KunTai’s four-server lineup: From the R722 K3 to the PoD2000 A5

Digital China’s KunTai introduced four systems at HUAWEI CONNECT 2026: the Kunpeng 950-based R722 K3 for general-purpose computing, plus three Ascend 950DT AI products. For buyers, the useful distinction is what needs to run together, at what scale, and with which cooling approach. 3132

The four deployment choices

• R722 K3 — general-purpose workloads. This Kunpeng 950 server targets core transactions, large in-memory databases and data warehouses. Reporting attributes 1.4× overall server performance and 2.25× memory bandwidth to the model, and says its Lingqu interface enables general-compute networking. Those gains are reported figures, not guarantees for every workload; the R722 K3 is not one of the three Ascend AI systems. 3032
• A989 I5 standard server — smaller AI deployments. Positioned for small-cluster inference and medium-model fine-tuning, this Ascend 950DT system is reported to use 16-card full-mesh networking. The cited reporting does not establish a comparable, independently verified throughput result for it. 32
• A989 I5 air-cooled supernode — a larger interconnected AI system. Reporting describes flexible networking of up to 768 cards, a maximum of 14.2 PFLOPS and inference throughput 1.5–2× that of the preceding A3 node. Treat those as attributed product claims rather than independently reproduced benchmarks. Its air cooling distinguishes it from the liquid-cooled PoD2000 A5, but the reporting does not specify site-level power or cooling requirements. 32
• PoD2000 A5 liquid-cooled supernode — large data-center AI. Intended for trillion-parameter-model training and centralized, high-concurrency inference, it is reported to integrate 1,024 accelerator cards in one supernode, support 256 TB of unified memory addressing and deliver 3 μs interconnect round-trip latency. KunTai also reports a 300-hour mean time between failures at thousand-card scale. These specifications are not independently verified training results or an availability guarantee. 12933

What does KunTai mean by a “true supernode”?

KunTai sets out three requirements: a dedicated high-speed hardware interconnect, such as Lingqu UnifiedBus, rather than only Ethernet/RoCE links; unified memory addressing that supports remote load/store access without copying data between nodes; and integration across chips, interconnect hardware, CCU, drivers, operating system and communication libraries. By that definition, linking many conventional servers is not enough. 3032

The same full-stack point matters for Ascend 950DT’s low-precision computing: a chip’s peak figure alone cannot establish application performance. The provided reporting offers no comparable, independently verified low-precision workload result across all four KunTai models. 3032

Don’t confuse these systems with Huawei’s other announcements

Huawei Cloud said its Ascend 950 cluster service would begin commercial operation in China on September 30, 2026, and expand globally on November 30, 2026. Those are not stated availability dates for each KunTai server. 6

Huawei separately unveiled the NPO-based Atlas 960E SuperPoD and said Ascend 960DT was scheduled for the first quarter of 2027. Neither announcement changes the specifications of KunTai’s Ascend 950DT lineup. 525