AI Compute Server Technology and Industry Chain Report

Overview

The compute scheduling layer can be likened to the "compute network brain" and is the core of the compute network. Beyond technical complexity, compute scheduling must balance multiple constraints such as cost and demand, making it a difficult multi-constraint optimization problem.

Technical Innovations

Key technical innovations in compute scheduling include cross-region compute dispatch, multi-level scheduling, unified allocation of compute resources, network protocols, visualization and monitoring, and intelligent operations and maintenance.

Demand Considerations

Different application scenarios have distinct compute and latency requirements. Scientific computing and critical national infrastructure projects need very high compute capacity, while ordinary commercial and consumer interactive scenarios typically emphasize low latency.

Cost Principles

The general principle of compute scheduling is to select the most economical compute network architecture and layout based on compute and latency requirements, in order to reduce compute costs, network costs, and operational costs.

Platform Construction

Development of compute scheduling platforms is active. Since the launch of China’s "East Data West Compute" initiative, regional organizations have carried out forward-looking research on compute measurement, compute interconnection, compute scheduling, and compute trading.

Compute platforms are often built by governments, state-owned enterprises, research institutions, data centers, and supercomputing centers, in cooperation with data operators, data service companies, digital transformation solution providers, compute service providers, equipment vendors, and large model providers.

The Compute Network

The goal of the compute network is on-demand availability of compute power. In 1961, US computer scientist John McCarthy suggested that compute power should be available like water and electricity on demand.

The compute network is divided into a resource layer, control layer, service layer, and orchestration and management layer. The international computing standard ITU-T Y.2501 defines the compute network functional architecture into these four modules: compute network resource layer, compute network control layer, compute network service layer, and compute network orchestration and management layer.

Orchestration and Scheduling Layer

The orchestration and management layer, i.e., scheduling, can be compared to the compute network brain. It is the core component responsible for intelligent orchestration and elastic scheduling of compute, network, and data resources across the entire network. Its main functions are:

1. Obtain real-time, global information on compute, network, and data resources, and the distribution across cloud, edge, and device, to construct a global situational awareness map.
2. Coordinate cross-domain scheduling by intelligently and automatically decomposing multi-domain collaborative scheduling tasks to enabling platforms, achieving resource scheduling for compute, network, and data.
3. Support multi-domain fusion orchestration by flexibly composing atomic capabilities of compute, network, and data to meet multi-domain fused service requirements.
4. Provide intelligent decision support by dynamically computing optimal collaborative strategies for compute, network, and data based on SLA requirements, overall network load, and the distribution of available compute resource pools.

Regional Imbalance and Economic Impact

Compute scheduling must consider demand and cost factors comprehensively. In China, an imbalance remains in the layout of new digital infrastructure across eastern, central, and western regions. The ratio of in-use racks between east and west is roughly 7:3. First-tier cities such as Beijing, Shanghai, Guangzhou, and Shenzhen face significant undersupply, with an average shortfall rate of 25%.

In the compute industry, every 1 yuan invested typically drives 3–4 yuan of economic output. Each percentage point increase in the compute power index can bring a 0.33% growth in the digital economy and a 0.18% growth in GDP.