Beijing / Shanghai, 18 September 2025 — Huawei has publicly laid out its most ambitious chipmaking and computing power plans to date, targeting a challenge to Nvidia’s leadership in the AI hardware space. At its Huawei Connect conference in Shanghai, Vice Chairman Eric Xu announced several new developments under its “Ascend” line of AI chips, along with large supernode architectures named “Atlas,” which are intended to deliver performance on scale.
After introducing the Ascend 910C earlier in 2025, Huawei revealed that the Ascend 950 (with two variants) will debut in 2026, followed by the Ascend 960 in 2027 and the Ascend 970 in 2028. A major technical leap is also coming in memory: Huawei claims it has developed its own high-bandwidth memory technology, a key area where, historically, China has depended on U.S. and South Korean suppliers.
Parallel to chip development, Huawei plans to launch new “supernode” computing systems – notably Atlas 950 later this year, which can support up to 8,192 Ascend chips, and Atlas 960 in 2027 with capacity for up to 15,488 chips. These systems are designed for high interconnectivity, enabling much greater compute scale and performance. Huawei claims that on certain metrics these systems already outperform Nvidia’s GB200 NVL72, a benchmark system relying on 72 GB200 GPUs.
Huawei frames these moves as part of a broader push for self-reliance in AI computing infrastructure, especially amid global supply chain pressures and export restrictions. By developing both chip architecture and supporting platforms (memory, interconnect, systems), Huawei is looking to reduce dependency and compete more fully in both domestic and global AI infrastructure markets.
Implications & Challenges
- Competition heats up: Huawei’s announcements sharpen the competition in AI hardware, particularly for clients in China seeking alternatives to Nvidia due to regulatory restrictions or geopolitical risk.
- Technical hurdles: Demonstrating superiority on benchmarks is one thing; matching the software ecosystem, energy efficiency, developer support, reliability and scale are other challenges.
- Timing and rollout: With new chips coming in 2026-2028 and supernodes debuting later this year (Q4 2025 for Atlas 950), there is a multi-year cusp period where Huawei must deliver not just prototypes but reliable, deployable systems.
- Global resonance: If Huawei can successfully package chip, system, memory, and interconnect capabilities, it could influence purchasers in APAC, Europe or other regions seeking alternatives to U.S.-based supply chains. However, export restrictions and trust / security perceptions will still play roles.
Source: Reuters
