SHANGHAI, CHINA – Media OutReach Newswire – 18 September 2026 – On September 18, Dr. Peter Zhou, Director of the Board at Huawei and CEO of Huawei Cloud, delivered a keynote titled “The Agentic Cloud for the Agentic World: Build Together, Grow Together” at HUAWEI CONNECT 2026. He shared Huawei Cloud’s latest progress. He announced the global launch of the latest AI Cluster Service (AICS), a key step in Huawei Cloud’s strategy to strengthen the silicon bedrock on the cloud and reinforce the foundation for agentic AI. The keynote also highlighted the Agentic Model as a Service (MaaS) platform, which brings together diverse models to accelerate model capabilities as services at scale. The AgentArts enterprise-grade agent platform already serves over 100 enterprises. The Industry AI Foundry has accumulated more than 1,000 industry assets, with over 1,000 projects deployed. The Industry AI Foundry now includes two new zones, Smart Government Zone and AI Hardware Zone, to drive agents toward large-scale adoption. These efforts, taken together, are building a thriving AI ecosystem on the cloud.
 




 

Building Agentic Infra for agents
Dr. Peter Zhou also highlighted that in the agentic era, infrastructure no longer simply exists to provide compute. Instead, its purpose has expanded to making every token more efficient, coordinating general and AI compute, enabling models to keep learning, and, more importantly, running agents securely and reliably within real production environments. Huawei Cloud has defined Agentic Infra, a new paradigm around efficient tokens, enhanced memory, unified general & AI scheduling, and secure autonomy. To date, Agentic Infra has served over 3,500 customers.

The latest AICS is built on a five-level fast recovery mechanism with full-chain observability. It supports over 40 days of stable training on cloud and is capable of fault recovery within 10 minutes. Furthermore, supported by coordinated optimization across scheduling, cache, and algorithms, it delivers 20% higher token throughput than the previous generation of compute service. The latest AICS will be commercially available in China on September 30 and in markets outside China on November 30.
The Context Memory Storage (CMS) solution addresses the needs of long-horizon agent tasks for memory capacity and access efficiency. It provides a petabyte-scale memory space, twice the storage capacity of comparable industry products, and supports high-speed terabyte-scale memory reads with 50% higher performance than industry peers.
Agentic MaaS brings together diverse models, enabling developers to invoke state-of-the-art (SOTA) models from leading providers with just one click and no deployment required. At the event, MiniMax demonstrated its advanced multimodal model, which, combined with Huawei Cloud, delivers open, cutting-edge multimodal capabilities.

Building an enterprise-grade agent platform to help enterprises develop and use agents effectively
Huawei Cloud is advancing its enterprise-grade agent platform through a dual approach: commercial and open source. Built on AgentArts and its open-source edition openJiuwen, the platform opens up more than 5,000 general Model Context Protocol (MCP) assets and over 1,000 industry-specific MCP assets, making it easier for enterprises to develop, use, and manage agents. To date, the platform has served more than 100 customers, including the Shenzhen Longgang District Government, China Southern Power Grid, Guangzhou Laboratory, the University of Science and Technology of China, Kingsoft Office, Sichuan Yingu Carbon Sink Renewable Resources Co., Ltd., Changsha Thunder Cloud Network Technology Co., Ltd., and KingMed Diagnostics. At the event, Kingsoft Office shared its implementation practices based on Huawei Cloud’s agent platform. By deeply integrating the agent platform with the WPS 365 Document Center and WPS Comate, Kingsoft Office has built vertical office agents that are now deployed across multiple industries, including finance and government.

Huawei Cloud has announced that AgentArts will be commercially available in markets outside China on December 30. To date, the openJiuwen open-source community has surpassed 50,000 stars and 3.29 million downloads. Together with Chinasoft International, iSoftStone Group, and Beiming Software, Huawei Cloud has launched partner commercial editions and shared a broad market space, helping partners thrive on Huawei Cloud’s silicon bedrock.

Advancing the Industry AI Foundry to accelerate AI adoption at scale across industries
Through the Industry AI Foundry, Huawei Cloud brings together assets such as industry scenarios, models, data, knowledge, and agents, turning industry know-how from isolated project experience into reusable industry capabilities. It has built five industry-specific zones — Smart Healthcare Zone, Embodied AI Zone, AI for Science (AI4S) Zone, Smart Manufacturing Zone, and Smart Finance Zone — with over 1,000 industry-specific assets and more than 1,000 deployed projects.

Two new zones have been launched within the Industry AI Foundry: Smart Government Zone and AI Hardware Zone. The Smart Government Zone has already brought together 24 founding partners, covering a range of scenarios such as government office operations, public services, urban governance, and industry development. The AI Hardware Zone has onboarded 15 core partners, covering more than 20 device types, including AI glasses, AI toys, and AI recording cards, along with over 10 scenario templates and more than 110 scenario skills.

From infrastructure and enterprise-grade agent platforms to industry ecosystems, Huawei Cloud is committed to building an open agentic cloud, with open infrastructure to power AI, an open platform to support enterprises in effectively developing and using agents, and an open industry ecosystem to accelerate AI adoption at scale. Ultimately, Huawei Cloud looks to move forward with customers and partners to build and grow together in the agentic era.

Themed Advancing the Agentic World, HUAWEI CONNECT 2026 will delve into AI across three dimensions: strategy, technology, and ecosystems. You can expect an in-depth look at our latest strategic initiatives, and we’ll also be unveiling our all-new digital and intelligent infrastructure products, scenario-specific solutions for industries, and development tools. The event will run from September 17 to 19 at the Shanghai World Expo Exhibition & Convention Center and Shanghai Expo Center. For more information, please visit HUAWEI CONNECT 2026 online at www.huawei.com/en/events/huaweiconnect
 




 

Key takeaways:

• Strategic focus: The rapid approach of an intelligent world is driving up demand for computing power. Huawei is focused on developing AI infrastructure, and is actively driving innovation in systems and architecture centered around SuperPoDs and SuperClusters. These efforts are aimed at building a solid silicon foundation for the intelligent world.
• Technological breakthroughs: Huawei unveiled the Atlas 960E SuperPoD, the first in the industry to use NPO. The company also launched an upgraded TaiShan 950 SuperPoD, as well as the OceanStor M900 (a memory context storage system). Interconnected with UnifiedBus, Huawei’s agentic SuperCluster can scale up to one million NPUs.
• Open ecosystems: Huawei is actively building out open computing ecosystems. To date, the Kunpeng ecosystem has attracted 4.16 million developers from around the world. CANN has moved to sustained, community-driven open-source development. Ascend now spans over 90 leading third-party open-source projects and is officially supported as a PyTorch accelerator backend.

SHANGHAI, CHINA – Media OutReach Newswire – 17 September 2026 – HUAWEI CONNECT 2026 kicked off today in Shanghai. The first keynote of the date was by David Wang, the Deputy Chairman of the Board and Rotating Chairman at Huawei. In his speech (Advancing the Agentic World, Building a Solid Silicon Foundation), Wang highlighted the work the company is doing alongside industry stakeholders to build powerful AI infrastructure, lay a solid computing foundation, and address the challenges and opportunities in the intelligent world to come.

AI is sweeping the world faster than any previous technological revolution. Today, foundation model parameters are rapidly approaching 10 trillion, and are projected to exceed 100 trillion by 2030. AI agents can now work on the same task continuously, for hours on end. By 2030, they will be able to handle tasks that span months.

In China alone, the average number of inference tokens consumed every day has surged to around 500 trillion, and is expected to reach quintillions (1018) by 2030.

On-device AI is also advancing rapidly. On-device models for smartphones have expanded from three billion parameters in 2024 to 30 billion today, and will push toward hundreds of billions in the near future.

These trends will set a much higher bar for the scale, performance, and reliability of underlying technical systems. Only by building powerful AI infrastructure can the industry lay a solid foundation for the future intelligent world.

An intelligent world is approaching – and faster than ever. To lead the charge into this new world, Huawei is laser-focused on building out AI infrastructure – the silicon foundation for the future to come.

In particular, Huawei’s AI strategy is centered on computing power, with a focus on monetizing hardware. The company is also sharpening its competitive edge through systems and architectural innovation. Centering these efforts on SuperPoDs and SuperClusters, the company aims to build a solid computing foundation and offer a new option for the world.

Huawei is a major contributor to open computing ecosystems, and will continue to support native training for mainstream foundation models on its systems, as well as supporting a vast range of models and applications.

For customers, Huawei provides flexible on-premises and cloud compute solutions for its customers to accelerate intelligent transformation across industries.

With diverse forms of compute, including solutions for micro-, low-tier, mid-range, and massive computing power – Huawei is driving the expansion of on-device and in-vehicle AI, making intelligence truly ubiquitous.

Additionally, Huawei is dedicated to building next-generation communications networks to bring readily available compute and intelligence to every person, home, and organization.

SuperPoDs gain broad consensus, with adoption growing in industries, academia, and research institutes

To date, over 1,000 Atlas 900 A3 SuperPoDs have been deployed, and Atlas 950 SuperPoD is seeing large-scale commercial use. While adoption continues to grow, SuperPoDs have gained broad acceptance across industry, academia, and research institutions as a key direction for AI infrastructure. Currently, a SuperPoD is explicitly defined as a computing system in which multiple computing nodes are tightly coupled through high-speed interconnect protocols, featuring unified memory addressing across physical nodes — functioning like a single logical computer.

SuperPoDs are the go-to choice for AI infrastructure buildout. Right now, 100k-NPU computing clusters have become the baseline for training SOTA models. However, traditional server architectures result in intra-cluster communications that account for over 40% of total training time, severely constraining Model FLOPs Utilization (MFU). Simulation results from Huawei’s Markov Lab show that a 100k-NPU cluster built with 4k-NPU SuperPoDs can deliver a 2.75x increase in MFU compared to a 100k-NPU cluster composed of 8-NPU servers.

11-chip UnifiedBus-powered portfolio for SuperPoDs and SuperClusters; the Atlas 960E SuperPoD –the industry’s first to use NPO

The Ascend series of chips is the most critical component in Huawei’s 11-chip UnifiedBus-powered portfolio for SuperPoDs and SuperClusters. Development on Ascend 960 has exceeded the company’s expectations, with performance doubling as planned. Ascend 960DT will be available in Q1 2027, three quarters ahead of the company’s original roadmap. And the Ascend 960PR will be ready in Q3 2027, one quarter ahead of schedule.

“We’re evolving our Ascend chip series on a one-generation-a-year cycle,” said Wang in his keynote. “In 2028 and 2029, we will roll out the Ascend 970 and 980 chips, respectively. Thanks to the Tau (τ) Scaling Law, not only will their compute specifications continue to double, but you can also expect to see huge improvements across the board in terms of memory bandwidth, memory capacity, interconnect bandwidth, and more.”

In addition to Ascend chips, Huawei has also developed a complete portfolio of chips for AI infrastructure, based on UnifiedBus, delivering key capabilities that cover computing, interconnect, storage, and management.

“SuperPoDs are designed to coordinate multiple NPUs through interconnect,” continued Wang. “We have developed a next-generation optical interconnect product based on near-packaged optics (NPO): the High-density Optical-interconnect-Node Engine (Hi-ONE).” Built on Huawei’s proprietary technologies, Hi-ONE has a multi-physics design for balancing optical, mechanical, electrical, electromagnetic, and thermal performance, realizing a transmission capacity of 7.2 Tbit/s per single engine.

“This is the industry’s first NPO product ready for mass production, delivering the largest transmission capacity. It is also the industry’s first NPO product with a built-in light source.”

This product combines high bandwidth and high reliability with low latency and low power consumption. This, coupled with UnifiedBus, will make it far easier to scale up SuperPoD interconnect systems.

Recently, Huawei submitted an implementation agreement (IA) on NPO to the Optical Internetworking Forum (OIF), a standards organization. The response from numerous industry partners has been widely positive. Huawei will continue its efforts to further refine the NPO industry ecosystem.

Using Ascend 960 chips and Hi-ONE, Huawei has developed the industry’s first NPO-based SuperPoDs: the Atlas 960E SuperPoDs. A single Atlas 960E SuperPoD can scale up to 4,096 NPUs, delivering 8 EFLOPS of FP8 compute performance, with up to 1 petabyte of HBM capacity. With 5,500 Hi-ONE units, this SuperPoD doesn’t need the 48,000 800G optical modules that would traditionally be required to connect all the NPUs. This cuts power consumption by over 550 kilowatts, while doubling the system’s fault-free operating time, achieving 99.8% system availability.

Combining the upgraded TaiShan 950 SuperPoD and context memory storage to power an ultrascale cluster with 1 million NPUs

As SOTA models scale to 10 trillion parameters, training and inference can no longer rely on a single AI server or AI SuperPoD – they require a more complex computing system. This system includes AI SuperPoDs, general-purpose SuperPoDs, and an interconnect system that features peer-to-peer interconnect and zero protocol conversion. For inference, including a petabyte-scale KV cache cluster is also a must.

To meet these demands, Huawei has fully upgraded its TaiShan 950 SuperPoD. Powered by UnifiedBus all-optical networking, this new SuperPoD supports up to 4,096 nodes with a unified memory pool of up to 256 TB. This setup significantly improves agent performance. For sandbox-intensive workloads, startup speeds for 100,000 sandboxes are 30 times faster than traditional servers, and sandbox density can be improved by an additional 25%. For vector search across 10 billion x 1,000-dimensional vectors, this SuperPoD delivers twice the search efficiency of traditional servers.

Huawei has also launched OceanStor M900 – a UnifiedBus-powered context memory storage cluster that delivers multi-tier KV caching for agent-heavy and longer-context workloads. Designed for agentic inference, this cluster supports one-hop direct access and provides a petabyte-scale KV cache for the L3.5 layer. OceanStor M900 also uses hybrid media and an optimized retention algorithm, extending SSD read/write lifespan by 16-fold. This ensures a higher KV cache hit rate alongside long-term stability and reliability from the ground up.

Combining its strengths in computing and communications, Huawei has built a brand-new agentic SuperCluster to accelerate training and inference for 10-trillion-parameter models. This SuperCluster uses UnifiedBus to consolidate multiple interconnect protocols into a single unified protocol, significantly reducing protocol conversion overhead. This delivers peer-to-peer interconnect between subsystems like Ascend SuperPoDs, Kunpeng SuperPoDs, and KV cache clusters. The SuperCluster also comes with a multi-tier, high-bandwidth, and large-capacity storage system that enables direct single-hop access for all KV cache tiers.

With a two-tier, four-plane Clos architecture, the SuperCluster can interconnect up to 512,000 NPUs. When combined with a multi-rail topology, this cluster can support up to one million NPUs.

One of Huawei’s core strategies: Going open source and open system to build out computing ecosystems

The Kunpeng ecosystem is driving digital and intelligent innovation across a wide range of industries. To date, the Kunpeng ecosystem has attracted over 4.16 million developers and more than 7,200 ecosystem partners from around the globe. The community currently supports over 560 open-source projects worldwide. openEuler has seen more than 20 million installations, securing the largest share in China’s server OS market.

The Ascend ecosystem has reached a new inflection point. The Compute Architecture for Neural Networks (CANN) is the foundation of the Ascend ecosystem. Today, CANN has moved to sustained, community-driven open-source development, which has brought the platform from usable to user-friendly.

External CANN developers now comprise 61% of all CANN developers, outnumbering internal developers for the first time. With over 5,200 monthly active developers, the CANN community has become the most vibrant open-source community in China. What’s more, over 40 models have been natively pre-trained on Ascend and CANN, making it the only proven domestic stack capable of model pre-training.

Ascend now supports over 90 leading third-party open-source projects, including PyTorch, Triton, vLLM, and veRL. With strong support from the Linux Foundation, Ascend is the first official Chinese compute platform on PyTorch’s website. This gives developers around the world ready access to new innovations in the Ascend ecosystem.

Diverse forms of compute for ubiquitous on-device and in-vehicle AI

AI is expanding faster into all kinds of devices. To deliver an unparalleled AI experience across all scenarios, Huawei will continue to strengthen capabilities in four key areas:

First, Huawei will combine Kirin and Ascend chips to drive self-reliance and autonomy in on-device compute.

Second, Huawei will bring together Pangu models and third-party models to make on-device intelligence better and easier to use.

Third, HarmonyOS, as an Agent OS for ubiquitous intelligence, will be completely redefined from the ground up – spanning system architecture, how it operates, and interaction logic – to enable human-agent collaboration.

Fourth, Huawei will keep cultivating a diverse AI ecosystem, which is the foundation for its system agent Celia to thrive.

Huawei plans to build four on-device computing platforms: for AI phones, AI PCs, vehicles, and homes. Through cross-device and device-cloud compute synergy, Huawei will be able to provide distributed swarm intelligence, delivering integrated and continuous intelligent services across personal mobile, office, vehicle, and home spaces, ultimately bringing intelligence to every person and every space.

Building next-generation communications that prioritize readily available compute, because without networks, all compute is siloed

Next-generation communications networks are crucial for fully unleashing the value of AI compute. We are driving the upgrade to networks that, in addition to connecting people, will prioritize delivering readily available compute. These networks will be underpinned by 5G-A/6G, 10-gigabit optical networks, and multi-tier, low-latency bearer networks, delivering intelligent connectivity across data centers, the edge, and devices.

Concluding his keynote, Wang expressed that AI “may well be the final technological revolution in human history,” noting that its impact is deeper and broader, and coming faster than anyone could have ever imagined. “No single company,” he said, “can build an intelligent world alone.”

He stressed Huawei’s ongoing commitments moving forward:
Huawei will remain committed to building a solid silicon foundation to make computing power readily accessible to all.
The company will continue to open source its software, helping developers unleash their full potential.
It will continue to embrace a wide range of models and applications, unlocking value in every form.

“And we will continue to work together to drive shared success, growing together with our customers and partners around the world,” Wang concluded. “Let’s work together to build a fully connected, intelligent world.”

Themed Advancing the Agentic World, HUAWEI CONNECT 2026 will delve into AI across three dimensions: strategy, technology, and ecosystems. You can expect an in-depth look at our latest strategic initiatives, and we’ll also be unveiling our all-new digital and intelligent infrastructure products, scenario-specific solutions for industries, and development tools. The event will run from September 17 to 19 at the Shanghai World Expo Exhibition & Convention Center and Shanghai Expo Center. For more information, please visit HUAWEI CONNECT 2026 online at www.huawei.com/en/events/huaweiconnect

FAQs:

Q1: What is a SuperPoD, and why is it becoming increasingly important?

A SuperPoD is a computing system in which multiple computing nodes are tightly coupled through high-speed interconnect, enabling them to share unified memory and function like a single computer. As foundation model training and inference continue to scale up, SuperPoDs can reduce communications overhead in large-scale clusters and improve Model FLOPs Utilization (MFU). They have gained broad consensus across industry, academia, and research institutes in AI infrastructure, and are the go-to choice for AI infrastructure buildout.

Q2: What makes the Atlas 960E SuperPoDs special?

The Atlas 960E SuperPoD is the industry’s first NPO-based SuperPoD. A single Atlas 960E SuperPoD can scale up to 4,096 NPUs, delivering 8 EFLOPS of FP8 compute performance, with up to 1 petabyte of HBM capacity. With 5,500 Hi-ONE units, this SuperPoD doesn’t need the 48,000 800G optical modules that would traditionally be required to connect all the NPUs. This cuts power consumption by over 550 kilowatts, while doubling the system’s fault-free operating time, achieving 99.8% system availability. Atlas 960E SuperPoDs can provide efficient and reliable computing power for large-scale AI training and inference.

Q3: What is NPO, and what role does Hi-ONE play in a SuperPoD?

NPO stands for Near-Packaged Optics, an optical interconnect technology designed for high-speed connectivity. Hi-ONE, developed by Huawei, is the industry’s first NPO product ready for mass production. It delivers the largest transmission capacity at 7.2 Tbit/s and is currently the industry’s only NPO product with a built-in light source. Hi-ONE, coupled with UnifiedBus, will make it far easier to scale up SuperPoD interconnect systems.