The brand KunTai from Digital China introduced a new product matrix featuring Kunpeng and Ascend 950 at the HUAWEI CONNECT 2026 event. This lineup covers both general-purpose systems and systems for artificial intelligence training and inference. Among the presented solutions are the KunTai R722 K3 server with Kunpeng 950 balancing, as well as three machines based on Ascend 950DT: the standard KunTai A989 I5 server, the liquid-cooled supernode KunTai PoD2000 A5, and the air-cooled supernode A989 I5.
Product specialist Sun Yanang described this event as the transition of supernodes from a marketing concept stage to ready-to-select SKUs for customers, applicable to both general computing and large model workloads.
The PoD2000 A5 supernode is aimed at large, new intelligent computing centers and the pre-training of trillion-scale Mixture-of-Experts models. According to stated specifications, it supports a cluster of 1024 cards, features a unified memory address space of 256 TB, and achieves an FLOPs utilization rate of about 45%, which is 30% higher compared to previous generations, according to KunTai. It is equipped with liquid cooling plates, rack-level orthogonal layout, Nebula 8000 LPO optical interconnects, and dual optical path protection.
The A989 I5 air-cooled supernode is designed for reusing existing air-cooled spaces without the need to upgrade liquid cooling piping. Its specifications include performance up to 14.2 PFLOPS at MXFP4, an NPU fabric interconnect bandwidth of approximately 13.4 TB/s, a declared latency of about 5 ms for multi-agent system coordination, and an inference throughput approximately 1.5–2 times higher than previous A3 nodes.
The standard 16-card A989 I5 server, featuring a full mesh topology, is used for smaller inference clusters and medium-sized fine-tuning, providing direct NPU-to-NPU connections.
KunTai asserts that this supernode must meet three strict criteria: the presence of a dedicated high-speed hardware fabric, such as Lingqu UnifiedBus, instead of Ethernet RoCE message stacking; a globally unified memory address space with load/save semantics and no copying overhead; and end-to-end design of software and hardware covering chips, interconnects, CCU, drivers, OS, and communication libraries.
Furthermore, the company emphasizes that the gain in low-precision accuracy of MXFP4 and MXFP8 on the Ascend 950DT requires full stack wiring through hardware, drivers, and the CANN stack, not just chip mathematical blocks.
In the general-purpose segment, the Kunpeng 950 on the R722 K3 is presented as a 3D-packaged CPU that offers increased memory bandwidth, SVE2 and INT8 matrix optimizations, and confidential computing options for the financial sector, telecommunications operators, and large databases.
The main focus of this material is on the OEM product matrix from Digital China KunTai and the engineering solutions of PoD2000 A5, rather than the commercial dates of Huawei Cloud Ascend 950 Lingqu or the announcements of Ascend 960 SuperPoD, which have already been covered by Pandaily.
