Inside Huawei’s UnifiedBus architecture that unites SuperPoDs for agentic AI

Details: Staff Writer; Category: Enterprise Ecosystems; 24 September 2026; 562 views

The swift approach of agentic AI is driving up demand for computing power. At Huawei Connect 2026 event in Shanghai, Huawei revealed its latest innovations in developing AI infrastructure centred around SuperPoDs and now SuperClusters, which the company sees as the necessary silicon foundation to support the era of agentic AI and beyond.

A crucial enabling component of that architecture is Huawei’s new UnifiedBus computing architecture, which was unveiled Huawei's Executive Director of the Board and CEO of the ICT Business Group, at Huawei Connect 2026.

Yang also announced a range of new UnifiedBus-powered devices that enable interconnection within cabinets, between cabinets, and across clusters, which unlocks elastic scaling from single cabinets to a new agentic AI SuperCluster that enables the diverse compute collaboration and tiered resource pooling needed to support agentic AI.

The challenge of agentic AI

Agentic AI outstrips generative AI in terms of traditional compute resources. Agentic AI systems can autonomously plan, invoke other tools, and execute tasks multiple times independently. That means it requires far more frequent data exchange between central processing units (CPUs), neural processing units (NPUs), memory, and storage devices compared to generative AI. This places higher demands on communications efficiency and resource coordination.

“With the adoption of agentic AI, parameter scale models are projected to grow from trillions to potentially 50 trillion parameters, while context lengths are expected to expand significantly, reaching tens of millions of tokens,” Yang said in his keynote.

In a traditional computing architecture, model FLOPs utilization (MFU) declines as a cluster scales. As little as 20% of a 100,000-NPU cluster's computing capacity is actually utilized by models, with a vast amount of compute sitting idle during data communications.

“A 10 trillion-parameter model, which requires massive amounts of intermediate data to train, far exceeds the capacity of the memory on a single accelerator under traditional architectures,” Yang said. “Furthermore, AI agents can have hundreds of interactions per hour with AI models, which means computing systems need to be able to flexibly utilize changing amounts of CPUs and NPUs.”

This issue is already being noted in the broader industry. A 2026 research report from Wedbush Securities found that the primary bottleneck in AI infrastructure has shifted from the chips themselves to the networking fabric required to connect them. A recent Forbes article similarly observed that connectivity has replaced compute and power as the defining constraint of this cycle.

Additionally, the resulting interactions between these processors generate KV cache that far exceeds memory capacity, which is driving up demand for data storage across compute systems and tiers. Indeed, various industry data platform vendors have described inference as more of a a GPU memory problem than a GPU cycles challenge, where a single 100,000-token sequence can consume roughly 40GB of GPU memory – nearly half of an enterprise GPU’s HBM capacity.

All of this means interconnection and communications between components within a cluster have become a key performance bottleneck for computing systems.

Inside the UnifiedBus

Huawei’s solution to this bottleneck is its UnifiedBus architecture, which facilitates collaboration between clusters and SuperPoDs.

UnifiedBus has four key features, starting with unified protocol and memory semantics. More than ten interconnect protocols have been combined under the UnifiedBus protocol, which increases interconnect bandwidth from the 100-Gbps level to the terabit range, and reduces round-trip time latency from 7 microseconds to 2 microseconds. The unified protocol also enables unified global memory addressing within SuperPoDs.

UnifiedBus also enables heterogeneous compute collaboration by directly interconnecting CPUs, NPUs, memory, and solid-state drives (SSDs) to enable decentralized, peer-to-peer access between them and flexible CPU-NPU mixing. Also, tiered hardware acceleration for Transformer enables Attention-FFN disaggregation (AFD).

This is key, because third-party analysis confirms a structural shift of CPU-to-NPU ratios is already underway. TrendForce reports that CPU-to-GPU ratios in AI server platforms are moving from roughly 1:8 toward 1:4 or even lower, as agentic AI workloads place greater demands on CPU scheduling and memory management. In fact, for agentic and multi-agent workloads, the CPU-to-GPU ratio could potentially flip in favour of CPUs.

UnifiedBus also features tiered storage with global pooling, using hybrid-media resource pooling to support the caching of activations. It uses Double Data Rate (DDR) memory as the alternative memory for NPUs – this reduces latency for search, recommendation, and advertising services, and doubling the performance of vector retrieval for 100 billion-level data items with thousands of dimensions. It also reduces the per-NPU HBM capacity requirements for the training of 10 trillion-parameter models, and improves cluster MFU.

UnifiedBus also makes use of superfast optoelectronic interconnect technology like Huawei’s new Hi-ONE 7.2T near package optics (NPO) product, enabling it to serve as a global ultra-high bandwidth, ultra-low latency "data highway", enabling flexible compute scaling.

Overall, UnifiedBus differs from existing interconnect approaches by extending unified protocol and memory semantics across CPUs, NPUs, memory, and storage – not just from accelerator to accelerator.

New UnifiedBus products

Yang also unveiled Huawei's latest UnifiedBus products that enable not only interconnection within cabinets, between cabinets, and across clusters, but also elastic scaling from a single cabinet to a million-NPU cluster.

For in-cabinet interconnect, the UnifiedBus LinkBlade eliminates circuit losses with a cable-free design, which reduces copper cabling within a 4,096-NPU SuperPoD by about 196 km.

For between-cabinet interconnection, the UnifiedBus LinkDevice supports 176 ports, with 1.6-Tbps bandwidth per port, delivering 280 Tbps of all-optical interconnection with an RTT latency as low as 2 microseconds.

For interconnect across clusters, the UnifiedBus UBG switch provides a massive radix fan-out capacity of up to 1,024 nodes. “That means it can support a million-NPU SuperCluster and evolution towards supporting tens of trillions of model parameters,” Yang said.

Agentic AI SuperCluster

This ties in to the launch of Huawei’s new agentic AI SuperCluster at Huawei Connect, which comprises newly launched individual solutions including the Atlas 960E SuperPoD (with the Hi-ONE NPO and Ascend 960 chips), the upgraded TaiShan 950 SuperPoD that now supports UnifiedBus, the OceanStor M900 context memory storage cluster that leverages UnifiedBus to deliver multi-tier KV caching for agent-heavy and longer-context workloads, and a Xinghe UBG switch.

The UnifiedBus architecture enables SuperCluster to deliver 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, it can support up to one million NPUs.

Compute appliances and open ecosystems

Beyond SuperPoDs, Huawei has also applied the UnifiedBus technology to its compute appliances.

For example, the company has developed a new series of UnifiedBus-based appliances (supporting one to eight NPUs) and Kunpeng and Ascend modules. Two Atlas 650E air-cooled servers can be directly linked through UnifiedBus to enable switch-free interconnect across 16 NPUs, allowing them to work seamlessly as a single unit. They can also scale smoothly to form a small-scale SuperPoD that enables small and medium-sized enterprises to run trillion-parameter models locally.

Yang also spoke about Huawei’s activities on the ecosystem front to enable development of open-source software and applications to run on its agentic AI architecture.

For example, Ascend is collaborating with open-source frameworks like DeepSeek Harness, OpenCode, and openJiuwen to develop fully open-source intelligent management, inference acceleration, and security framework agent plugins. Ascend has also worked with more than 50 customers, partners, and universities to create 26 industry-specific operator libraries, and has collaborated with over 90 mainstream open-source communities to help partners quickly develop industry-specific agents and integrate them with IT systems.

Emerging AI applications are making compute the cornerstone of industrial intelligent transformation, Yang observed.

“Huawei is committed to innovating at the system level, building a computing-foundation product portfolio, and going open source and open system,” he said. “We will continue working with customers, partners, and developers to help this ecosystem thrive and offer the world a new option for compute.”