Groundbreaking SuperPoD Interconnect: Leading a New Paradigm for AI Infrastructure

At Huawei Connect 2025, Huawei revealed an aggressive AI infrastructure roadmap, introducing new Ascend AI chips (950, 960, 970 series through 2028) and powerful SuperPoDs and SuperClusters. Key announcements include the Atlas 950 SuperPoD (Q4 2026) claiming 8 EFLOPS (FP8) and superior performance over NVIDIA's future systems, and the Atlas 950 SuperCluster (Q4 2026) projected to surpass xAI's Colossus with 524 EFLOPS (FP8) from over 500,000 NPUs. This strategy, bolstered by the open-source UnifiedBus 2.0 interconnect and new general-purpose TaiShan 950 SuperPoD for sectors like finance, underscores Huawei's intent to lead the global AI compute market by leveraging proprietary hardware and ecosystem development despite semiconductor manufacturing limitations.

Analysis

Huawei has unveiled a highly aggressive multi-year roadmap aimed at establishing leadership in the AI and general-purpose computing infrastructure markets, directly challenging incumbents like NVIDIA and Oracle. The strategy pivots from competing on leading-edge semiconductor process nodes, where Huawei acknowledges a persistent disadvantage, to achieving superior performance through massive system-level integration. Central to this is the new open-source UnifiedBus 2.0 interconnect protocol, designed to link over 10,000 NPUs to function as a single computer. The company detailed a clear chip release cadence with its Ascend 950, 960, and 970 series, promising to double compute power annually through 2028. The upcoming Atlas 950 SuperPoD, slated for Q4 2026, is explicitly positioned to outperform NVIDIA's planned NVL144 system by delivering 6.7 times more computing power (8 EFLOPS in FP8). Furthermore, the Atlas 950 SuperCluster is projected to surpass xAI's Colossus with over 500,000 NPUs. In a significant move into the enterprise space, Huawei is also launching the TaiShan 950 SuperPoD, which leverages its Kunpeng processors to directly target the replacement of legacy mainframes and Oracle's Exadata systems in the lucrative finance sector. This hardware-focused monetization is supported by a strategic decision to open-source key software stacks like CANN and foundation models, aiming to accelerate adoption and build a competitive ecosystem.

Key Decisions for Investors

• Investors in NVIDIA (NVDA) should view Huawei's detailed roadmap as a significant long-term competitive threat, particularly in China and aligned markets; monitor the Q4 2026 launch and performance benchmarks of the Atlas 950 SuperPoD, as successful execution could challenge NVIDIA's market share and pricing power.
• The introduction of the TaiShan 950 SuperPoD presents a direct challenge to Oracle's (ORCL) high-margin Exadata business, and investors should assess the risk of market share erosion in the enterprise database hardware segment, especially within the financial services industry.
• The decision to open-source the UnifiedBus 2.0 interconnect protocol is a critical move to build a rival ecosystem to NVIDIA's NVLink and InfiniBand; watch for industry adoption announcements, as a successful coalition could fragment the AI hardware market and impact the entire data center supply chain.
• While the announcements are ambitious, they carry substantial execution risk, given Huawei's reliance on system integration to compensate for lagging chip manufacturing; investors should treat these forward-looking statements with caution and discount for potential delays or underperformance relative to the aggressive targets.