How Alibaba Cloud’s AI Infra Predictable Network Is Redefining Large-Model Training

In September 2025, the Alibaba Cloud Yi Conference opened in Hangzhou with the theme “Cloud‑Intelligence Integration, Carbon‑Silicon Co‑existence”, featuring the AI Infra Predictable Network forum.

The forum highlighted full‑stack AI infra optimization from architecture evolution, ecosystem partner innovation, to joint model‑framework optimization.

Alibaba Cloud’s network R&D leader Cai Dezhi presented future plans, unveiling the next‑generation HPN8.0 architecture supporting up to 100,000 cards and the UPN512 optical interconnect Scale‑up architecture, guiding the direction of Scale‑out and Scale‑up designs.

Senior experts Shi Yongqing and Fu Binzhang detailed the design considerations behind HPN8.0 and UPN512, and shared innovations in high‑performance network protocols, communication libraries, and ecosystem collaborations.

Speakers from Broadcom and NVIDIA discussed the roadmap of leading network chips and aggressive AI product plans, underscoring Ethernet’s thriving ecosystem.

Alibaba Cloud specialists Zhang Pengcheng, Li Qingxu, Huang Weixiao, and community member Liang Jiexin presented optimization practices at the model‑framework and communication‑library layers, exploring trends in AI framework optimization.

The closing round‑table, chaired by Su Yuanchao, brought together five guests to discuss key challenges and breakthroughs for AI infra networks.

The accompanying “AI Infra Predictable Network” exhibition showcased products ranging from data‑center to WAN solutions, including 800 Gbps and 1.6 Tbps optical modules, and demonstrated a full upgrade from DCN to DCI to meet exponential traffic growth.

Alibaba Cloud chose NPO technology, launching a 102.4 Tbps device built from four 25.6 Tbps domestic chips.

Today, AI Infra Predictable Network serves as a critical backbone for large‑model training and inference, acting as a high‑speed conduit and a “neural hub” that ensures stability and optimal efficiency through iterative architecture, hardware‑software co‑design, ecosystem integration, and distributed inference.

These advances position China at the forefront of AI infrastructure technology, turning the network into an active engine driving the intelligent era.

Reference: UPN512 Technical Architecture Whitepaper.