Traditional infrastructure—highways, high-speed rail, airports, and urban facilities—forms the connective arteries of the physical world. Today, China is embarking on a significant wave of new infrastructure construction. This new infrastructure encompasses three major categories: information, integration, and innovation. Its core is a digital foundation represented by 5G and data centers, the integrated application of intelligent technology to upgrade traditional facilities, and scientific and educational facilities that support cutting-edge R&D.

　　Simultaneously, the explosive growth of artificial intelligence has propelled electricity supply to center stage. The electricity consumption for training a single large AI model can reach thousands or even tens of thousands of megawatt-hours. This has given rise to "computing-power and electricity synergy," where computing tasks are actively shifted to regions with lower electricity prices and a greater supply of green energy. The ultimate goal is to build an "intelligent computing power network" as convenient as the power grid, enabling a continuous flow of high-quality tokens throughout society and driving productivity transformation.

　　In this multi-dimensional race, Alibaba Cloud, as a global technology company with top-tier capabilities in large language models, cloud computing, and chips, is playing a pivotal role.

Digital Foundation: The Full-Stack Self-Developed "Computing Power Engine"

　　If the computing power network is likened to the future "power grid," then Alibaba Cloud's role is not just that of a "meter seller" or "pole installer," but a systemic builder encompassing everything from power generation units to the transmission network.

　　Chip Layer: The "Zhenwu" Computing Power Breakthrough by T-Head

　　In early 2026, Alibaba's chip subsidiary T-Head publicly unveiled the "Zhenwu 810E" computing chip. It features a self-developed parallel computing architecture and inter-chip interconnection technology, with 96G HBM memory and inter-chip interconnect bandwidth reaching 700GB/s. Combined with a full-stack self-developed software stack, it achieves deep hardware-software integration. This chip is widely applicable to the training and inference of various large models, including autonomous driving models, generative models, multi-modal models, perception models, and search/recommendation models.

　　Currently, Zhenwu PPUs have been deployed in multiple ten-thousand-card clusters within Alibaba Cloud. Alibaba Cloud serves over 400 customers, including XPeng Motors and Sina Weibo. As of February 2026, cumulative volume deliveries of Zhenwu PPUs had reached 470,000 units, with over 60% adopted by external customers. This large-scale commercial adoption has fully validated the adaptability and stability of Zhenwu PPUs in real, complex business scenarios.

　　Cloud OS: Apsara

　　Apsara is Alibaba Cloud's self-developed cloud operating system and the only one of its kind in China. It orchestrates and schedules millions of servers, with a single cluster managing over 100,000 units. It possesses EB-level data storage capabilities and, through the CIPU, has pioneered virtualization with "zero" performance loss, delivering industry-leading computing performance. This builds a cloud computing network with AI service capabilities and global technological competitiveness.

　　In the era of large AI models, the true performance ceiling of a cluster is not determined by compute units alone, but by the network connecting them. Alibaba Cloud has adopted an end-network integrated HPN architecture, self-developed 400G/800G intelligent network cards and 51.2T/102T ultra-high-capacity switches, and launched the high-performance communication library ACCL. The new generation HPN 8.0 uses a multi-plane CLOS architecture with IPv6 Native design, supporting the scale of hundreds of thousands of cards within a single availability zone and cross-regional RDMA interconnection. Concurrently, UPN employs LPO/NPO all-optical interconnection technology to build a high-speed communication network within a single rack, improving system availability by over three times while reducing costs by 30%.

Computing-Power and Electricity Synergy: Following Green Power

　　The energy consumption challenges posed by AI development are escalating from "background noise" to a core strategic issue. The China Academy of Information and Communications Technology predicts that by 2030, China's data center electricity demand will reach 300 to 700 billion kilowatt-hours, accounting for 2.3% to 5.3% of total national electricity consumption. Computing-power and electricity synergy—actively shifting computing tasks to locations with cheaper electricity, cooler climates, and more green energy—has become the key to resolving the "computing-power-electricity-carbon emissions" trilemma.

　　Synergy 2.0: A Multi-Scale Collaborative Paradigm

　　At the 30th United Nations Climate Change Conference (COP30), Alibaba Cloud was the first to propose the "Computing-Power and Electricity Synergy 2.0" paradigm for intelligent data centers. At the computing operations level, AI-driven dynamic scheduling intelligently shifts tasks to times and locations with better electricity pricing and greater green energy availability. At the planning level, it involves a thorough assessment of energy resources and power reliability in key regions to support the spatial and temporal planning of business development.

　　The "Five Greens" System and Green Data Centers

　　Alibaba Cloud has built a "Five Greens" system encompassing green energy, green products, green architecture, green operations, and green services, providing low-carbon solutions to 5 million global customers. Through AI-driven digital twin technology for operational optimization, Alibaba Cloud's self-built data centers can achieve a PUE as low as 1.09, averaging below 1.2, a leading level in Asia.

　　Beyond its self-built and operated five major super data centers in locations like Ulanqab, Zhangjiakou, Hangzhou, Shaoguan, and Nantong, Alibaba Cloud is also supporting three major telecom operators in building green data centers under the national "East Data, West Computing" strategy:

　　China Telecom:Collaborated to launch the first "Zhenwu" ten-thousand-card cluster in the Guangdong-Hong Kong-Macao Greater Bay Area in Shaoguan. Cluster resources are listed on "Guangdong Telecom's Computing Power Supermarket," offering retail computing services to small and medium-sized enterprises on a per-card, per-hour billing basis.

　　China Mobile:Participated in the "Xinjiang Computing to Chongqing" project to build a China Mobile-Alibaba Cloud Intelligent Computing Center. The first phase is planned for completion by the end of 2026, aiming to meet the 100,000-card computing power demand in Chongqing. The two regions have already established the country's first cross-provincial 100G computing power transmission line based on a V2V visual network, with a two-way latency of 32.4 milliseconds.

　　China Unicom:Collaborating to build the Sanjiangyuan Intelligent Computing Center in Qinghai. Upon full completion of its four phases, the center's total computing power is expected to reach 150,000 PFLOPS. The park achieves 100% green power supply through a green microgrid and has been selected as a "National Green Data Center" and a "Typical Case of Computing-Power and Electricity Synergy."

AI Competitiveness: From "Computing Power Race" to "Token Race"

　　In an era where AI competition is a core driver of new infrastructure, Alibaba Cloud's competitive edge lies not only in the scale of its computing power but also in its full-stack engineering capabilities to efficiently convert that power into intelligence.

　　Infrastructure Layer:Alibaba Cloud offers the Intelligent Computing Lingjun (PAI-Lingjun), designed for high-density computing needs like large model training, autonomous driving, and scientific intelligence. A single Lingjun cluster can support the collaborative work of 100,000 cards, with a single task supporting ten-thousand-card training. Node-to-node communication latency is in microseconds, and linearity for scaling to ten-thousand cards exceeds 96%. Its fully parallel storage CPFS achieves a single cluster throughput of up to 2 TB/s and 3,000万 IOPS, eliminating the performance bottleneck of Checkpoint during large model training. Lingjun also incorporates a full-link fault self-healing system, reducing fault recovery time by 80% and ensuring 7×24-hour stable operation of large clusters.

　　Model Layer:Alibaba Cloud was the first major Chinese tech company to open-source its self-developed large models and is the only global cloud vendor actively developing advanced AI models with a comprehensive open-source strategy, pioneering "full-size, full-modality" open-source in the industry. Since 2023, Alibaba has open-sourced over 400 models, including two major foundation model series: the large language model Qwen and the visual generation model Wan. The Qwen series has repeatedly won top honors in authoritative leaderboards like Chatbot Arena and OpenCompass, and has frequently topped trending charts on HuggingFace and GitHub. Currently, the number of derivative models based on Qwen has exceeded 200,000, with cumulative global downloads surpassing 1 billion.

　　Application Layer:Leveraging its full-stack AI capabilities of "large models + cloud + chips" and deep integration with Alibaba's business ecosystem, Alibaba Cloud continues to advance in both AI to Business (to B) and AI to Consumer (to C) directions. In November last year, Alibaba launched the Qwen App, which now has over 300 million Monthly Active Users (MAU) across all platforms, ushering in an era of large-scale AI agents for handling tasks. In January this year, Alibaba Cloud launched a full suite of OpenClaw cloud services, allowing users to quickly deploy OpenClaw on Alibaba Cloud's lightweight application servers or Wuying Cloud desktops, with on-demand access to over a hundred Qwen-series models on Alibaba Cloud Bailian. During the Spring Festival, Alibaba Cloud Bailian introduced its strongest "Coding Plan," featuring API services for four top open-source models: Qwen3.5, GLM-5, MiniMax M2.5, and Kimi K2.5. In March, Alibaba launched "Wukong," the world's first enterprise-grade AI-native work platform, enabling every team and company to have a "Lobster Legion" available 24/7.

　　As a leader in China's AI cloud market, Alibaba Cloud has built a comprehensive service ecosystem from foundational computing power to upper-layer applications, making it a true "Token Factory."

Conclusion: A Leading Player in the Intelligent Computing Power Era

　　In the AI era, the goal is to enable efficient data flow and make intelligence accessible anytime, anywhere. Alibaba Cloud's role in this transformation extends far beyond just "providing cloud computing services." It aims to deliver intelligent computing power, convert computing power into high-quality tokens, deliver them efficiently to businesses and users, and ensure these tokens are utilized effectively to drive tangible productivity transformation.

　　Competition in the AI era is not just about computing power infrastructure; it is increasingly about the "token economy" centered on "intelligence output per watt of electricity," stable transmission, and efficient utilization. Alibaba Cloud, along with other Chinese technology companies, is actively participating in this global competition and emerging as a leading player.