September 18, 2026, (Inside AI) — Huawei unveiled the Ascend 960 SuperPoD at its 2026 Connect conference, a massive AI computing cluster that links 15,488 Ascend AI processors across 220 cabinets. The system occupies 2,200 square meters and operates as a single logical machine. It marks Huawei's most aggressive move yet to build independent computing infrastructure free from US semiconductor restrictions.
The SuperPoD is not a single chip but a full system. It connects thousands of processors using Huawei's upgraded UnifiedBus interconnect, which can link up to 4,000 processors simultaneously. The company says the system delivers three to four times higher training and inference performance than earlier SuperPoD generations.
Eric Xu, Huawei's rotating chair, said the Atlas 950 SuperPoD would deliver 6.7 times greater computing power and 15 times more memory capacity than NVIDIA's corresponding generation, reaching 1,152 terabytes of total memory. That claim, if verified, would represent a significant shift in the global AI hardware balance.
Huawei's three-year roadmap runs from 2026 through 2028. The Ascend 950 series launches in 2026, the Ascend 960 arrives in Q4 2027, and the Ascend 970 follows in 2028. Each generation aims to double computing capacity. The company is guided by the Tau Scaling Law, an internal principle that pushes each new iteration beyond the previous generation's limits.
Two technical innovations define the Ascend 960. Near-Packaged Optics (NPO) integrates a light source directly into the package, enabling faster optical connections while cutting energy use. The upgraded UnifiedBus interconnect allows thousands of processors to work as one. Huawei also announced plans for mass production of the industry's first NPO product with integrated light sources, signaling commercial readiness rather than lab experimentation.
David Wang Tao, Huawei's rotating and acting chairman, emphasized that supernodes represent an inevitable progression toward ultra-large-scale data centers. "Supernodes represent the inevitable progression toward ultra-large-scale data centers," said David Wang Tao, Huawei's rotating and acting chairman.
The Ascend 950 series introduces new low-precision formats, stronger vector processing, and refined memory access granularity from 512 bytes down to 128 bytes. It also features 2 terabytes-per-second interconnect bandwidth, which is 2.5 times that of the Ascend 910C. The Ascend 950PR, available in card and SuperPoD server formats starting Q1 2026, supports prefill, recommendation, and agent-based applications. It uses HiBL 1.0, Huawei's proprietary low-cost HBM technology that the company says is more cost-effective than competing solutions like HBM3E and HBM4E.
The Ascend 950DT, set for launch in Q4 2026, targets training and decoding workloads with higher memory and bandwidth demands. It integrates HiZQ 2.0 HBM, offering 144 gigabytes of memory and 4 terabytes-per-second memory access bandwidth. The jump from the Atlas 950's 8,192 Ascend NPUs to the Ascend 960's 15,488 chips represents nearly a doubling of computing density within the same architectural framework.
Huawei also filed research on arXiv describing a 256K-node-class single AI computing framework. That framework extends capabilities far beyond individual SuperPoD installations. The company has introduced a SuperCluster architecture that connects multiple SuperPoD units. Huawei describes the expanded system as operating as one unified machine rather than a network of over 200,000 separate machines. This design addresses latency and synchronization challenges that plague conventional distributed computing at extreme processor counts.
The previously announced Atlas 900 A3 SuperPoD, equipped with 384 Ascend 910C NPUs, achieved processing power of 300 PFLOPS. That system demonstrated Huawei's escalating capability trajectory. However, building such systems without access to NVIDIA GPUs or advanced US semiconductor technology requires alternative component strategies. Huawei has spent years developing its own HBM, interconnect, and packaging technologies to fill those gaps.
The Ascend 960 builds directly upon the Atlas 950, which houses 8,192 Ascend NPUs spanning 128 computing racks across approximately 1,000 square meters. The architectural approach differs fundamentally from distributed systems. Huawei is developing a large-scale AI computing cluster similar to NVIDIA's DGX SuperPOD but using Huawei's proprietary Ascend series AI chips. The company positions the SuperPoD as a single logical machine, not a network of separate servers.
Industry analysts note that Huawei's claims remain difficult to verify independently. The company operates under US export controls that limit its access to advanced chipmaking tools. Whether Huawei can mass-produce the Ascend 960 at scale without those tools remains an open question. The company has not disclosed pricing or specific customer commitments for the SuperPoD.
Huawei's push comes as global demand for AI computing infrastructure continues to surge. Data center operators face growing pressure to secure enough processors for training and inference workloads. If Huawei's SuperPoD delivers on its promises, it could offer an alternative for buyers outside the US sphere of influence. The company's next major milestone will be the Ascend 960's Q4 2027 launch, followed by the Ascend 970 in 2028.