2025-2027年人工智慧运算与资料中心领域的颠覆性趋势

分散式人工智慧基础设施、新型经营模式和ESG倡议推动未来成长潜力

颠覆性人工智慧模式正日益挑战人们对「传统」云端资料中心的固有认知。随着训练丛集规模扩大到数千个加速器，耗电量高达数十兆瓦，传统的以CPU为中心、采用风冷的单体伺服器模式已永续。本研究探讨了人工智慧基础设施分拆这一大趋势，即从传统的伺服器机箱转向由加速器、记忆体、储存、冷却和电力等元件组成的互联丛集。我们分析了运算密度、记忆体和I/O瓶颈的日益严峻，以及碳排放和水资源限制如何迫使营运商重新思考资料中心的基本设计，包括从传统伺服器转向人工智慧专用丛集和可组合丛集架构。未来三到五年，这些变革预计将在决定人工智慧如何在满足更严格的永续性和监管标准的同时，实现经济高效且可靠的扩展方面发挥关键作用。

本报告涵盖技术发展及其对应用和经营模式的影响，包括以下模组：

• 人工智慧基础设施解耦—大趋势概述
• 技术概述、架构和分类方案
• 人工智慧基础设施的转型主题
• 五大主题：
  • 异质与专用人工智慧加速器
  • 记忆体分散式高频宽架构
  • 冷却和电源作为核心设计变量
  • AI原生编配与自主资料中心
  • 永续性、位置和电力系统集成
• 技术进步应用案例
• 新兴经营模式与实施模式
• 人工智慧资料中心区域发展趋势
• 战略机会与未来展望

目录

调查范围

• 分析范围

变革性成长：人工智慧运算与资料中心

• 为什么经济成长变得越来越困难？
• The Strategic Imperative 8-TM
• 我们的大趋势宇宙概览
• 我们的大趋势领域—人工智慧运算与资料中心
• 主要发现

生态系统：人工智慧运算与资料中心

• 解耦人工智慧基础架构：从伺服器到 Fabric 连线池
• 结构性限制要求重新设计人工智慧基础设施
• AI Pods 作为一种新型设计单元：架构与战略意义

技术趋势与创新驱动因素

• 人工智慧资料中心和人工智慧运算栈的演进
• 人工智慧资料中心架构分类：运算与资源的结合
• 以人工智慧为中心的资料中心架构分类：部署、电源和冷却

人工智慧基础设施的转型主题：主要趋势和次要趋势

• 主题一：异质与专用人工智慧加速器
• 主题二：记忆体分布与高频宽架构
• 主题三：以散热和电源为核心设计变数的方法
• 主题四：人工智慧原生编配与自主资料中心
• 主题五：永续性、位置与电网集成

企业行动 (C2A)

• 案例研究 1：基于 CXL 的 LLM 训练单元记忆体分配
• 案例研究 2：针对人工智慧/高效能运算 (HPC) 和资料库工作负载的运算储存磁碟机和资料编配
• 案例研究3：低碳人工智慧区域及区域供热
• 案例研究 4：能源效率和气候限制下的主权百亿亿百万兆级人工智慧

生态系统：由人工智慧运算和资料中心驱动的新型经营模式

• 一种新的经营模式：加速器舱即服务
• 新兴经营模式：冷气即服务
• 新兴经营模式：遥测和数据驱动的货币化

生态系统：人工智慧运算与资料中心的区域趋势

• 人工智慧运算和资料中心的区域发展趋势

成长来源：趋势吸引力分析

• 趋势吸引力分析

成长机会分析

• 趋势、机会、影响分析与确定性分析
• 趋势机会革新指数
• 趋势分析吸引力评分
• 趋势成长指数
• 成长吸引力评分
• BEETS如何影响人工智慧运算与资料中心

成长机会领域

• 成长机会 1：人工智慧单元和组合式基础架构园区
• 成长机会 2：面向人工智慧资料中心的液冷与热回收平台
• 成长机会 3：碳感知编配与遥测平台

成长机会分析：成长的关键成功因素

• 成长的关键成功因素
• 结论

附录

• 我们的大趋势宇宙

下一步

• 成长机会带来的益处和影响
• 下一步
• 图表清单
• 免责声明
• 人工智慧基础设施解耦—大趋势概述

技术概述、架构和分类方案

人工智慧基础设施的变革性主题

• 五个详细主题：
• 异质与专用人工智慧加速器
• 记忆体分散式高频宽架构
• 冷却和功率是核心设计变量
• AI原生编配与自主资料中心
• 永续性、位置和电力系统集成

技术进步应用案例

新兴经营模式和发展模式

人工智慧资料中心区域发展趋势

战略机会与未来展望

AI Infrastructure Unbundling, Emerging Business Models, and ESG Commitments Driving Future Growth Potential

Disruptive AI models are increasingly challenging traditional assumptions about "classic" cloud data centers. As training clusters scale up to thousands of accelerators and consume tens of megawatts of power, the conventional CPU-centric, air-cooled, monolithic server model is no longer sustainable. This research study explores the megatrend of AI infrastructure unbundling, which signifies a shift from traditional server boxes to fabric-connected pools of accelerators, memory, storage, cooling, and power. The study analyzes how growing compute density, memory, and I/O bottlenecks, and constraints related to carbon and water are compelling operators to rethink the fundamental design of data centers. This includes moving from traditional servers to AI pods and composable cluster fabrics. Over the next 3 to 5 years, these disruptions are expected to play a critical role in determining how effectively AI can be scaled in an economical and reliable manner while adhering to stricter sustainability and regulatory standards.

The report, covering technological developments and their impact on deployment and business models, includes the following modules:

• AI Infrastructure Unbundling - Megatrend Overview
• Technology Overview, Architecture, and Taxonomy
• Transformational Themes in AI Infrastructure
• Five Deep-Dive Themes:
  • Heterogeneous and specialized AI accelerators
  • Memory disaggregation and high-bandwidth fabrics
  • Cooling and power as core design variables
  • AI-native orchestration and autonomic data centers
  • Sustainability, siting, and grid integration
• Technological Advancement Use Cases
• Emerging Business and Deployment Models
• Regional Trends in AI-Centric Data Centers
• Strategic Opportunities and Future Outlook

Table of Contents

Research Scope

• Scope of Analysis

Transformational Growth AI Computing and Data Centers

• Why is it Increasingly Difficult to Grow?
• The Strategic Imperative 8-TM
• Our Megatrend Universe-Overview
• Our Megatrend Universe-AI Computing and Data Centers
• Key Findings

Ecosystem: AI Computing and Data Centers

• AI Infrastructure Unbundling: From Servers to Fabric-Connected Pools
• Structural Constraints Forcing AI Infrastructure Redesign
• AI Pods as the New Unit of Design: Architecture and Strategic Impact

Technology Landscape & Innovation Drivers

• AI Data Center Evolution and the AI Compute Stack
• Taxonomy of AI-Centric Data Center Architectures: Compute and Resource Coupling
• Taxonomy of AI-Centric Data Center Architectures: Deployment, Power, and Cooling

Transformational Themes in AI Infrastructure: Key Megatrends and Sub-Trends

• Theme 1: Heterogeneous and Specialized AI Accelerators
• Theme 2: Memory Disaggregation and High-Bandwidth Fabrics
• Theme 3: Cooling and Power as Core Design Variables
• Theme 4: AI-Native Orchestration and Autonomic Data Centers
• Theme 5: Sustainability, Siting, and Grid Integration

Companies to Action (C2A)

• Case Study 1: CXL-based Memory Disaggregation for LLM Training Pods
• Case Study 2: Computational Storage Drives and Data Orchestration for AI/High Performance Computing (HPC) and Database Workloads
• Case Study 3: Low-Carbon AI Region with District-Heating Heat Reuse
• Case Study 4: Sovereign Exascale AI Under Energy-Efficiency and Climate Constraints

Ecosystem: Emerging Business Models Driven by AI Computing & Data Centers

• Emerging Business Model: Accelerator Pods-as-a-Service
• Emerging Business Model: Cooling-as-a-Service
• Emerging Business Model: Telemetry and Data-Driven Monetization

Ecosystem: Regional Trends for AI Computing & Data Centers

• Regional Trends in AI Computing and Data Centers

Growth Generator: Trend Attractiveness Analysis

• Trend Attractiveness Analysis

Growth Opportunity Analysis

• Trend Opportunity Impact and Certainty Analysis
• Trend Opportunity Disruption Index
• Trend Disruption Attractiveness Score
• Trend Opportunity Growth Index
• Growth Attractiveness Score
• BEETS Implications for AI Computing and Data Centers

Growth Opportunity Universe

• Growth Opportunity 1: AI Pods & Composable Infrastructure Campuses
• Growth Opportunity 2: Liquid First Cooling & Heat Reuse Platforms for AI Data Centers
• Growth Opportunity 3: Carbon Aware Orchestration & Telemetry Platform

Growth Opportunity Analysis: Critical Success Factors for Growth

• Critical Success Factors for Growth
• Conclusion

Appendix

• Our Megatrend Universe

Next Steps

• Benefits and Impacts of Growth Opportunities
• Next Steps
• List of Exhibits
• Legal Disclaimer
• AI Infrastructure Unbundling - Megatrend Overview

Technology Overview, Architecture, and Taxonomy

Transformational Themes in AI Infrastructure

• Five Deep-Dive Themes:
• Heterogeneous and specialized AI accelerators
• Memory disaggregation and high-bandwidth fabrics
• Cooling and power as core design variables
• AI-native orchestration and autonomic data centers
• Sustainability, siting, and grid integration

Technological Advancement Use Cases

Emerging Business and Deployment Models

Regional Trends in AI-Centric Data Centers

Strategic Opportunities and Future Outlook