On October 26 the China Computer Federation (CCF) announced its 2026 science and technology awards, giving the first prize to "key technologies and applications of high-performance networks for large-scale intelligent computing centers", a joint project by Aliyun and Tsinghua University, per Sina Tech and Xinhua, with an "internationally leading" appraisal.

__FIG__

1. The network is the hidden bottleneck

At the scale of ten-thousand-GPU clusters, GPUs themselves are often not the scarcest resource: one network hiccup can idle thousands of cards during cross-node training, and finding a single faulty cable among hundreds of thousands of devices is like finding a needle in a haystack. Bandwidth supply, traffic scheduling and fault localization are the three problems this project targets. CCF's appraisal committee credited creative results in Ethernet-based intelligent-compute networking, naming three internationally leading innovations: a flat two-tier architecture, end-network cooperative path selection, and root-cause localization via multi-dimensional trace correlation.

2. From three tiers to two flat planes

Concretely, Aliyun's HPN compresses the traditional three-tier, single-uplink data center network into a "dual-plane, high-density multi-rail" design: every server connects into multiple planes at once, so any link failure fails over in seconds, with flatter bandwidth, shorter paths and smaller blast radius. Building on Ethernet rather than proprietary interconnects means the results live in an open ecosystem that smaller players can reuse.

3. From award to deployment

What makes the project notable is that it ships: the technology is already in large-scale commercial deployment through Aliyun, supporting LLM training, inference and industry AI applications. The industry is learning to price network engineering properly, in the first prize it just received. Effective compute equals chips times network efficiency, and efficiency gains eventually flow through as lower training and inference costs. NineZenith's industry model and agent work rides the same infrastructure-cost curve: the more efficient the fabric, the lower the barrier for vertical AI adoption.

(Compiled from public reports by Sina Tech and Xinhua)