Core views

The 2024 Yunqi Conference focused on AI and discussed and shared topics such as AI-related technology, big model development, and AI's promotion and reshaping of multiple industries. Alibaba Cloud is upgrading its product line with AI as the center and strives to seize the opportunities of this new technological revolution. In terms of models, Alibaba Cloud creates a large family of full-scale, full-modal, multi-scene models.

In terms of related facilities, the AI Infra series products are innovated and upgraded with AI as the center to create the strongest infrastructure adapted to AI to meet rapidly growing AI infrastructure needs. In terms of databases, develop cloud-native databases driven by Data+AI to meet the one-stop management of customer multi-modal data. In terms of cloud security, a full product line was upgraded and NDR, a cloud-native network detection and response product, was launched for the first time.

occurrences

2024 Alibaba Yunqi Conference was held in Hangzhou on September 19-21 under the theme “Cloud is smart, industry is changing”. The Yunqi conference set up three main forums and nearly 400 parallel topics and forums. The exhibition hall brought more than 1,000 technology exhibits around the three themes of “Artificial Intelligence +,” “Computing,” and “Frontier Applications,” and presented a feast covering various cutting-edge fields such as chips, big models, cloud storage, cloud computing, robotics and intelligent driving.

Brief review

The conference focused on AI, pointing out that the rapid iteration of AI will profoundly affect products and productivity in the physical world, and that the accompanying fundamental changes in computing architectures will bring about a large number of AI-related infrastructure requirements. Wu Yongming, CEO of Alibaba Group and Chairman and CEO of Alibaba Cloud Intelligence Group, stated, “In the past 22 months, AI has developed faster than any historical period, but it is still in the early stages of AGI transformation.” Currently, big model technology is rapidly iterating, and the cost of model inference is falling exponentially, far exceeding Moore's Law, but to achieve the next generation of AGI models, the investment threshold for advanced models worldwide will reach more than a billion dollars. In addition to the digital world, new products with AI capabilities will also be formed in the physical world, and there will be a synergy effect with cloud AI, which will revolutionize industry and productivity. Automobiles and robotics are industries that can be predicted to undergo drastic changes.

Unlike computing systems dominated by CPUs in the past, in the future, almost all software and hardware computing cores will use a computing model where GPU AI computing power is the main focus, and traditional CPU computation is supported.

More than 50% of the new demand is driven by GPU AI computing power, and a large number of existing applications are also being rewritten using GPUs. In order to meet the blowout of demand for infrastructure with higher performance, larger scale, and better suited to AI, Alibaba Cloud has expanded a single network cluster to the 100,000 card level, and is focusing on investing in and building advanced infrastructure adapted to future AI in various aspects, from chips, servers, networks, and storage to cooling, power supply, and data centers.

Alibaba Cloud's AI Infra series products are innovated and upgraded with AI as the center to create the strongest infrastructure adapted to AI. Through full-stack optimization, Alibaba Cloud has created a stable and efficient AI infrastructure to achieve continuous training effective time greater than 99% and increase model computing power utilization by more than 20%. The Panjiu AI server supports 16 stand-alone cards, 1.5T video memory and 3.2T bandwidth, and can provide AI algorithms to predict GPU failures, with an accuracy rate of 92%; HPN7.0, a high-performance network architecture designed for AI, can be stably connected to more than 0.1 million GPUs. Among them, the integrated communication performance is increased by more than 1 times, and the model's end-to-end training performance is increased by more than 10%; Alibaba Cloud CPFS file storage, data throughput is 20 Tb/s, providing exponentially expanded storage capacity for AI intelligence. Through the upgrading of infrastructure such as underlying computing, storage, and networking, Alibaba Cloud's Lingjun cluster can provide intelligent computing power of up to 0.1 million cards. The linearity of 10,000 card scale performance exceeds 96%, the parallel storage throughput is 20Tb/s, and the network bandwidth utilization rate exceeds 99% at the 10,000 card scale, which can support the 100,000 card level AI computing power scale of a single cluster. For AI businesses, Alibaba Cloud computing products have also evolved significantly. The container service ACK has achieved a major upgrade for AI. The cold start delay for large model applications has been reduced by 85%, and it can support 15,000 hyperscale nodes. At the same time, the container computing service ACS will launch GPU container computing power for the first time, which can improve computational affinity and performance through topology-aware scheduling; the artificial intelligence platform PAI has achieved 10,000 card level training inference unified flexible scheduling, and the effective utilization rate of AI computing power exceeds 90%, which can comprehensively improve the efficiency of model training and inference. Alibaba Cloud released a new elastic computing product family, based on the fully upgraded cloud infrastructure processor CIPU2.0. The three elastic computing product lines, including general computing, accelerated computing, and container computing, all achieved significant performance improvements. In order to meet the rapidly growing demand for accelerated computing represented by GPUs, Alibaba Cloud released the CIPU 2.0 chip, which increased overall machine stability by 20%, IOPS increased to 3.6 million, and increased bandwidth from 2 million to 4 million. While improving inference efficiency, it also has full security capabilities The orientation was enhanced. In terms of general computing, two 9th-generation ECS enterprise-class instances were released in collaboration with AMD and Inte l. The Web application was increased by 20%, database capacity increased by 17%, and search recommendations increased by 30%. Big data increased by 20%. At the same time, Yitian ECS achieved full-stack optimization, and the big data scenario and video transcoding scenario increased by 14% and 30% respectively. In terms of container services, enterprise-grade containers have achieved full-scene coverage, reduced cold start time by 85%, and capacity expansion efficiency increased by 25%. In terms of storage, OSS optimized for AI/ML workloads, and achieved a 30-fold increase in performance for Python & Go languages.

Vigorously develop basic models, model services, model applications and model community complexes to create a full-scale, full-modal, multi-scene generic model family. The Tongyi Qianwen ecosystem has gradually been formed, and the total number of downloads of the Qwen2 model exceeds 40 million; the newly upgraded open source model Qwen2,5 can generate 10,000 word reports, and math and code capabilities have been greatly improved. Combining large language models, multi-modal models have been launched in multiple vertical fields: including Qwen-VL, a visual language model that can relate text and images and has strong comprehension ability; Qwen-Au d io, a large-scale audio language model that can fully understand conversation, music, and noise; Qwen-Math, a large mathematical model with mathematical abilities that exceed the capabilities of GPT 4O and other open source models; and Qwen-Coder with the best coding ability of the same size. Qwen Tongyi Wanx's ability to visually generate large models has also been fully upgraded. Wensheng's video capability has been newly launched, which can generate 1080p videos of 20 seconds, covering a variety of scenes and art styles. In terms of model application, the Tongyi Spirit Code upgrade expands AI programmers' capabilities, and can independently complete the R&D process from requirement understanding and coding implementation to independent testing; Tongyi Hearing and Tongyi will separately upgrade real-time recording and real-time reading assistant capabilities; Tongyi AI Translation Assistant will launch simultaneous and interactive real-time interpretation functions; and Tongyi Video Call combined with language models to launch chat partners. In terms of model services, the Alibaba Cloud Model Service Platform combines model calls and enterprise-level calls, provides a model-centered development paradigm, including enhancing workflow and intelligent orchestration capabilities, and refining operation tools, auxiliary observation, and optimization of application effects. Meanwhile, Tongyi Qianwen's recommended models Qwen-Plus and Qwn-Turbo will reduce prices across the board, up to 85%. In terms of the model community, the Magic Match Community has become the largest model community in China, with more than 1000 models and 6.9 million developer users.

From data to intelligence, Alibaba Cloud creates a cloud-native database driven by Data+AI. Alibaba Cloud has built DM S, a one-stop multi-mode data management platform. Through DMS+X, it can achieve open, unified and open source multi-mode data governance. At the same time, the flow of data between engines can achieve data processing, feature extraction, and one-stop management of orchestration services, thereby helping enterprises build cross-cloud, cross-environment, cross-engine, and cross-data type databases. Under the DMS framework, Alibaba Cloud completely decouples memory, storage, and computing, and uses a full stack pool as the core concept to create a differentiated engine for different needs.

Cloud Native Serverless has been fully upgraded with the core of faster bounce, more stable bounce, wider bounce, and finer bounce. The memory pool can already support 100TB of remote sharing and consistent read/write access, and won the first ICDE 2024 Best Industry Paper Award in mainland China. PolarDB AlwaysOn achieved large-scale commercialization of multi-master and multi-write for the first time in the world, fully demonstrating the advantages of a three-tier decoupling architecture. Lindorm has made significant progress in 2024, covering 65% of domestic car companies and more than 60% of generative AI unicorns, including many companies such as Ideal, NIO, and Baichuan Intelligence. From a customer perspective, Lindorm can truly achieve unified entry, unified storage, and unified management of multi-modal data, thereby supporting the development of more complex and diverse intelligent applications. The Tair serverless KV service was launched to cater to the market environment of the explosion of AI applications and the growth of inference services, and cache acceleration for large model inference services.

All cloud-native security capabilities have been upgraded, and the cloud-native network detection and response product NDR has been released for the first time. Complex IT systems in the cloud era, fragmented security tools and traditional protection ideas, and multiple changes brought about by new technologies and new threats are pain points in the field of cloud security. In response to these problems, Alibaba Cloud proposed a “three-in-one” strategy for cloud security: integration of infrastructure security, integration of security technology domains, and integration with office safety and production safety. Native NDR comprehensively enhances full-traffic defense capabilities in cloud environments. Compared with traditional third-party products, it can be launched immediately without deployment, and can forensically store attack incidents and traffic before and after the attack occurred, taking into account retention requirements and cost investment, and then perform traceability and correlation analysis to help customers discover advanced cyber threats faster. Alibaba Cloud has also strengthened capabilities such as WAAP, Cloud Security Center, and DDoS protection, and has also upgraded the security capabilities of cloud-native products such as databases, network CDN, computing, and storage. At present, Alibaba Cloud has become one of the vendors with the most complete global security capabilities recognized by the three international authorities Fo rrester, Ga rtner, and IDC.

Profit forecast and valuation: Alibaba's revenue growth rates for FY2025-2026 are expected to be 8.6% and 8.62%, respectively, and non-GAAP net profit growth rates of -2.5% and 18.22%, respectively. Using the segmented valuation method, the corresponding ADS valuations for Taotian Group and Cloud Intelligence Group are 103.17 US dollars and 13.34 US dollars, respectively. The target price is 116.51 US dollars/ADS, corresponding to 113.48 HKD/share for Hong Kong stocks.

Risk warning: Shareholder returns fall short of expectations, low dividend expectations, SBC exceeds expectations; macroeconomic and social zero growth; weak macroeconomic and social zero growth; the Federal Reserve's interest rate cut process is low; domestic e-commerce market share continues to decline; CMR growth is low, and Taotian Group's profit margin expectations are low; overseas e-commerce market competition is intense, AIDC losses exceed expectations; overseas tariff policy adjustments affect AIDC's growth rate; Cainiao Group's revenue growth rate continues to be low; Alibaba Cloud's revenue growth rate continues to fall short of expectations; profit margin growth falls short of expectations; AI drives cloud computing's limited business; Industry regulation Risks; progress in the implementation of Ant Group's credit reporting license and financial control license fell short of expectations; uncertainty about the development of Sino-US relations; risk of Chinese stock delisting.