibaba Cloud and Volcano Engine have continuously emphasized the AI cloud narrative, using daily token call volume and MaaS revenue scale as new growth anchors. Even large model vendors such as The Dark Side of the Moon, DeepSeek, and Zhipu have kept pushing down inference prices. Across the entire industry, the key words are model call volume and scale.

Huawei Cloud chose to enter this crowded battlefield in another way. Huawei Cloud simultaneously released a batch of the most AI-intensive new products since last year: AICS Lingqu Intelligent Computing Cluster, AMS Agentic Memory Storage, CCE Volcano Next integrated scheduling engine, AgentSphere secure autonomous operating foundation, and also ModelArts Next and the enterprise-level intelligent agent platform AgentArts (open-source version openJiuwen). It also packaged these into a new paradigm called “Agentic Infra.”

The KPI defined by Zhou Yuefeng for Huawei Cloud is not the number of tokens, but whether “each token behind it truly improves productivity.” In a window period when domestically produced computing power supply is still constrained and business models are still being reshaped, Huawei Cloud has pulled itself out of the “fight for second place in AI cloud.”

Not competing on token scale

At the meet-and-greet event, Zhou Yuefeng responded, unusually directly, to the differences between Huawei Cloud and Alibaba Cloud and Volcano Engine. He said there are three reasons.

First is the computing power route. Huawei Cloud uses a fully domestically developed computing power stack of both hardware and software—Ascend, Kunpeng, CANN, Euler, and an entire self-developed system. This path is more winding because Huawei has no way to use someone else’s computing power; it can only turn domestication into an industry-level answer.

As a result, Huawei Cloud must build a second computing power plane—offering a different ecosystem choice beyond the globally dominant computing power path formed by NVIDIA plus mainstream public clouds. Huawei Cloud is unable to, and does not intend to, “reconcile accounts” with competitors on computing power scale using “universal” hardware. Zhou Yuefeng said, “I don’t want to compare revenue or whether we’re second, third, or fourth in scale with other cloud companies. That’s meaningless.”

Second is the business focus. Internet-based cloud providers naturally depend on C-end traffic and developer ecosystems, while Huawei Cloud places heavy resources on government and enterprise, and on industries related to the country’s economy and people’s livelihoods. For example, Huawei’s hybrid cloud has ranked first for years in market share across government, finance, and central/State-owned enterprise segments, serving more than 5,500 global customers.

Zhou Yuefeng said that the iteration speed of models and computing power is so fast that it may be outdated by the time it is deployed. Therefore, he advised government and enterprise customers not to build massive “ten-thousand-card” clusters themselves, but instead use local data plus remote public cloud AI computing power/model services, combined with technologies such as confidential inference, confidential training, and confidential computing—so that a balance can be formed between data sovereignty and shared computing power. Essentially, this means delivering the iteration dividend of public cloud to customers that cannot fully move to public cloud.

Third is the ecosystem strategy. Huawei Cloud has executed open source very thoroughly. Ascend CANN, the Euler operating system, CCE Volcano scheduling, and the ModelArts toolchain are all open sourced; the open-source version openJiuwen of the intelligent agent platform AgentArts shares more than 90% code similarity with the commercial version.

At the conference, it also jointly launched the “One Hundred Models, Thousand Varieties—Cloud Gathering for Win-Win” initiative together with more than 20 leading model vendors, including Zhipu, DeepSeek, MiniMax, Kimi, Jiayue Xingchen, Baidu, Meituan LongCat, and iFlytek Spark.

When domestically produced computing power remains limited in capability and supply, the larger the ecosystem is spread and the more model choices there are, the second computing power plane can stand firm.

Agentic Infra: moving the battlefield from selling tokens to selling productivity

If the computing power route determines what Huawei Cloud “doesn’t compete on,” then Agentic Infra determines what it “wants to compete on.”

Zhou Yuefeng put forward an assessment of how the AI industry is evolving: four years ago, building AI meant buying computing cards; three years ago, it was about training large models; this year, it is about using intelligent agents. Computing power and models are receding behind the scenes, while intelligent agents are stepping onto the stage.

The competitive focus of AI cloud is shifting from token throughput to whether intelligent agents can truly run in enterprises.

Huawei Cloud’s product matrix is also rearranged based on this judgment. The “four-piece set” of Agentic Infra—high-efficiency token factories, continuous learning, intelligent one-piece scheduling, and secure autonomous operation—maps to the engineering problems that enterprises inevitably face when deploying intelligent agents.

AICS Lingqu compresses token latency for a 100,000-card cluster to within 10 milliseconds; AMS uses NPU direct access to CMS to build PB-scale memory space, solving the long-range task memory bottleneck for Agents; CCE Volcano Next improves resource utilization by more than 30% through joint pools for training and inference; AgentSphere uses a lightweight sandbox to achieve 100-millisecond-level startup and creation at a scale of 100,000 per minute.

ModelArts Next also reshapes the playbook of MaaS. Its model routing supports three strategies: cost-priority, effect-priority, and balanced routing. It has already integrated more than 15 SOTA models, with a scheduling accuracy exceeding 95%, and it has reduced average calling costs by 20%.

But Huawei Cloud’s truly differentiated bet is on industry zones. At this conference, Huawei Cloud launched, all at once, four “industry AI dream factories” zones: smart healthcare, embodied intelligence, intelligent manufacturing, and scientific computing.

Among them, in the Smart Healthcare zone, the RuiPath large model jointly built with Shanghai Ruijin Hospital has brought more than 20 top-tier, prefecture-level, and county-level hospitals from Handan, Rui’an, Qianxinan, and Wuan to participate. This means that capabilities that highly depend on expert experience—such as pathology diagnosis—are being scaled to county hospitals in a “cloud service” form for the first time.

The Embodied Intelligence zone launched CloudRobo, the world’s first full-process embodied intelligence development platform, aiming to carry end-to-end tool-chain needs for more than 300 Chinese embodied intelligence startup companies.

Zhou Yuefeng said that healthcare and finance are China’s most digitally mature and data-rich industries. “If we can’t make AI work in these industries, it will be even harder in other industries.” In these fields, the yardstick for measuring AI value should not be daily active users or token counts, but rather the proportion of financial risk prevention, improvements in credit efficiency, and the probability that remote patients receive accurate diagnoses.

Linking these clues together makes Huawei Cloud’s strategic outline clear: using fully domestically produced computing power plus an open-source ecosystem as the foundation; covering government and enterprises with hybrid cloud and confidential computing; and shifting competition from “selling tokens” to “selling productivity” through Agentic Infra plus industry zones.

This path is much slower than pursuing MaaS revenue and it is also harder to tell a beautiful year-over-year comparison figure, but it sidesteps the most intense price-red-ocean in today’s AI cloud. It bets on a market that has not yet been priced. Once intelligent agents truly enter the industry, whoever can secure the position of the underlying infrastructure will win.

In this AI cloud race track, Huawei Cloud can only take another solution. Zhou Yuefeng concluded, “I have no way to build a silicon-based ‘black land’ like those universal brands.” While other cloud providers compete on whose token cost-effectiveness is higher, Huawei Cloud is competing on whether this domestically produced computing power system can meet the truly real future needs of China’s industrial AI.

(Author | Zhang Shuai, Editor | Yang Lin)