Huawei Unveils First Air-Cooled AI Supernode

Huawei presented an AI supernode that requires no liquid cooling at the World Artificial Intelligence Conference in Shanghai. The Atlas 850E is billed as the industry's first enterprise-grade system to combine supernode architecture with air cooling, allowing it to be deployed in an ordinary server room without a costly overhaul of cooling infrastructure.

What Huawei Showed in Shanghai

Huawei's booth featured two distinct products that the company presents as complementary solutions for different deployment scales. The Atlas 950 SuperPoD is the flagship, liquid-cooled supernode, designed to link up to 8,192 Ascend processors into a single logical computer via Huawei's proprietary UnifiedBus interconnect protocol, also known as Lingqu. Alongside it sat the Atlas 850E, an air-cooled version aimed at customers who lack large-scale liquid cooling infrastructure.

The publicly demonstrated Atlas 950 SuperPoD configuration linked 1,024 NPUs spread across 16 racks. According to figures provided by Huawei, this setup delivers 1 EFLOPS of performance at FP8 precision, 2 EFLOPS at FP4, 256 terabytes of globally addressable memory, and transmission latency of 3 microseconds.

Air Cooling as a Market Edge

According to Huawei, the Atlas 850E uses an improved VCE cooling technology designed to ensure stability, high throughput, and low latency for agentic inference workloads. The configuration scales from 8 to 1,024 Ascend processors, letting customers start with a small inference deployment and gradually expand it to cluster scale without changing the cooling architecture.

That marks a significant difference from most competing supernode-class systems, including Nvidia's, which typically require liquid cooling at comparable compute density. Air-cooled server rooms are cheaper to build and operate, and their numbers still far exceed facilities equipped for liquid cooling, especially outside the largest hyperscalers.

Competing With Nvidia

Huawei has spent months positioning its Atlas lineup as an alternative to Nvidia chips in markets where the American company faces restricted access due to export controls. At earlier presentations of the Atlas 950 supernode, Huawei representatives claimed it delivers 6.7 times the computing power of Nvidia's NVL144 system, until now regarded as one of the most powerful GPU solutions for AI.

The company previously announced that new generations of its Ascend processor line will launch annually, with computing power doubling every cycle. The full version of the Atlas 950 SuperPoD is expected to go on sale in the fourth quarter of 2026.

Implications for the Data Center Market

For data center operators outside China, the Shanghai presentation signals that pressure to cut AI infrastructure costs no longer stops at the processors themselves but now extends to how they are cooled. If the Atlas 850E genuinely holds performance close to liquid-cooled systems, it could push suppliers in Asian and African markets, where Chinese chips remain available without US export restrictions, toward cheaper AI deployments that skip investment in new cooling installations.

For Polish companies and integrators, the direct impact is limited for now, since European data centers mostly rely on Nvidia and AMD chips, and exporting Chinese Ascend processors to the EU is not standard commercial practice. Still, the event points to the direction competition over energy and cooling costs is heading in the global race for AI infrastructure, which indirectly affects component and energy prices in Europe as well.