Huawei has unveiled a new line of AI infrastructure that aims to boost computing power and position the company as a formidable challenger to Nvidia’s dominance in the AI chip market.
Quick Summary – TLDR:
- Huawei introduced its new Atlas SuperPoDs and SuperClusters at HUAWEI CONNECT 2025 in Shanghai, calling them the most powerful systems in the world.
- The launch comes just as China restricts domestic firms from buying Nvidia hardware, intensifying demand for local alternatives.
- The Atlas 960 SuperPoD can connect up to 15,488 Ascend NPUs, and the Atlas 960 SuperCluster exceeds one million NPUs.
- A new interconnect protocol called UnifiedBus 2.0 was also introduced to overcome hardware connectivity limits.
What Happened?
Huawei held its annual HUAWEI CONNECT 2025 conference in Shanghai, where the company rolled out a suite of AI infrastructure products, including its new Atlas SuperPoDs and SuperClusters. The announcement came shortly after China enforced a ban on Nvidia hardware, further fueling Huawei’s urgency to fill the gap with homegrown solutions.
Huawei’s Answer to AI Compute Demand
Huawei’s Deputy Chairman Eric Xu took the stage to reveal what the company claims are the world’s most powerful SuperPoDs and SuperClusters. He emphasized that computing power is now the critical driver behind AI progress, especially in China where international supply chains face increasing scrutiny.
Huawei’s newly announced infrastructure includes:
- Atlas 950 SuperPoD with 8,192 Ascend NPUs
- Atlas 960 SuperPoD with 15,488 Ascend NPUs
- Atlas 950 SuperCluster with over 500,000 Ascend NPUs
- Atlas 960 SuperCluster surpassing one million Ascend NPUs
These systems are built using semiconductor process nodes available domestically in China, making them a sustainable alternative amid ongoing trade restrictions.
SuperPoDs and SuperClusters: How They Work
A SuperPoD is essentially a logical machine made up of multiple physical units that work together like a single brain. By clustering AI chips such as Huawei’s Ascend NPUs, these setups allow for higher computing efficiency and larger-scale AI model training.
“With the Atlas 950 SuperPoD & Atlas 960 SuperPoD, we are confident in our ability to provide abundant computing power for rapid AI advancements, both today & tomorrow,” said Eric Xu, Huawei’s Deputy Chairman of the Board & Rotating Chairman at HUAWEI CONNECT 2025 in Shanghai. pic.twitter.com/FRJAkbPjeD
— Huawei (@Huawei) September 19, 2025
According to Xu, these SuperPoDs outperform any competing infrastructure in key metrics like:
- Total number of NPUs
- Aggregate computing power
- Memory bandwidth
- Interconnect performance
SuperClusters scale this model even further, aggregating multiple SuperPoDs to reach levels of performance previously unseen in public infrastructure. This makes Huawei’s new products particularly attractive for hyperscale data centers handling vast AI workloads.
Tackling the Interconnect Challenge
One of the biggest hurdles in large-scale AI computing is how to reliably connect thousands of chips without losing speed or increasing latency. Huawei has responded with a new protocol called UnifiedBus, which ensures consistent, high-speed communication between chips and clusters.
At the event, Huawei also introduced UnifiedBus 2.0 and shared its technical specifications publicly. The goal is to foster an open ecosystem where industry partners can build compatible products, expanding the reach of Huawei’s interconnect standard.
A Push into General-Purpose Computing
Huawei is not stopping at AI-specific systems. It also unveiled the TaiShan 950 SuperPoD, the world’s first general-purpose computing SuperPoD. Designed to run everything from enterprise databases to traditional workloads, it combines with Huawei’s distributed GaussDB to rival mainframes and Exadata-class systems.
This move positions Huawei not only as an AI infrastructure leader but also as a broader enterprise computing solution provider.
SQ Magazine’s Takeaway
I’ve been watching Huawei’s push into AI hardware for a while, but this is easily their boldest move yet. With China cutting off Nvidia imports, Huawei is stepping into the vacuum with infrastructure that doesn’t just fill the gap but tries to leap ahead. The sheer scale of these SuperClusters is mind-blowing, and the UnifiedBus tech feels like a direct response to a real bottleneck in today’s AI systems. If Huawei can deliver on the performance claims and get ecosystem partners on board, this could shift the balance of AI computing power globally. I’d say this is a serious play and not just for show.
