WAIC Spotlight: China's AI Chipmakers Shift from Flexing Specs to Building Full-Stack Systems, with SuperPods and Open-Source Software Stacks Taking Center Stage

“We have always believed that the true explosion of the AI era will not come from one or two chips, but from a more open, coordinated, and efficient full-stack computing system,” said Gao Hui, Vice President of T-Head Semiconductor, at the 2026 World Artificial Intelligence Conference (WAIC) on July 18. Her remarks precisely captured the core narrative driving China's domestic AI chipmakers this year.

Unlike previous years, when vendors were keen on comparing peak single-card computing power, this year's WAIC sent a clear signal: competition in China's domestic AI computing sector has fully entered the "second half," with the focus shifting from standalone hardware performance to system-level, full-stack efficiency. SuperPoD clusters, open-source AI software stacks, and infrastructure innovations for the agentic AI era became the hottest keywords at the conference.

Open-Source Software Stacks: Unblocking the Bottlenecks to Unleashing Computing Power

In recent years, China's domestic AI chips have made significant progress at the hardware level, but efficiently converting on-paper specs into real-world application performance has remained a major hurdle for developers. At this year's conference, T-Head provided its answer.

Gao Hui announced that T-Head is officially open-sourcing its proprietary AI software stack, T-Head SAIL ("SAIL"), which is the core low-level software for its Zhenwu AI chips. According to Gao, the SAIL software architecture covers the complete chain from the OS layer to the SDK and interface layers. Its core mission is to maximize chip computing power release and efficiently support upper-layer application demands. In terms of ecosystem compatibility, SAIL is fully adapted to mainstream AI ecosystems and is compatible with over 260 mainstream training and inference frameworks.

"Open-sourcing the AI software stack is just the beginning. T-Head hopes to work with global developers, research institutions, and industry partners to jointly refine toolchains, enrich operator ecosystems, and optimize performance experiences," Gao said. This move is seen by the industry as a critical step for China's domestic AI computing sector to address its software ecosystem shortcomings. If an AI chip is the "engine," then the AI software stack is the "transmission, drivetrain, and control system"—its quality directly determines the overall driving experience.

SuperPods Take the Spotlight: A Systemic Evolution from Single Machines to Clusters

If the software stack represents invisible soft power, then SuperPoD products were the tangible symbols of hard power at this year's WAIC. As large model competition shifts from parameter scale to agentic AI application deployment, the challenges facing AI infrastructure have changed dramatically. Scenarios involving ultra-long context, multi-turn reasoning, and high-frequency multi-agent collaboration can easily lead to idle AI chips, reduced computing utilization, and a surge in wasted energy. Simply stacking more chips can no longer solve these systemic problems, giving rise to the SuperPoD architecture.

Huawei: Ascend 950 Takes the Baton, Over 750 384 SuperPoDs Deployed

Huawei publicly exhibited its Ascend 950 SuperPoD (Atlas 950 SuperPoD) physical machine for the first time at the conference. According to information gathered by Yicai from the booth, the product uses a single cabinet with 64 cards as its basic unit, enabling high-speed interconnection of 1,024 Ascend NPU cards. It delivers 1 EFLOPS of FP8 and 2 EFLOPS of FP4 computing power, with 256TB of globally unified memory addressing space. Huawei stated that the product, leveraging TB-level NPU interconnect bandwidth, 3-microsecond RTT latency, and unified memory addressing capabilities, provides robust support for trillion-parameter large model training and high-concurrency inference scenarios.

Regarding commercial deployment, Huawei revealed that over 750 units of the Ascend 384 SuperPoD, released in 2025, have been commercially deployed, covering numerous industries including the internet, telecommunications, finance, and education. The company emphasized that this is the only SuperPoD in China to have trained a state-of-the-art (SOTA) model. Additionally, Huawei showcased the Atlas 850E air-cooled SuperPoD, which utilizes its proprietary VCE phase-change cooling technology and can be deployed directly in traditional air-cooled data centers, reducing the retrofitting costs for enterprise AI infrastructure.

Alibaba's T-Head: Backed by 560,000 Chip Deliveries, Public Cloud Debuts SuperPoD Service

Alibaba's T-Head and Alibaba Cloud also presented a heavyweight combination. The Zhenwu M890 × Panjiu AL128 SuperPoD exhibited at the conference was selected as one of the top ten "Treasures of the Exhibition" at WAIC 2026.

According to an introduction by Alibaba Cloud staff to Hongxing Capital Bureau, the Panjiu AL128 SuperPoD server adopts a high-density design with 128 cards per cabinet. Through an orthogonal cable-free architecture and the proprietary ALink System interconnect protocol, it achieves Pb/s-level bandwidth and sub-hundred-nanosecond latency. Notably, the server uses a multi-dimensional decoupling architecture, completely separating CPU, GPU, and ALink switching nodes. It is compatible with China's domestic AI chips and mainstream industry GPUs, offering a gradual path toward autonomous and controllable infrastructure. For equivalent AI computing power, its inference performance can be improved by 50% compared to traditional architectures.

The Alibaba Cloud Lingjun Zhenwu M890 SuperPoD instance was also officially launched at the conference, marking the first time Alibaba Cloud has offered SuperPoD-form AI computing services through its public cloud. The hardware foundation is T-Head's self-developed Zhenwu M890 chip. This chip natively supports multiple data precisions from FP32 to FP4, covering all scenarios from high-precision training to ultra-low-precision inference. As of April this year, cumulative deliveries of Zhenwu AI chips have exceeded 560,000 units, serving over 400 customers across more than 20 industries, with applications spanning autonomous driving, finance, telecommunications, and embodied intelligence.

Elsewhere, MetaX launched its next-generation AI SuperPoD product, the Xijing S600, at the conference. ZTE also released its new ZTE OEX SuperPoD and collaborated with Biren Technology and MetaX to create the Matrix SuperPoD.

A Paradigm Shift in Competition: System-Level Efficiency Will Determine the Winners

The collective shift by China's domestic AI chipmakers toward system-level competition is driven by profound changes in the market landscape. Guosheng Securities noted in a recent research report that China's domestic SuperPoDs are entering their first year of mass production. The SuperPoD architecture expands effective computing power through system engineering, signaling a gradual paradigm shift in computing power competition. This means that cluster engineering capabilities and the software ecosystem have become the new core competitive advantages.

Huatai Securities estimates that the market size for China's domestic SuperPoDs could reach 341.4 billion yuan (approximately $50.4 billion) by 2028. TrendForce projects that in 2026, domestic players like Huawei and Cambricon will collectively hold nearly 80% of China's AI server chip market share, with the share held by overseas vendors declining further.

However, moving from exhibition displays to large-scale commercial use, China's domestic SuperPoDs still need to bridge the gap of engineering deployment. Challenges such as ultra-high-bandwidth cross-node interconnects, heat dissipation and energy management for large-scale clusters, and the appeal of the software ecosystem to developers are all areas that require continuous effort. As Gao Hui stated, the true explosion of the AI era will not rely on breakthroughs in just one or two chips, but on whether an open, coordinated, and efficient full-stack system can be built. The 2026 WAIC witnessed this system accelerating from blueprint to reality.