As the brutal price war over AI model tokens in China reaches its logical extreme, with prices approaching zero, Huawei Cloud is attempting to shift the competitive battlefield entirely. At the 2026 Huawei Cloud Inspire conference in Shanghai on June 5, CEO Peter Zhou unveiled the company’s new “Agentic Infra” paradigm, a strategic pivot to provide the foundational computing power and software architecture required to support the next generation of autonomous AI agents.
The move comes after more than a year of intense price competition among China’s tech giants. In May 2024, ByteDance triggered a race to the bottom by pricing its Doubao model at just RMB 0.0008 per 1,000 tokens, following the release of the highly efficient DeepSeek V2. Alibaba, Baidu, Tencent, and iFlytek quickly followed suit, slashing prices to capture market share. The result has been the commoditization of basic inference, making it increasingly difficult to build a sustainable business on token pricing alone.
“What we care about is the health of tokens produced by domestic computing systems, and whether they represent productivity gains rather than just emotional value,” Zhou told the conference, drawing a clear line between Huawei’s infrastructure-first strategy and the volume-focused approach of its competitors.
The “Token Factory” Architecture
Huawei Cloud’s response to the post-price-war landscape is what it calls a “token factory,” a massive AI cluster service designed to maximize the output of its domestic Ascend chips. The infrastructure can scale up to 100,000 accelerator cards, delivering 200 exaflops of computing power.
Crucially for agentic workloads, which require rapid, iterative interactions between multiple AI models and tools, Huawei claims the system achieves latency of less than 10 milliseconds and a throughput of 5 million tokens per second per 1,000 cards, with an availability rate of 99.95%. These specifications are designed to meet the demanding requirements of enterprise agentic applications, where reliability and speed are more important than raw cost per token.
To extract maximum performance from its hardware, Huawei has deployed its CCE Volcano scheduling engine, which the company says improves resource utilization by more than 30%. This software optimization is critical for Chinese cloud providers, who must compensate for the performance gap between domestic Ascend chips and restricted Nvidia hardware through superior software efficiency. The approach mirrors the broader strategy of Chinese AI labs that have achieved competitive performance through software innovation despite hardware constraints.
A Full-Stack Agent Platform
Beyond raw compute, Huawei Cloud introduced a comprehensive suite of platforms designed to accelerate agent development across the enterprise. AgentSphere provides a secure runtime environment for agents with a startup time of just 100 milliseconds, enabling the rapid deployment of complex multi-agent workflows.
The new ModelArts training and inference platform connects to more than 15 advanced models, boasting a routing accuracy of over 95% and reducing average call costs by 20%. For enterprise users, Huawei opened public testing for AgentArts, a dedicated agent development platform, alongside OpenJiuwen, an open-source version that shares over 90% of its architecture with the enterprise edition — a strategy designed to build developer loyalty and ecosystem lock-in.
Huawei is also targeting the booming embodied AI sector with CloudRobo, a development platform specifically for robotics startups. “China has more than 300 embodied intelligence startups, and most are not large,” Zhou noted, highlighting the need for accessible cloud infrastructure to support their development without requiring massive upfront capital investment.
In the healthcare sector, Huawei Cloud has partnered with Ruijin Hospital to develop a pathology large model. The application addresses a critical bottleneck in China’s healthcare system, which has only about 20,000 doctors qualified to read pathology slides, a severe shortage that AI could help alleviate at scale.
Building an Ecosystem Around Ascend
To build momentum for its Agentic Infra strategy, Huawei Cloud announced a collaboration plan that has already attracted over 20 model developers, including prominent startups such as Zhipu AI, DeepSeek, Moonshot AI, and StepFun, as well as established players such as Baidu.
The partnership with DeepSeek is particularly notable. DeepSeek’s R1 and V3 models have been deployed on Huawei’s Ascend CloudMatrix 384 supernode, where inference efficiency reportedly matches that of Nvidia’s H800 chips. This collaboration demonstrates that domestic hardware can support frontier-level models when paired with highly optimized software, and it provides a powerful marketing proof point for Huawei’s cloud ambitions.
By positioning itself as “the most open cloud in the agent era,” Huawei is seeking to build an ecosystem resilient to US export controls and capable of supporting the next phase of AI commercialization. As the token price war fades into the background, the competition among China’s cloud providers will increasingly be defined by who can offer the most robust, reliable, and efficient infrastructure for the agentic future, and Huawei is making a strong early claim to that position.