The Chinese tech giant Huawei unveils new AI Infrastructure that will allow the company to compete with Nvidia and boost the computer power. Curious about Huawei AI-powered infrastructure? Let’s explore the key components and technology behind this new development.
Huawei Announces New AI Infrastructure
Huawei has organized Huawei Connect 2025 in Shanghai, China, from 18 to 20 September 2025, where it dug deeper into AI intelligence across all dimensions, technology, strategy, and ecosystem. The Deputy Chairman, Eric Xu, announces the SuperPoD Interconnect technology that will lead to AI infrastructure.
Last year, Huawei discussed that they would develop computing SuperPoDs, SuperClusters, and a new computing architecture to fulfill the demand for computing power, and now they are back with what they have promised.
The company understands that physical AI and Artificial General Intelligence would need massive computer power. The Ascend chips serve as the foundation of their AI computing strategy, and with SuperPoD and SuperClusters, the company is reshaping the future of AI infrastructure.
What is Huawei SuperPoD and its New Products?
The SuperPoD interconnect technology will link thousands of Ascend AI chips into a unified system. It is a single local machine, which will be made of multiple physical machines, that can reason, think, and learn at once.
This year, in March 2025, the company released the Atlas 900 A3 SuperPoD, which is the largest in the world as it links 384 Ascend chips and can deliver 300 PFLOPS of computing power. Now, let’s understand the new SuperPoD product and its launch:
Atlas 950 SuperPoD:
- Built on Ascend 950DT chips and can connect up to 8192 chips, which is 20 times more than the current SuperPoD product.
- It will have 160 cabinets, 32 for communication and 128 for computing, and they are connected using optical cables.
- Performance-wise, it will deliver 16 EFLOPS computing power.
- Compared to the previous SuperPoD, it will have a 17-fold improvement and deliver 4.91 million tokens per second.
- Launch: IV quarter of 2026
Atlas 960 SuperPoD
- It will be charged with Ascend 960 chips and link up to 15,488 chips, which will take another peak in the AI SuperPoDs.
- It will have 220 cabinets, 44 for communications, and 176 for computation, which will fit into 2200 square feet.
- It will deliver 60EFLOPS in FP4 and 30 EFLOPs in FP8 with 4460 Tb memory and bandwidth of 34 PB/s.
- Launch: IV quarter of 2027
TaiiShan 950 SuperPoD
- It will be built on the Kunpeng processor, which will be launched in the 2026 first quarter and become the first general-purpose computing SuperPoD.
- It will make the real-time data processing 30% faster and memory utilization 20% in the virtual ecosystem.
- Launch: 2026 First Quarter
What is Huawei SuperCluster?
The SuperCluster will connect the SuperPoDs, which will offer a network of 1 million chips, increasing the computing power. The SuperCluster and SuperPoDs will help in building new AI infrastructure that will handle large AI tasks.
The SuperCluster product that Huawei introduced at Connect 2025 is as follows:
Atlas 950 SuperCluster
- It will connect 64 Atlas 950 SuperPods, that is, more than 520,000 Ascend 950DT chips.
- It will deliver 524 EFLOPS in FP8 and support RoCE and UBoE protocols, where, with UBoE, the users will be able to use the Ethernet switches.
- This cluster will outperform the xAI’s Colossus, with 1.3 times the computing power and 2.5 times the NPUs.
- Launch: 2026 IV quarter
Atlas 960 SuperCluster:
- The Atlas 960 SuperCluster will integrate around 1 million MPUs, and it will deliver 4 ZFLOPS in FP4 and 2 ZFLOPS in FP8.
- It will increase the performance and let users up their innovation game and improve the AI intelligence.
- Launch: 2027 IV quarter
How will it give competition to Nvidia?
The company has not clarified how the new system will compare to the Nvidia chips, but claimed the new SuperPoDs and SuperClusters will be the world’s most powerful supernodes that will deliver high computing power.
However, the Vice Chairman and rotating Chairman say their Atlas 950 Supernode will deliver 6.7 times higher computing power compared to Nvidia’s upcoming NVL144 system that will launch next year.
The Chinese market has been urging for self-reliance on Chip technologies for a very long time, and now Huawei has taken a step to build a new AI infrastructure with its interconnect technology.
Nvidia has acknowledged the competition and hopes customer will choose the best technology, as they are still confident about their open-source AI models.
Huawei is building the new AI infrastructure with SuperPods and SuperCLusters, which will be powered by the UnifiedBus to meet the computing power demand and sustain the AI development.
