Huawei Unveils New Ascend AI and Kunpeng Server Chips

▼ Summary
– Huawei has revealed its long-term chip strategy to launch powerful Ascend AI chips and Kunpeng server processors, aiming to reduce reliance on foreign suppliers like Nvidia.
– The Ascend 950 will be released next year in two variants, followed by the Ascend 960 in 2027 and Ascend 970 in 2028, with a goal to double compute annually.
– Huawei plans to introduce high-speed computing supernodes, including the Atlas 950 in Q4 2026 supporting 8,192 chips and the Atlas 960 in Q4 2027 supporting 15,488 chips.
– The Kunpeng 950 server processor will launch in Q1 2026 in two variants: one with 96 cores/192 threads and another with 192 cores/384 threads, featuring enhanced security.
– This announcement breaks years of secrecy around Huawei’s chip operations and raises stakes in the U.S.-China tech rivalry, following U.S. sanctions in 2019.
Huawei has taken the wraps off its ambitious long-term semiconductor strategy, signaling a major push to develop some of the world’s most advanced computing systems. This move underscores China’s determined effort to lessen its dependence on foreign chip suppliers like Nvidia. For the first time, the company has broken its longstanding silence on chip development, publicly outlining release timelines for its Ascend AI and Kunpeng server processors. This announcement is likely to intensify the ongoing technological competition between the U.S. and China.
Eric Xu, Huawei’s rotating chairman, shared the roadmap during Huawei Connect 2025, emphasizing a commitment to annual product cycles and doubling computational power with each new iteration. He stated that the company will continue refining its Ascend chips to improve usability, support more data formats, and boost bandwidth, all to keep pace with the exploding global demand for artificial intelligence processing.
Huawei first entered the chipmaking arena back in 2018 but went quiet after facing U.S. sanctions the following year. Accused of posing national security risks, the company has since become a central player in China’s quest to build a self-reliant semiconductor industry. Earlier this year, it released the Ascend 910C AI chip. Looking ahead, the Ascend 950 is set to debut next year in two versions, with the Ascend 960 following in 2027 and the Ascend 970 in 2028.
In addition to individual processors, Huawei is developing high-speed computing supernodes, rack systems that cluster large numbers of chips. The Atlas 950, scheduled for Q4 2026, will support up to 8,192 Ascend chips. Xu expressed confidence that this system will outperform rival products across all major performance indicators. Another supernode, the Atlas 960, is planned for late 2027 and will accommodate an impressive 15,488 chips. These build on the existing Atlas 900, also known as CloudMatrix 384, which integrates 384 of the current 910C chips.
On the server front, Huawei’s Kunpeng processors are also evolving, with a focus on supporting SuperPoDs, higher core counts, and improved performance through the company’s proprietary dual-threaded LinxiCore architecture. The Kunpeng 950 is slated for launch in Q1 2026 and will come in two configurations: one with 96 cores and 192 threads, and a more powerful variant with 192 cores and 384 threads. This chip will enable general-purpose SuperPoDs and introduce four-layer security isolation for confidential computing, a first for Kunpeng’s data center offerings.
Looking further ahead, Huawei has plans to continue advancing Kunpeng’s microarchitecture and packaging, with two new models expected in early 2028.
(Source: Economy Middle East)





