Huawei Join 2025 was greater than displaying off new merchandise. Eric Xu defined from the stage that export controls and provide chain constraints have compelled Huawei to construct its personal home provide chain and outline its personal path to marketplace for tech infrastructure or danger obsolescence. This turned Join bulletins into statements of strategic autonomy from Western know-how hegemony, for Huawei and its clients.
Geopolitical Constraints Catalyze Jap Innovation
Huawei’s push into AI infrastructure has been formed by US export controls that block entry to Nvidia’s high-end GPUs and ASML’s EUV lithography machines. These constraints prompted Huawei to innovate throughout the stack with:
- Ascend AI NPUs: The Ascend 910C powers Huawei’s CloudMatrix 384 “supernode” cluster, rivaling Nvidia’s GB200 NVL72 in compute efficiency by stacking 384 chips with high-speed interconnects, and Huawei has an extended roadmap of enhancements.
- ARM-Based mostly Kunpeng Processors: Huawei’s in-house ARM CPUs now energy laptops, servers, and AI infrastructure, lowering reliance on x86 and enabling tighter integration with its software program ecosystem.
- HarmonyOS: Huawei AI-native working system powers cellular gadgets, laptops, even AI infrastructure elements. It helps real-time inference, clever scheduling, and context-aware providers, making it a strategic asset of their push for technological sovereignty.
- UnifiedBus: Huawei opened up technical specs for UB 2.0. The brand new spec is designed to hyperlink greater than 10,000 NPUs in a SuperPoD and can scale to one million NPUs in future SuperClusters. UB is supposed to outscale current NVLink and RoCE based mostly interconnects.
An AI Future With Ascend, Kunpeng, TaiShan, and Atlas
Ascend NPU Roadmap
The Ascend NPU roadmap extends into 2028 and every chip has corresponding SuperPoD and SuperCluster configurations. One of many dazzling guarantees embody the Atlas 960 SuperCluster anticipated to attain a Zettaflop of compute energy with nearly one million clustered NPUs in 2027.
- Ascend 950PR, Q1 2026: (prefill/suggestion) 1 PFLOPS FP8, 2 PFLOPS FP4, 2 TB/s interconnect, 128GB reminiscence at 1.6 TB/s
- Ascend 950DT, This fall 2026: (decode/coaching) 1 PFLOPS FP8, 2 PFLOPS FP4, 2 TB/s interconnect, 144GB reminiscence at 4 TB/s
- Ascend 960, 2027: 2 PFLOPS FP8 , 4 PFLOPS FP4, 2.2TB/s interconnect, 288GB reminiscence at 9.6 TB/s
- Ascend 970, 2028: 4 PFLOPS FP8 , 8 PFLOPS FP4, 4TB/s interconnect, 288GB reminiscence, at 14.4TB/s
TaiShan Product Portfolio
The TaiShan 950 SuperPoD is a general-purpose compute pod designed to deal with enterprise-grade workloads like software servers, database operations, and enterprise workloads with AI-inferencing help. It will likely be powered by Kunpeng 950 processors, out there in 96-core/192-thread and 192-core/384-thread variants. Anticipated in Q1 2026.
Atlas SuperPoD Servers and NPU Playing cards
Atlas SuperPoD nodes are targeted on enterprise AI adoption and can compete with Nvidia-based AI Servers from Dell, HPE, and Lenovo. The product line maps to Ascend NPU availability and features a PCIe 5.0 card possibility for upgrading current servers.
- Atlas 850 SuperPoD Server: Air-cooled design powered by eight Ascend NPUs, clustering as much as 128 nodes for a complete of 1024 NPUs in a single SuperPoD. Anticipated in Q1 2026.
- Atlas 860 SuperPoD Server: Iteration on the 850 with larger reminiscence bandwidth, and enhanced interconnect for extra demanding inference duties. Anticipated in Q1 2026.
- Atlas 350 NPU Card: Powered by the Ascend 950PR, it helps standalone use or related through UnifiedBus in 2 or 4 card configurations permitting as much as 512GB of reminiscence for giant parameter fashions. Anticipated in Q2 2026.
Storage Innovation With AI, Flash, and Improved Information Resilience
Huawei’s storage portfolio displays rising capabilities in enterprise storage, knowledge safety, and fleet degree datacenter administration with DME:
- OceanStor A600: Excessive efficiency AI-native storage delivering 160 GB/s and three.6M IOPS. It helps vector retrieval and multi-level KV cache to scale back each hallucinations and latency
- OceanStor All-Flash Dorado v7: Up to date platform with PCIe 5.0, scalability to 500PB and 100M IOPS, and help for 32 controllers utilizing SmartMatrix 4.0 for improved reliability.
- OceanStor Pacific 9926: Bringing all-flash efficiency to the Pacific scale-out storage portfolio. 8PB in a 2U chassis utilizing 61TB SSDs, Anticipated September 30, with 122TB SSDs in 2026.
- OceanProtect Equipment: Enhanced knowledge safety answer deployed as software program and now in a brand new equipment providing.
- OceanCyber: Information safety equipment for Huawei storage platforms. Unifies safety coverage administration throughout gadgets utilizing DME, enhanced ransomware detection, and safe restoration orchestration for OceanStor and OceanProtect programs.
Networking Is AI-Centric Switching and DPUs
Huawei’s networking improvements are designed to help large AI clusters:
- Xinghe Clever Cloth 2.0: AI-Centric networking that features network-scale load balancing, self-healing, and built-in safety.
- CloudEngine Switches: 128×800GE switches with StarryLink optical modules allow high-speed SuperPoD interconnect.
- DPUs and Community AI Brokers: Community-focused compute offload for CPUs and NPUs, with AI brokers designed to automate community operations and improve safety.
What It Means For Enterprise IT Consumers
Huawei’s vertically built-in method spans chip design, working programs, and O&M software program mirroring the ecosystem benefits seen in distributors like Apple and Oracle. Whereas elements can run independently, the portfolio is optimized for joint deployment, as demonstrated on the exhibition ground by use-case targeted options.
For AI enablement, Huawei’s open-sourced AI stack with MindSpore, CANN, and openEuler come collectively to kind a full-stack different to Nvidia’s CUDA, aiming to foster a collaborative ecosystem the place others like AMD and Intel have struggled.
Huawei’s convention drew curiosity from clients effectively past APAC, with attendees from Latin America, Africa, Europe, and Center East collectively to discover sensible metropolis, grid, and enterprise IT options all through the present ground. Prospects cited affordability in case research in addition to dependable platforms and robust partnership from Huawei for deployment. It was clear that for a lot of enterprises Huawei is a critical infrastructure contender. Nonetheless, businesses working throughout areas or topic to compliance audits in areas the place Huawei know-how is restricted could face regulatory challenges. In such instances, Huawei’s platform may very well be extra of a compliance headache than a cost-saving panacea.
Subsequent Steps
In case you’re evaluating Huawei as a part of an upcoming infrastructure refresh or long-term tech technique, attain out to Forrester for tailor-made steering. Our analysts may also help you assess Huawei’s match for your corporation, navigate regional compliance dangers, and benchmark its choices in opposition to rivals.
