Huawei used its flagship HUAWEI CONNECT 2025 event in Shanghai to outline a sweeping roadmap for AI computing, anchored by its next-generation Ascend processors, a family of massive SuperPoDs and SuperClusters, and a new all-optical interconnect protocol called UnifiedBus. Eric Xu, Huawei’s rotating chairman, described the announcements as a “new paradigm for AI infrastructure” that can sustain long-term growth in compute demand.
Huawei confirmed that its monetization strategy in AI remains hardware-focused, with a broad commitment to open sourcing its software stack, including the CANN compiler interfaces, Mind toolchains, and openPangu foundation models by the end of 2025. Xu said the company’s core mission is to build sustainable computing power within the constraints of process nodes available in China.
At the chip level, Huawei revealed a three-generation roadmap beyond the current Ascend 910C. The Ascend 950 series will debut with two models: the Ascend 950PR (prefill and recommendation) shipping in Q1 2026, and the Ascend 950DT (decode and training) shipping in Q4 2026. Both share a new die design supporting FP8, MXFP8, MXFP4, and Huawei’s proprietary HiF8 format. They deliver up to 1 PFLOPS (FP8/HiF8) and 2 PFLOPS (MXFP4), with 2 TB/s interconnect bandwidth. The 950PR integrates HiBL 1.0 HBM for cost-efficient memory, while the 950DT incorporates HiZQ 2.0 HBM with 144 GB capacity and 4 TB/s bandwidth. The Ascend 960, due Q4 2027, doubles compute and bandwidth relative to 950, and adds HiF4 (optimized 4-bit format). The Ascend 970, targeted for Q4 2028, will again double compute and interconnect, scaling to 8 PFLOPS (FP4) and 4 PFLOPS (FP8).
Huawei also unveiled its next wave of AI infrastructure. The Atlas 950 SuperPoD, built with up to 8,192 Ascend 950DT chips, will deliver 8 EFLOPS (FP8) and 16 EFLOPS (FP4) with 16 PB/s interconnect bandwidth. It enters the market in Q4 2026. Xu compared it directly against NVIDIA’s NVL144, claiming 6.7x more compute, 15x more memory, and 62x more interconnect bandwidth. By Q4 2027, the Atlas 960 SuperPoD will scale to 15,488 Ascend 960 chips, doubling compute, bandwidth, and memory to 30 EFLOPS (FP8) and 60 EFLOPS (FP4). Both platforms use a new all-optical interconnect fabric that lowers inter-cabinet latency to 2.1 µs and boosts reliability 100x.
Complementing AI infrastructure, Huawei announced Kunpeng 950 general-purpose processors (96-core/192-thread and 192-core/384-thread variants) in Q1 2026, with a 2028 roadmap toward 256-core CPUs. These will power the TaiShan 950 SuperPoD, a general-purpose cluster targeting finance workloads and mainframe replacement, available in Q1 2026. Hybrid SuperPoDs combining Kunpeng and Ascend are also in development for next-generation generative recommendation systems.
At the interconnect layer, Huawei launched UnifiedBus 2.0, a peer-to-peer, bus-grade optical protocol supporting over 10,000 NPUs per system. Huawei is releasing UB2.0 specifications to industry partners to build an open ecosystem. The Atlas 950 SuperCluster, with 64 SuperPoDs and more than 520,000 Ascend 950DT chips, will debut in Q4 2026 with 524 EFLOPS (FP8). The Atlas 960 SuperCluster, slated for Q4 2027, will scale to 1 million NPUs and 2 ZFLOPS (FP8).
• Ascend 950PR: optimized for prefill/recommendation, Q1 2026 launch, HiBL 1.0 HBM
• Ascend 950DT: optimized for decode/training, Q4 2026 launch, HiZQ 2.0 HBM
• Ascend 960: doubles compute, adds HiF4, Q4 2027 launch
• Ascend 970: doubles again, Q4 2028 launch
• Atlas 950 SuperPoD: 8 EFLOPS FP8, 8,192 NPUs, Q4 2026
• Atlas 960 SuperPoD: 30 EFLOPS FP8, 15,488 NPUs, Q4 2027
• Kunpeng 950 CPUs: 96/192-core variants, Q1 2026; 256-core model by 2028
• TaiShan 950 SuperPoD: finance/general-purpose focus, Q1 2026
• UnifiedBus 2.0: open all-optical interconnect protocol, scales beyond 10,000 NPUs
• Atlas 950 SuperCluster: 524 EFLOPS FP8, Q4 2026
• Atlas 960 SuperCluster: 2 ZFLOPS FP8, Q4 2027
“We believe that computing power is – and will continue to be – key to AI. UnifiedBus and our new SuperPoDs are designed to connect more than 10,000 NPUs as one machine, redefining the paradigm for AI infrastructure,” said Eric Xu.
🌐 Analysis: Huawei’s roadmap signals a determined push to remain competitive in AI computing under domestic semiconductor constraints. The Ascend 950–970 roadmap, with FP8 and FP4 specialization, aligns with global trends toward lower precision for inference. UnifiedBus 2.0 directly addresses interconnect scaling challenges that NVIDIA, AMD, and startups such as Cerebras and Groq are also working on. Huawei’s strategy of open-sourcing software while doubling down on hardware reflects the broader Chinese AI ecosystem’s pivot after DeepSeek-R1.
🌐 We’re tracking the latest developments in AI infrastructure. Follow our ongoing coverage at: https://convergedigest.com/category/ai-infrastructure/
