Huawei Unveils Atlas 850E: The New Air-Cooled AI Supernode Revolutionizing Large Language Models

Tech giant Huawei has officially introduced its groundbreaking Atlas 850E air-cooled AI supernode at the prestigious World Artificial Intelligence Conference (WAIC) 2026. Designed to address the critical challenges of thermal management in high-performance computing, this innovative hardware breakthrough focuses on optimizing heat dissipation while significantly boosting the execution of next-generation large language models (LLMs).

• ✨ Revolutionary Cooling: Utilizes self-developed VCE phase-change heat-dissipation technology to eliminate the need for complex liquid-cooling upgrades.
• ✨ High-Speed Memory: Features an impressive 4.0TB/s high-speed memory bandwidth to handle intensive data operations.
• ✨ Ultra-Low Latency: Delivers rapid inference performance at the 10ms level, perfect for real-time agentic services.
• ✨ Seamless Scalability: Expands a single air-cooled supernode to support up to 96 high-performance AI accelerator cards.

Unveiling the Powerhouse: Atlas 950 and Atlas 850E at WAIC 2026

The WAIC 2026 (World Artificial Intelligence Conference), running from July 17 to July 20, serves as a global stage for showcasing breakthrough AI strategies and computing infrastructure. At this landmark event, Huawei turned heads by unveiling two dynamic AI supernode products: the liquid-cooled Atlas 950 and the air-cooled Atlas 850E. While the Atlas 950 boasts up to 1024 Ascend 950 NPU chips (scalable to 8192 NPUs for massive performance), the Atlas 850E stands out as a highly practical, easily deployable air-cooled solution.

The Atlas 850E integrates Huawei's proprietary VCE phase-change heat-dissipation technology and the advanced Lingqu interconnect protocol. The primary advantage of this system is that it requires absolutely zero modifications to existing data center liquid cooling infrastructure. Enterprises can deploy it directly into standard air-cooled racks, saving enormous capital expenditure and setup time.

Enterprise Benefits: Speed, Scale, and Efficiency

By expanding a single air-cooled supernode to support a commercial cluster of 96 AI accelerator cards, the Atlas 850E delivers exceptional performance characterized by low latency, high throughput, and the ability to process long token sequences. This makes it a highly viable choice for modern enterprises running large-scale Agentic inference services.

Furthermore, the system supports a comprehensive range of low-precision calculation formats, including FP8, mxFP8, HiF8, INT8, and mxFP4. This ensures versatility across various model architectures and optimization pipelines. Key hardware highlights include:

• ✨ 4.0TB/s High-Speed Memory Bandwidth: Ensures rapid data transfer, minimizing bottlenecks during heavy workloads.
• ✨ 10ms-Level Low-Latency Inference: Perfectly matches the rapid response demands of multi-turn dialogues and real-time AI agents.
• ✨ Optimized KV Cache Read/Write: Seamlessly handles high-frequency memory access patterns common in agentic workflows.

Huawei’s new air-cooled AI supernode enables efficient parallel inference across multiple chipsets and memory resources. It also supports ultra-long context processing up to the Million (M) token level, allowing AI models to comprehend massive documents and complex datasets in a single pass.

As Huawei continues to expand its Ascend ecosystem, its hardware solutions now power over 60 real-world commercial operations. The company has set ambitious goals for the future of its AI hardware division, planning further expansions that will continue to challenge global computing standards.

What makes the Huawei Atlas 850E different from traditional AI supernodes?

The Atlas 850E utilizes Huawei's self-developed VCE phase-change heat-dissipation technology, allowing it to run as a high-performance air-cooled supernode. This eliminates the need for expensive liquid-cooling infrastructure modifications, making deployment much faster and cheaper for enterprises.

What kind of performance and memory bandwidth does the Atlas 850E offer?

It offers an impressive 4.0TB/s high-speed memory bandwidth and supports ultra-low latency inference at the 10ms level, which is ideal for real-time agentic services and multi-turn interactive dialogues.

How many AI accelerator cards can the Atlas 850E scale up to?

The Atlas 850E can expand a single air-cooled supernode into a commercial deployment of 96 AI accelerator cards, offering high throughput and handling ultra-long context processing up to the Million (M) token level.

Which data precision formats are supported by this new supernode?

It supports a full range of low-precision formats designed to optimize modern AI workloads, including FP8, mxFP8, HiF8, INT8, and mxFP4.

🔎 In conclusion, the debut of the Huawei Atlas 850E air-cooled AI supernode at WAIC 2026 marks a significant milestone in AI hardware accessibility. By combining cutting-edge phase-change cooling technology with top-tier performance metrics, Huawei has successfully bridged the gap between high-power AI computing and standard data center environments. As the demand for large language models and real-time AI agents continues to surge, solutions like the Atlas 850E will play a crucial role in democratizing powerful, scalable, and cost-effective AI operations worldwide.