At Huawei Connect 2025 in Shanghai, Huawei Cloud detailed its latest developments in artificial intelligence computing and cloud infrastructure. In a keynote speech, Zhang Ping’an, executive director of the board and chief executive officer of Huawei Cloud, outlined new products and services aimed at expanding compute capacity, advancing AI models, and supporting enterprise adoption of AI-driven technologies.
Huawei Cloud introduced an upgraded AI Compute Service powered by CloudMatrix384. The company said its supernode specifications will expand from 384 cards to 8,192, with support for hyperscale clusters ranging from 500,000 to one million cards. Huawei also launched an Elastic Memory Service designed to reduce latency in multi-round interactions with foundation models. To support these capabilities, Huawei Cloud has built liquid-cooled data centers in Guizhou, Inner Mongolia, and Anhui, with each cabinet capable of dissipating 80 kilowatts of heat and operating at a power usage effectiveness ratio of 1.1.
The company formally launched its AI Token Service, designed to simplify access to inference computing by abstracting underlying complexity and pooling compute, memory, and storage resources. According to Huawei Cloud, CloudMatrix384 delivers three to four times the average inference performance per card compared with its H20 system.
Huawei also highlighted its Pangu Models, which it continues to develop for both open-source and closed-source applications. The models have been applied in more than 500 use cases across over 30 industries, including finance, healthcare, government services, and transportation. The company said it intends to expand investment in Pangu to help enterprises train and deploy industry-specific AI models.
In robotics, Huawei Cloud announced the CloudRobo Embodied AI Platform, which deploys computational tasks to the cloud to enable lighter and more adaptable robots. To standardize connectivity between robots and cloud infrastructure, Huawei also introduced the Robot to Cloud (R2C) Protocol, with 20 partners already onboard.
Other product updates included an expansion of Kunpeng Cloud Services, where the number of ARM-based compute cores on Huawei Cloud increased from 9 million to 15 million in the past year. The company said its Kunpeng platform is now compatible with more than 25,000 applications. Huawei also promoted its GaussDB databases, which it claims can process 5.4 million transactions per minute using supernode-based clusters, more than doubling the performance of non-supernode deployments.
Huawei Cloud further presented its distributed cloud solution, which integrates services across central, regional, and edge sites under offerings branded CloudOcean, CloudSea, CloudLake, and CloudPond. In addition, the company introduced Versatile, a platform for building and running enterprise-grade AI agents, which Huawei said reduces the complexity of generating custom agents for business applications.
