At the 2026 Apsara Conference in Hangzhou, Alibaba’s semiconductor arm T-Head introduced the Zhenwu V900, a next-generation AI chip designed to accelerate both large-model training and inference. This launch marks a strategic expansion toward a full-stack AI chip architecture addressing computing, storage, and networking demands.
- Zhenwu V900 delivers triple the performance of its predecessor
- Supports large-scale AI workloads with up to 1,000 chips interconnected
- Roadmap includes enhanced Yitian server CPUs scheduled for 2027
What happened
During the 2026 Apsara Conference in Hangzhou, T-Head launched its Zhenwu V900 AI chip, targeting efficient AI model training and inference. The chip boasts significant upgrades including 216GB memory and a 1,200GB/s inter-chip bandwidth, supporting advanced precision formats like FP8 and FP4 to optimize computing efficiency and cost. The V900 aims to operate in large-scale AI systems, interconnected via T-Head’s proprietary ICN Switch technology that allows over a thousand chips to work cohesively.
In addition, T-Head presented its comprehensive AI hardware development roadmap, which includes new Yitian server CPUs focused on single-core performance and energy efficiency improvements, expected by Q3 2027. Alibaba also showcased a next-generation supernode server platform incorporating the V900 chip, networking NICs, and smart storage controllers, collectively enhancing AI computing, storage, and network integration.
Why it matters
T-Head’s unveiling of the Zhenwu V900 and the broader infrastructure elements underscores a pivotal shift from isolated AI accelerators to a full-stack solution that tightly integrates computing, storage, and networking. This holistic approach is critical as AI models grow exponentially in size, demanding cohesive system-level design rather than piecemeal chip upgrades. It positions Alibaba to more effectively support next-generation AI workloads, including trillion-parameter models and advanced AI agents.
Furthermore, the roadmap detailing future Yitian CPUs that natively interface with AI accelerators highlights Alibaba’s investment in optimizing server architectures for AI. This integrated strategy addresses inefficiencies in current AI infrastructure and strengthens Alibaba Cloud’s position as a premier provider of scalable AI computing resources in China and beyond.
What to watch next
Industry observers should monitor the market deployment of the Zhenwu V900 supernodes and their adoption across sectors such as autonomous driving, finance, and large language models. Alibaba’s ability to scale clusters up to half a million accelerators will be a key indicator of their infrastructure leadership and potential competitive advantage in AI services.
Attention should also focus on the launch and performance of the Yitian 720 and 730 CPUs in 2027, especially how they enhance integration with AI chips for more efficient processing. Developments in the ICN interconnect protocol and its adoption in the Yitian 750 CPU will be important for understanding Alibaba’s evolving chip ecosystem and its impact on AI server design.