Tech war: Huawei’s computing cluster to power AI projects, challenge US platforms
News Summary
Huawei Technologies is leveraging its latest supernode computing cluster systems to bolster its artificial intelligence projects, aiming to challenge rival platforms from Nvidia and xAI. Huawei Deputy Chairman Eric Xu Zhijun stated at the Huawei Connect 2025 conference that the Atlas 950 and Atlas 960 SuperPoDs, along with the Atlas 950 and Atlas 960 SuperClusters, are designed to help the Shenzhen-based firm “circumvent the limitations in China’s chip manufacturing process.” The company plans to release the Atlas 950 SuperPoD, powered by 8,192 Ascend neural processing unit (NPU) cards, by the fourth quarter of next year. The Atlas 960 SuperPoD, featuring up to 15,488 Ascend 960 NPU cards, is slated for availability by the end of 2027. Larger computing clusters, the Atlas 950 SuperCluster and Atlas 960 SuperCluster, will eventually integrate over 500,000 and 1 million Ascend NPUs respectively, with the Atlas 950 SuperCluster available by the end of 2026.
Background
Since 2019, the U.S. government, particularly under President Donald J. Trump's administration, has imposed stringent trade restrictions and technology export controls on Huawei, aiming to curtail its access to advanced semiconductor technology. These sanctions have significantly hampered Huawei's ability to design and produce high-end chips, especially in areas like artificial intelligence and high-performance computing. In response, Huawei has been heavily investing in domestic research and development, leveraging its Ascend series NPUs to develop localized alternatives, striving for technological self-sufficiency and reducing reliance on foreign supply chains.
In-Depth AI Insights
Can Huawei's computing clusters genuinely challenge US platforms, or is this primarily a strategy for domestic self-sufficiency? - Huawei's Atlas SuperPoD and SuperCluster series, integrating hundreds of thousands or even over a million Ascend NPUs, indicate a move to build a massive and independent AI computing infrastructure. This isn't solely about domestic self-sufficiency; it's also about leveraging scale and vertical integration within a constrained environment to minimize the performance gap with US AI computing leaders (like Nvidia) in specific application scenarios. However, the breadth, ecosystem maturity, and cutting-edge performance required to broadly challenge US platforms remain limited by advanced process technology constraints in the short term. What are the deeper investment implications for China's AI ecosystem and technological autonomy? - Huawei's actions will accelerate the build-out and standardization of China's domestic AI computing infrastructure, offering local AI companies an alternative to US technology stacks, thereby further reducing reliance on US tech. This could foster an independent "China AI Computing Alliance," attracting more local software, algorithm, and application developers to innovate on the Ascend platform, forming a unique domestic ecosystem. For investors, this implies more investment opportunities within China's AI sector centered on indigenous hardware, particularly in industry-specific applications and vertical solutions. What are the primary risks and challenges Huawei faces with this large-scale deployment of proprietary NPU clusters? - Pace of Technological Iteration: The US lead in AI chip architecture and manufacturing processes means faster technological advancement. Huawei will need to invest immense R&D resources to keep pace. Future US export controls could also restrict more related technologies. - Ecosystem Maturity: While Huawei is actively building the Ascend ecosystem, it still needs time to attract enough developers and applications compared to Nvidia's mature CUDA ecosystem, which may impact market penetration and real-world effectiveness. - Cost-Effectiveness: Under constrained manufacturing processes, the cost-effectiveness of mass-producing and deploying these NPU clusters, and their competitiveness in actual commercial applications, still needs to be proven in the market.