Huawei pulls Ascend 960DT forward by three quarters and launches the industry's first NPO-based SuperPoD

Start with the optical line, because it is the most concrete piece of the launch. Huawei names its next-generation NPO optical interconnect product Hi-ONE, the High-density Optical-interconnect-Node Engine. It is built on proprietary technology with a multi-physics design balancing optical, mechanical, electrical, electromagnetic and thermal performance, and reaches a transmission capacity of 7.2 Tbit/s per single engine. Huawei's claim is that this is the industry's first NPO product ready for mass production and the first NPO product with a built-in light source. Combined with UnifiedBus, its purpose is to make scaling SuperPoD interconnect systems far easier. Huawei recently submitted an implementation agreement on NPO to the Optical Internetworking Forum and says the response from industry partners has been widely positive.

That optical work directly shapes the Atlas 960E. A single Atlas 960E SuperPoD can scale up to 4,096 NPUs, delivering 8 EFLOPS of FP8 compute with up to 1 petabyte of HBM capacity. The detail that matters is that it uses 5,500 Hi-ONE units where connecting those NPUs the traditional way would require 48,000 800G optical modules. Huawei's conclusion is that this cuts power consumption by more than 550 kilowatts while doubling fault-free operating time, achieving 99.8% system availability. In other words, the value of NPO is not a faster single engine but an order-of-magnitude reduction in the number of interconnect parts, and part count directly determines the power, failure points and cabling complexity of a large cluster. At the tens-of-thousands-of-cards scale, all three are costs you can only discuss once they are solved.

Beyond the chip roadmap, Huawei filled in a relative weak spot: memory and context storage built for agents. The TaiShan 950 SuperPoD has been fully upgraded on UnifiedBus all-optical networking, supporting up to 4,096 nodes with a unified memory pool of up to 256 TB. Huawei offers three comparisons: for sandbox-intensive workloads, startup speed for 100,000 sandboxes is 30 times faster than traditional servers with sandbox density improved by a further 25%; and for vector search across 10 billion 1,000-dimensional vectors, search efficiency is twice that of traditional servers. Alongside it, OceanStor M900 is a context memory storage cluster providing multi-tier KV caching for agent-heavy and longer-context workloads, with one-hop direct access, a petabyte-scale KV cache at the L3.5 layer, and hybrid media plus an optimised retention algorithm that Huawei says extends SSD read/write lifespan 16-fold. What these figures share is that they all answer one question: as contexts grow longer and agents multiply, what holds the intermediate state.

What ties the systems together is the SuperCluster. It uses UnifiedBus to consolidate multiple interconnect protocols into a single unified protocol, significantly reducing protocol conversion overhead, which enables peer-to-peer interconnect between Ascend SuperPoDs, Kunpeng SuperPoDs and KV cache clusters, aimed at training and inference for 10-trillion-parameter models. Huawei also cites simulation results from its Markov Lab: a 100k-NPU cluster built from 4k-NPU SuperPoDs can deliver a 2.75x increase in MFU compared with a 100k-NPU cluster composed of 8-NPU servers. One counterpoint has to travel with all of this. China tech analyst Rui Ma noted on social media that Huawei had previously said its Atlas 960 SuperPoD would scale to 15,488 Ascend 960 chips, while this week's announcement referred to a system with 4,096 chips; her assessment was that the chip itself is coming way earlier, but the SuperPoD announced is much smaller than originally laid out. That gap is central to reading the launch: a pulled-forward roadmap and a shrunken system happened at the same time, and citing only half of it distorts the picture.