Huawei presents its Atlas SuperPoDs portfolio — the liquid-cooled Atlas 950 SuperPoD and the air-cooled Atlas 850E SuperPoD—alongside a powerful open-source software stack that pioneers the AI Infra 2.0 reference architecture. The announcements reaffirm Huawei's commitment to delivering scalable, stable, and efficient computing infrastructure for the Agentic AI. Mr. Hu Zhengce, Vice President, ICT MSSD Regional Support Center (Middle East & Central Asia), said, "As large models advance toward a scale of one hundred trillion parameters, they will accelerate the transition of AI Infra into the 2.0 era. SuperPoD is an inevitable choice for the construction of ultra-large-scale AI infrastructure. Huawei introduces a leading AI Infra solution that integrates Huawei's proven ‘electricity-compute-network-storage-cloud’ capabilities. To ensure out-of-the-box usability, we also provide an open-source software stack tailored for training, inference, and Agent scenarios, accelerating intelligent innovation across industries. " Targeting large-scale data center construction, trillion-parameter model training, and high-concurrency central inference scenarios, the Atlas 950 SuperPoD integrates a maximum of 1,024 NPU cards per SuperPoD. Through systematic architectural optimization, the entire SuperPoD cluster operates as one unified computer. Designed for traditional enterprise data centers, the Atlas 850E and Atlas 650E deploys directly into standard racks without requiring liquid-cooling retrofitting. This brings SuperPoD capabilities into existing air-cooled DC facilities. Atlas 950 SuperPoD, Atlas 850E and Atlas 650E stably deliver ms-level inference latency, perfectly matching the workloads of multi-turn Agent conversations and high-frequency KV cache read/write operations. By enabling highly efficient multi-card parallel inference and memory resource pooling, they support million-level ultra-long context windows, providing efficient and reliable computing power for AI application scenarios across industries such as internet, telecom, finance, government, and education. Unifying more than ten interconnection protocols, Unified Bus protocol increases interconnection bandwidth to TB level, compresses RTT communication latency to microseconds, and supports a unified global memory addressing space within the SuperPoD. Huawei has also introduced the SuperPoD architecture into general-purpose computing systems. Powered by Unified Bus, the general-purpose computing SuperPoD supports scaling up to thousands of nodes. This enables higher-density sandboxes for Agents, faster sandbox boot speeds, and improved Agent vector retrieval performance. Meanwhile, a multi-tier, high-bandwidth, and large-capacity storage system is provided to enable all types of KV caches to be reached in a single hop. On top of this hardware infrastructure, a comprehensive, open-source software stack powers the entire lifecycle of AI training, inference, and Agent development—all anchored by the openEuler SuperPoD OS. This foundational layer unlocks the full potential of SuperPoD hardware through microsecond-level global scheduling, dynamic resource allocation, and ultra-high-performance communication. To accelerate Agentic innovation, the stack coordinates critical components like sandbox runtimes, Kubernetes, data services, the Witty RAG Core, and the Agent framework into a complete development suite. Simultaneously, it streamlines AI training and inference deployment using optimized Docker engines, MindCluster, and high-efficiency runtimes. Driven by the AI wave, computing power has become the core cornerstone of intelligent transformation across all industries. Hu Zhengce stated: "Huawei will persist in innovation to build a diversified computing power foundation. Embracing open source and openness, Huawei will collaborate with customers, partners, and developers to co-create a thriving ecosystem, providing the world with a new choice."