Huawei와 Alibaba가 AI 가속기, 클러스터 인터커넥트, 스토리지 기능에 대한 진전을 발표합니다.

Chinese tech companies are advancing AI infrastructure at scale, with Huawei and Alibaba making significant announcements on chips, clusters, and models.

At its Huawei Connect 2026 conference in Shanghai on September 17, Huawei unveiled products designed to deliver AI inference at scale and enable training of AI models with up to 10 trillion parameters. These announcements are part of the company's new Peerium Computing Architecture, spanning servers, storage, and networking.

On the storage front, Huawei introduced the OceanStor M900 Context Memory Storage, a system designed to deliver petabyte-scale key-value caches with global one-hop connections for AI models running on Huawei SuperPoDs. Supporting up to 64 PB of SSD storage, the M900 combines CPU, network controller unit, and NAND controllers into a single unit, delivering 40 TB/s of aggregate bandwidth in a single cluster. Each Huawei Ascend neural processing unit (NPU) connected to an M900 can support terabytes of KV cache, enabling single context windows to exceed one million tokens.

The Peerium architecture centers on the Unified Bus, a new interconnect shipping in the OceanStor M900 for scale-up, scale-out, and scale-across connections. The Unified Bus consolidates multiple network protocols and delivers terabytes per second of bandwidth across copper and optical connections.

For scale-up interconnection, Huawei offers the UnifiedBus LinkBlade, a cable-free design connecting up to 4,096 NPUs in a SuperPoD. Scale-out is handled by the UnifiedBus LinkDevice, an optical switch delivering 1.6 Tbit/s bandwidth per port across 176 ports for 280 Tbit/s aggregate bandwidth. The UnifiedBus UBG switch connects up to 1,024 end devices, enabling Huawei to build its million-NPU SuperCluster and allowing customers to train AI models with up to 10 trillion parameters.

Huawei rolled out new AI SuperClusters, including the TaiShan 950 and Atlas 960 SuperPoDs, which combine the OceanStor M900 and Xinghe UBG switch. The Atlas 960 SuperPoD will be the industry's first server to feature near-packaged optical (NPO), the company said.

The company also introduced UnifiedBus-based appliances, Kunpeng and Ascend, supporting one to eight NPUs. Smaller operations requiring AI model training up to 1 trillion parameters can now create a small-scale SuperPoD using two Atlas 650E air-cooled servers connected directly through UnifiedBus with switch-free interconnect across 16 NPUs.

More than 1,000 Atlas 900 A3 SuperPoDs have been deployed to date, and the Atlas 950 SuperPoD is seeing large-scale commercial use. Huawei is currently developing the latest Ascend 960 chips, with delivery expected in early 2027.

"We're evolving our Ascend chip series on a one-generation-a-year cycle," said David Wang, Huawei's deputy chairman of the board and rotating chairman. "In 2028 and 2029, we will roll out the Ascend 970 and 980 chips, respectively. Thanks to the Tau Scaling Law, not only will their compute specifications continue to double, but you can also expect to see huge improvements across the board in terms of memory bandwidth, memory capacity, interconnect bandwidth, and more."

Huawei is among China's leading semiconductor companies and the top phone maker. The company's products are banned in the US but are exported to Europe and other regions. In May, then-rotating chairman Xu Zhijun commented on the US export restrictions: "If the United States hadn't forced our country, our companies, and our industry, we wouldn't have done something like this. But we are also grateful to the US for enabling our country's semiconductor industry chain to truly grow."

At its annual Apsara Conference, Alibaba, China's number-one public cloud provider, also announced significant chip and AI developments.

Alibaba's T-Head chip design unit unveiled the Zhenwu V900, its latest AI training and inference processor. The Zhenwu V900 delivers 216 GB of GPU memory and 1.2 TB/s of inter-chip bandwidth, three times that of its predecessor, the Zhenwu M890. The chip supports FP8 and FP4, handling mixed AI workloads, and is scheduled for release in the first quarter of 2027.

Alibaba also unveiled an upgraded "supernode server" integrating the Zhenwu V900 processor, ICN Switch, Panmai SmartNIC, and Zhenyue SSD controller chip, enabling clusters composed of up to 500,000 computing cards.

The company shared its roadmap for the next-generation Yitian 720 CPU, scheduled for launch in 2027, which features enhanced single-core performance, higher core density, and increased energy efficiency compared to the Yitian 710. Alibaba also announced the Yitian 730, developed on T-Head's proprietary microarchitecture, which will deliver a 40% performance increase over the Yitian 710.

On the AI front, Alibaba announced that its next-generation model, Qwen 4, is currently in training. The company stated that upcoming Qwen 4.5 and Qwen 5 model series will scale up to 10 trillion parameters.

Qwen models are improving through Alibaba's focus on recursive self-improvement (RSI). Qwen 3.8-Max completed more than 33 RSI cycles over a month of fully automated training runs, boosting its Artificial Analysis score from 40 to 45. In chip design applications, Alibaba subjected Qwen to more than 60 hours of RSI across 10,000 electronic design automation (EDA) tool calls, resulting in a chip with 42% reduced size "with zero compromise in performance."

Several new Qwen AI models were introduced at the conference: Qwen 3.8-LiveTranslate, Qwen-Audio-3.1-TTS (Text-to-Speech), Qwen-Audio-3.1-Realtime, Qwen-Audio-3.1-TTS-Next, Qwen-Audio-3.1-ASR, and Qwen-Image 3.1.