随著生成式人工智慧在商业化应用加深，昔日被大型语言模型与专用晶片术语取代的旧式 IT 正在回潮：以中央处理器为核心的伺服器、储存与传统软体堆叠重新被列入 AI 支出版图。投资人目前关心的是，这些旧 IT 在 AI 主要运算工作中的角色会不会重新变得核心，以及其供应商能否在新价值链中拿到可观分成；若可证明其不可替代，旧 IT 仍有显著定价能力。

文章指出 AI 由「训练」转向更多查询推理（inference）是主因之一，CPU 在载入模型、管理输出中扮演关键角色。ARM 与 AMD 近期显示出强劲 CPU 订单，AMD 甚至预估未来数年 CPU 销售将有约35%复合成长，约为六个月前预估的一倍。若「AI agents」广泛采用，人机协作工作会扩大至与 GPU 相匹配的 CPU 需求，Intel CFO 表示训练阶段只需每8颗 GPU 配1颗 CPU，agent 工作则可能达到 CPU 与 GPU 1:1，这与 Intel 股价自上季美国政府入股以来约400%的弹升一致。受惠情绪亦扩及 Seagate，上月涨幅约60%。

更关键是价值如何分配。传统硬体可借由参与核心流程保住议价能力；若仅做背景服务则难以受惠。软体端最受考验，Salesforce 已成为首个面向 AI agents 以非浏览器（headless）模式提供软体的大型企业，借由长期沉淀的客户资料建立情境理解并编排多个 agents，但其商业模式仍高度依赖使用者人数，且 AI 供应商凭借与人类并行运行有机会快速累积历史经验，直接威胁其「资料护城河」。市场暂未完全认可这些模型，但随著 AI 软体供应链成形，旧 IT 角色的重新评估正在进行中。

The swing in the narrative is from AI’s chip-specialized boom back toward pre-existing “old IT” layers. The article says that while servers, storage, CPUs for general computing, and enterprise software were sidelined during the ChatGPT era, they are now being reconsidered as inference and AI-agent workloads become more central to business. The key uncertainty is what slice of AI demand these layers can capture: not just technical necessity, but whether legacy suppliers can claim meaningful value and preserve pricing power in the new stack.

Demand is shifting from model training toward inference, where CPUs are heavily used to load models and manage outputs. Recent disclosures from ARM and AMD point to a rebound in CPU demand, and AMD now expects compound growth of around 35% in CPU sales over the next few years—roughly double its outlook only six months earlier. Intel executives argue that training can run on one CPU per eight GPUs, while agent workloads could require CPUs in a 1:1 ratio with GPUs. Sentiment has been reflected in equity reactions: Intel has rallied roughly 400% since the U.S. government investment last summer, and Seagate has risen about 60% over the past month.

The main issue is value allocation. Legacy hardware can benefit if positioned as central, but not if treated as background infrastructure. In software, Salesforce is the first major firm to push a headless model for enterprise software not centered on human operators, betting on long-held customer data to orchestrate teams of AI agents. Yet its pricing model still depends on active human users, and AI vendors can potentially build historical operational context by running alongside workers. Wall Street has not fully granted the benefit of the doubt, and the reshaping of AI software supply chains is likely to force a reassessment of which “old” technologies become indispensable.