Huawei Rotating Chairman Xu Zhijun: Solving China's Computing Power Shortage through “Architectural Innovation”
Xu Zhijun(徐直军), Huawei's Vice Chairman and Rotating Chairman, delivered a speech titled "Embracing the Era of Full Intelligence." at the 2024 Huawei Connect Conference On September 19 . Xu emphasized that the continuous advancement of AI technology is driving industries towards deeper levels of intelligence, eventually leading to comprehensive intelligentization.
Notably, Xu highlighted China's computing power issue during his speech. According to Xu, Huawei views intelligence as a long-term process, with computing power being the critical foundation of this development. However, due to U.S. sanctions on AI chips and the limitations of China’s semiconductor manufacturing capabilities, Huawei must rely on the available chip manufacturing processes to build sustainable computing power solutions.
The core of Huawei’s strategy is to seize the opportunities presented by the AI revolution by fostering collaborative innovation across computing, storage, and networking technologies. They aim to develop computing architectures by creating a “super node + cluster” system to provide long-term, scalable computing solutions.
At the same time, Huawei believes not every company needs to build large-scale AI computing power. Some reasons include:
AI servers have high demands for power, cooling, and data center environments, and as AI models grow larger, the need for greater computing power will increase rapidly, which may result in underutilized resources or unmet demands in data centers.
The fast-paced development of AI hardware, with new products launched every 1-2 years, makes it difficult for small-scale computing operations to keep up. Mixing multiple generations of hardware can complicate resource management and negatively impact large model training capabilities.
AI technology is still in a growth phase with rapidly changing technologies, making it difficult for companies with only traditional IT maintenance capabilities to manage these shifts.
Additionally, Huawei argues that not every company should train its own foundational AI models, reasons include:
Training foundational models requires vast amounts of high-quality data, which is costly and difficult for many companies to gather.
The growing complexity of large models means that iterations and optimizations can take months to years, delaying the application of AI to core business areas.
Foundational models require cutting-edge, ever-evolving technologies, and there are few experts with hands-on experience, making it challenging for companies to build the necessary technical talent.
Xu suggests that companies should choose the most appropriate way to access computing power based on their unique business needs, rather than blindly building their own AI infrastructure or training foundational AI models.
The full transcript of Xu’s speech:
Embracing the Era of Full-Scale Intelligence
— Keynote Speech by Xu Zhijun at Huawei Connect 2024
Ladies and gentlemen, old friends and new friends, good morning! Welcome to the 2024 Huawei Connect Conference. I hope you all have an enjoyable stay in Shanghai. At the 2018 Huawei Connect Conference, I introduced Huawei’s AI development strategy and our full-stack, all-scenario AI solutions, positioning AI as a general-purpose technology. In 2021, I discussed how the Pangu large model enabled industries to achieve intelligent transformation. Since 2018, AI has developed rapidly and garnered significant attention from global investors, industries, and governments. Huawei has been steadily advancing its AI strategy since then, and at last year’s Huawei Connect Conference, we further clarified our strategy for full-scale intelligence. When it comes to intelligence, every industry and company has its own explorations. We’ve seen great results but also encountered many challenges. Today, I would like to take this opportunity to share some of our observations, thoughts, strategies, and practices.
1. AI: The Most Impactful Technology Across Industries
Let’s start with the business progress of AI. From a commercial application perspective, no technology has advanced as rapidly as AI, producing such a profound impact in such a short period of time. Research by McKinsey and Stanford University shows that AI applications are primarily concentrated in product development, marketing, and business operations across industries. From the perspective of corporate executives, Gartner's research indicates that CEOs have a highly positive outlook on AI. In summary, the continuous progress of AI technology is deepening intelligent transformation across industries, moving us towards a fully intelligent world.
2. A Vision for Enterprises in the Intelligent Era
As we stand at the dawn of the intelligent era, every enterprise not only hopes to quickly leverage AI to create value today but also to gain a competitive edge in the future. This has been a focal point of our thinking. I believe it’s essential to first understand the future direction of enterprises in the intelligent era and then work backwards to strategize and plan today’s actions. Based on Huawei’s own intelligent transformation practices and our years of support for industry-wide intelligent transformation, I’d like to share our vision of what enterprises in the intelligent era might look like and the characteristics they will need.
We believe that enterprises in the intelligent era should have six "A" characteristics. The first four "A"s represent the effects of intelligence:
The first "A" addresses how companies will serve their customers in the future, which we define as Adaptive User Experience. This means that intelligent enterprises should be able to perceive and understand users' behaviors, needs, interests, preferences, and environmental changes, and proactively adjust to provide services that best meet their needs. Such companies should be able to simultaneously and appropriately meet the personalized and unique demands of large numbers of customers. Products must be designed from the outset with this in mind, rather than merely being adapted after the fact. For example, an AI-powered learning device could automatically adjust its content and difficulty level based on a student's age, progress, comprehension, and feedback from tests, ensuring that each student receives a tailored learning experience at any given moment. The shift from offering predetermined, fixed experiences to adaptive experiences represents a major leap, and every company must offer customer experiences that align with the needs of the intelligent era.These six “A” characteristics summarize our preliminary thinking, based on our experience, on how to fully leverage AI in enterprises. We hope they will inspire more companies to become winners in the intelligent era.
The second "A" answers the question of what kind of products companies will need to create in the future. We believe this refers to Auto-Evolving Products. In the intelligent era, products will have the ability to learn autonomously, continuously iterate, and adapt to changes. These products will be able to self-optimize and evolve over time. For example, autonomous vehicles will improve the more they are driven, becoming easier and more efficient to operate. The shift from digitized products to intelligent products represents a major leap forward, fundamentally changing competition. Every company will need to consider how to incorporate intelligent capabilities into their products.
The third "A" addresses the future of daily operations in companies, which we define as Autonomous Operation. This refers to the need for businesses to achieve highly autonomous operational flows, from sensing, planning, and decision-making to execution, forming an end-to-end autonomous closed loop. For example, a port could use an intelligent planning platform to automatically generate operational plans, with autonomous trucks handling container transportation. For many years, companies have been striving for operational automation, and achieving autonomy in operations represents another leap in efficiency. Each company must think about how to use AI to empower and transform their operations across a broader and deeper scope.
The fourth "A" relates to the future of employee work experience and methods, which we define as Augmented Workforce. This means that every employee will have an intelligent assistant that “understands” them, enabling them to complete tasks efficiently and with high quality. For instance, telecom field maintenance workers could use a maintenance assistant app to quickly locate faults, identify root causes, and receive recommendations for handling the issues. The ultimate purpose of AI is to benefit humanity, and improving employee work experiences will be a key foundation for every company’s competitiveness in the intelligent era.
The next two "A"s represent the foundations of intelligent transformation:
The fifth "A" is All-Connected Resources, which means achieving full interconnection of a company’s assets, employees, customers, partners, and ecosystem. Every business object, process, and rule must be digitized, not just to increase the quantity of information but to improve the quality of that information. This provides the data and information foundation that is necessary for intelligence—through deep, comprehensive digitization.
The sixth "A" is AI-Native Infrastructure, which refers to the systematic construction of ICT infrastructure that can support the needs of intelligent applications. This is referred to as "ICT for Intelligence." Additionally, the operation, maintenance, and experience management of the infrastructure itself must also be fully intelligent, which is called "Intelligence for ICT."
These six "A" characteristics represent our preliminary thoughts based on our own practice and understanding of intelligence. We hope these insights will help others consider how to effectively leverage AI. Ultimately, we aim for every company to become a winner in the intelligent era.
3. Advancing the Full-Scale Intelligence Strategy
In response to the coming era of full-scale intelligence, Huawei introduced the full-scale intelligence strategy at the 2023 HC Conference. This strategy covers a wide range of areas, and today, I’d like to share some thoughts in seven key areas.
(1) Sustainable Computing Power Solutions through Architectural Innovation
Let’s first talk about computing power. Intelligent transformation is bound to be a long-term process, and computing power is the critical foundation for this transformation—both in the past and in the future. Therefore, sustainable intelligent transformation first requires sustainable computing power. Computing power relies on semiconductor processes, but we must face the reality that U.S. sanctions on AI chips for China are unlikely to be lifted in the long term, and China’s semiconductor manufacturing processes will remain behind for quite some time due to these sanctions. This means the sophistication of the chips we can produce will be constrained. This is a challenge we must address in building sustainable computing power solutions.
For China, only computing power based on the semiconductor processes that we can realistically access will be sustainable in the long term—otherwise, it’s unsustainable. Huawei has recognized this challenge, but also sees opportunities and possibilities, which have further ignited our passion for innovation. As AI becomes the dominant driver of computing power demand, the entire computing system is undergoing structural changes. What’s needed is system-level computing power, not just the computing power of individual processors. These structural changes present us with opportunities to innovate through architectural breakthroughs and pave the way for an independent, sustainable computing industry.
Our core strategy is to seize the opportunities presented by AI’s transformative moment. By leveraging the semiconductor processes we can realistically access, and driving coordinated innovation in computing, storage, and networking technologies, we aim to create new computing architectures. We are developing “super node + cluster” system computing power solutions that will meet long-term demand in a sustainable way.
The breakthrough in large model technology has significantly accelerated the process of intelligent transformation. Over the past period, almost every industry has been discussing large models, with many rushing to build AI computing power and train large models. For Huawei, as a provider of computing power, this is undoubtedly a major benefit. However, from a long-term development perspective, we firmly believe that only the sustained success of our customers can ensure Huawei’s long-term growth. Today, I’d like to share a few thoughts on some key issues.
First, not every company needs to build large-scale AI computing power. We all know that AI servers—especially AI computing clusters—are very different from general-purpose x86 servers. They have extremely high requirements for power supply, cooling, and data center environments. As large models grow larger, AI computing power will need to scale accordingly, and the pace of change is rapid. AI servers are being upgraded quickly, creating a dilemma for data centers: either their infrastructure is underutilized, or they can’t meet the demands.
Additionally, the industry is releasing new AI hardware products every one to two years, with a fast pace of iteration. Compared to public cloud providers, companies with small-scale computing power are less capable of handling the fast-changing landscape of large models. They often find it difficult to complete work with each new generation of computing hardware, requiring multiple generations to work together for model training. This increases the complexity of resource scheduling, and the “short-board” effect of older-generation products can drag down the performance of newer ones, affecting the ability to train large models effectively.
Lastly, there are operational and maintenance challenges. AI technology is still in its growth phase, and rapid technological changes, coupled with the coexistence of multiple generations of products, require high levels of expertise, making it difficult to operate and maintain. For companies that only have traditional IT maintenance capabilities, this presents a major challenge. Since these challenges will persist for some time, I believe each company should carefully consider how to best obtain AI computing power, rather than simply building its own.
Second, not every company needs to train its own foundational large model. The key to training foundational models is data, and gathering enough high-quality data is a significant challenge. Pre-training foundational models now require datasets at the scale of tens of trillions of tokens. This not only involves high costs but also raises the question of whether companies can acquire sufficient data.
Moreover, training models is difficult. The parameter size of foundational models continues to increase, and model iteration and optimization have become more challenging, often requiring several months to years for a complete training iteration. Companies should focus on their core businesses, as training foundational models independently could delay AI from empowering their core operations.
Finally, acquiring talent is difficult. The technology related to foundational models is evolving daily, and there are few technical experts with real-world experience. For companies, building a sufficient pool of technical talent is another challenge.
Third, not all applications need to pursue "large" models. From Huawei Pangu’s industry practices, we’ve found that models with billions of parameters can meet the needs of business scenarios such as scientific computing and predictive decision-making, for example, in rainfall prediction, drug molecule optimization, and process parameter prediction. On endpoint devices like PCs and smartphones, billion-parameter models are also widely applicable. Models with tens of billions of parameters can meet the needs of specific domains such as NLP, computer vision (CV), and multi-modal tasks, for example, in knowledge Q&A, code generation, customer service assistance, and security detection. More complex tasks in NLP and multi-modal applications may require models with hundreds of billions of parameters.
Therefore, we believe that companies need to choose the most suitable models for their various business scenarios. By using a combination of multiple models, they can solve problems and create value.
(2) Huawei Cloud’s Full-Stack AI Upgrade: Empowering Industry Intelligence
Based on the points I just mentioned, I believe that for many companies that lack the capability to build their own AI infrastructure or train foundational large models, choosing cloud services is a more reasonable and sustainable option. Huawei Cloud has been upgraded to address these challenges, focusing on AI, and is committed to enabling every company to train models and apply model inference efficiently and on-demand.
Firstly, through continuous development of the Ascend Cloud service, Huawei Cloud allows companies to access powerful AI computing resources with a single click, without the need to modify or build their own data centers, or manage and maintain AI computing infrastructure. With coordinated integration of computing, storage, and network resources, Huawei Cloud has already achieved 40 uninterrupted days of training for models with hundreds of billions of parameters.
4. Huawei Cloud’s Comprehensive Security Capabilities: Ensuring Safe AI Model Training and Inference
Training and inference of large models in the cloud bring new security challenges. To address these challenges, Huawei Cloud has significantly enhanced its security capabilities to ensure the safety of large model training and inference. Key aspects include:
Security Philosophy: Huawei Cloud adopts a "defense against extreme attacks" philosophy in its security design. Based on zero-trust architecture, it has built a seven-layer defense system covering physical security, identity, network, application, host, data, and operations, along with a security operations center. This system successfully defends against up to 1.2 billion attacks daily, ensuring that operations remain uninterrupted, data is not lost, and regulatory compliance is maintained.
Security Mechanisms: Huawei Cloud provides tiered cloud services to create a secure digital space for customers. It supports physical or logical isolation, and its cloud platform operations are transparent and auditable, ensuring that customers can use the cloud with confidence.
Security Technology: Huawei Cloud offers an end-to-end full-stack data security protection solution, ensuring comprehensive security throughout the data lifecycle—from hardware, software, and applications to data flow, large model training, and inference data. This ensures that the training data and generated content are compliant with security regulations throughout the entire process.
Intellectual Property: If a customer using Huawei Cloud’s large model services generates content that infringes on third-party intellectual property, Huawei will bear the cost of legal defense and compensate for any losses, costs, or expenses resulting from a final court ruling or settlement with the third party. The specific terms will be outlined in the contract.
(3) Building HarmonyOS Native Intelligence for an All-Scenario Smart Experience
In the intelligent era, terminals are an indispensable part of the ecosystem. Huawei was one of the earliest companies to integrate AI into smartphones. As early as 2017, Huawei launched the Mate10, which featured a built-in AI chip, and introduced AI-powered capabilities such as intelligent image processing and translation to smartphones, marking the start of the Mobile AI era. Today, as AI enters the large model era, we have integrated AI deeply into the HarmonyOS system architecture across devices, chips, and cloud, reconstructing HarmonyOS around AI. This allows us to achieve full-scale intelligence, from the kernel to system applications, while fostering more open ecosystem collaboration and providing stronger privacy protection.
Huawei will upgrade its virtual assistant "Xiaoyi" into a more advanced intelligent agent, enabling more natural multi-modal interactions and comprehensive sensory integration. It will precisely understand the user, the digital world, and the physical world, providing users with intelligent, personalized services across all scenarios. Additionally, we will work with HarmonyOS ecosystem partners to co-create smart capabilities for future products, tailored to the full range of consumer needs in work, learning, life, and entertainment. We will also fully open up HarmonyOS, from AI model capabilities to AI components, empowering third-party applications and fostering a thriving HarmonyOS native app ecosystem.
5. Terminal AI Should Focus on User Experience, Not Computing Power
We have also observed that incorporating AI capabilities into various terminals has become a common trend, such as the development of AI Phones and AI PCs. As a result, there are many perspectives in the industry on how to define intelligent terminals in the AI era. At Huawei, we firmly believe that user experience should be the top priority. Consumers often find it difficult to understand technical specifications like chip processes, TFLOPS of computing power, or model parameters. Instead, they focus more on how these technologies enhance their personal experience.
Therefore, we advocate that terminal AI should focus on user experience, rather than computing power. Based on this philosophy, to give consumers a clearer and more intuitive understanding of AI terminal capabilities, and to reach a unified consensus within the industry on the evolution of AI terminal capabilities, we have collaborated with the Tsinghua University Institute for AI Industry Research to propose a classification standard for AI terminals, ranging from L1 to L5 levels of intelligence. This standard is driven by user experience, aiming to quantify the intelligent experience of users. By continuously improving the intelligence level, we aim to provide users with better experiences. We also encourage our industry peers to work together to refine and optimize this classification standard, promoting the orderly development of terminal AI across the industry.
(4) Reshaping Network Experience and Operations with Autonomous Driving Networks
Since 2018, Huawei has been pioneering the application of AI in telecommunications networks and introduced the concept of the autonomous driving network architecture. We are now incorporating large communication models and network digital twins. Together with partners such as TM Forum and China Mobile, we are driving highly autonomous networks based on value scenarios, progressively moving toward L4 level autonomy and aiming for full autonomy in the future. Through operator network autonomous driving, we are committed to delivering an ultimate user experience with zero wait times, zero interruptions, and zero-touch operations, as well as ultra-simplified network operations with self-configuration, self-healing, and self-optimization.
At the same time, we are extending the concept of autonomous driving networks to enterprise networks, which face similar challenges in terms of operations and maintenance. The rapid shift toward wireless office environments, cloud-based and video-heavy applications, has made it difficult to fully ensure the user experience. In addition, as the scale of enterprise networks continues to grow—covering offices, production sites, data centers, branch locations, and multi-cloud connections—network equipment becomes more diverse, and the scope and complexity of day-to-day maintenance increase. Today, we are proposing the concept of enterprise network autonomous driving with the goal of achieving zero lag in business operations, zero network interruptions, zero waiting for network deployment, and zero security risks.
(5) Developing Autonomous Driving Solutions with a Focus on Safety and User Experience, Ultimately Achieving Driverless Cars
Automotive autonomous driving solutions have been one of the key areas where Huawei first invested in AI, as the ultimate goal of autonomous driving is driverless cars—one of the most challenging application scenarios for AI. Our ADS 3.0 version enables more accurate decision-making, more efficient traffic flow, a more human-like driving experience, and enhanced safety. The solution also achieves one-click parking, providing a seamless experience from public roads to campus roads and even underground parking lots. We have further upgraded the all-direction collision avoidance system, expanding its speed range and achieving comprehensive obstacle avoidance.
These advancements allow consumers to truly experience the safety and improved user experience brought by intelligent driving. Today, Chinese consumers are highly familiar with intelligent driving, and the proportion of new cars equipped with advanced intelligent driving features is very high. In fact, intelligent driving capabilities have become a key factor for Chinese consumers when purchasing new cars.
Looking ahead, we will continue to evolve our autonomous driving solutions based on integrated perception, progressively achieving: On highways: get in the car and relax, even take a nap on long journeys; On city and suburban roads: easy and safe driving that rivals that of experienced drivers; On rural roads and mountainous terrain: confident, all-terrain, all-weather driving; In parking scenarios: drop-off parking with zero scratches and no stalling.
In terms of safety, we aim to achieve comprehensive, all-direction proactive safety, focusing primarily on reducing accidents for which the driver is responsible and mitigating secondary responsibility. With these key milestones achieved, we will ultimately realize driverless cars.
(6) Building a Unified Developer Platform and Ecosystem for Win-Win Growth
Ecosystem development has always been a key part of Huawei’s strategy. We consistently work with partners to build ecosystems and create a unified developer platform that drives mutual success. From 2017 to 2019, Huawei initiated the construction of the Huawei Cloud, Ascend, Kunpeng, and HarmonyOS ecosystems. In 2024 and over the next five years, Huawei will make significant strategic investments in ecosystem development. By developing these ecosystems, we aim to promote and drive growth in the computing and terminal industries, offering the world an alternative choice in computing and providing the world with a third mobile operating system.
(7) Advocating and Practicing "AI for Good": Enhancing Human, Social, and Environmental Well-being
Finally, AI has limitless applications, but at its core, it must serve humanity. We advocate for and practice "AI for Good" and believe that:
AI should serve people, improving work efficiency and quality of life. Through AI-enabled industry digitization, AI can transform how industries operate, becoming the core engine for industries as they transition to the intelligent world. AI technology barriers should be lowered so that everyone—individuals, families, and organizations—has equal access to and use of AI.
AI should be used to create broader societal benefits. During the design, development, and application of AI, we will carefully assess the long-term and potential impacts of AI on society, avoiding the misuse of AI technology.
AI should be applied to environmental protection and sustainable development, actively contributing to global issues such as the United Nations' Sustainable Development Goals.
The era of full-scale intelligence has arrived, bringing new opportunities and challenges to every person and every enterprise. Let’s work together to promote full-scale intelligence so that everyone has their own intelligent assistant, every company can become an intelligent enterprise, and every car can achieve autonomous driving. Thank you!