Memory has grown to nearly two-thirds of AI chip component costs

内存成本已占 AI 芯片组件成本的近三分之二

High-bandwidth memory (HBM) has grown from 52% to 63% of total AI chip component spending between Q1 2024 and Q4 2025. These estimates are an average across all AI chips designed by Nvidia, AMD, Google, and Amazon, weighted by production volume. 从 2024 年第一季度到 2025 年第四季度，高带宽内存（HBM）在 AI 芯片组件总支出中的占比从 52% 增长到了 63%。这些估算数据是基于英伟达、AMD、谷歌和亚马逊所设计的所有 AI 芯片的平均值，并按产量进行了加权。

As a share of spend, logic dies stayed roughly flat near 13%, while advanced packaging fell from 19% to 15% and auxiliary components fell from 15% to 9%. In absolute terms, HBM spend across these four designers grew from roughly $12 billion in 2024 to $32 billion in 2025, a faster year-over-year increase than any other component. 从支出份额来看，逻辑芯片（logic dies）基本保持在 13% 左右，而先进封装从 19% 下降至 15%，辅助组件从 15% 下降至 9%。从绝对值来看，这四家设计商在 HBM 上的支出从 2024 年的约 120 亿美元增长到 2025 年的 320 亿美元，其同比增幅超过了任何其他组件。

HBM will likely account for an even larger share in 2026 as memory supply remains tight and prices rise. Hyperscalers are already anticipating this in capex guidance: Microsoft’s $190 billion FY2026 capex outlook includes about $25 billion from higher component prices, while Meta raised its 2026 capex range by $10 billion, citing higher component prices. 随着内存供应持续紧张且价格上涨，HBM 在 2026 年的占比可能会进一步提升。超大规模数据中心运营商（Hyperscalers）已在资本支出指引中预见到了这一点：微软 2026 财年 1900 亿美元的资本支出预期中，约有 250 亿美元源于组件价格上涨；Meta 也以组件价格上涨为由，将其 2026 年的资本支出范围上调了 100 亿美元。

For each AI chip designed by Nvidia, AMD, Google, and Amazon, we estimate the per-chip cost of four component categories: memory (HBM), logic dies, advanced packaging (CoWoS), and auxiliary components. We then multiply those per-chip costs by estimated quarterly production volumes to get total component spending in each category, and compute each category’s share of total component spending per quarter from Q1 2024 to Q4 2025. 针对英伟达、AMD、谷歌和亚马逊设计的每一款 AI 芯片，我们估算了四类组件的单片成本：内存（HBM）、逻辑芯片、先进封装（CoWoS）和辅助组件。随后，我们将这些单片成本乘以每季度的预估产量，得出各类别组件的总支出，并计算出 2024 年第一季度至 2025 年第四季度期间，每季度各类别在总组件支出中的占比。

We find that memory’s share rose from 52% to 63% over this period, while packaging fell from 19% to 15% and auxiliary components from 15% to 9%. Logic die share stayed roughly constant near 13–14%. Total component spend on AI chips grew from approximately $22 billion in 2024 to $52 billion in 2025, with HBM spending alone accounting for roughly $20 billion of that increase. 我们发现，在此期间内存的占比从 52% 上升至 63%，而封装占比从 19% 降至 15%，辅助组件从 15% 降至 9%。逻辑芯片的占比基本保持在 13%–14% 左右。AI 芯片的组件总支出从 2024 年的约 220 亿美元增长到 2025 年的 520 亿美元，其中仅 HBM 的支出就贡献了约 200 亿美元的增长。

Data

数据

Component cost estimates are drawn from our AI Chip Components explorer, which builds chip-level bills of materials from financial disclosures, supplier filings, and analyst reports. Four component categories are tracked: 组件成本估算源自我们的“AI 芯片组件浏览器”（AI Chip Components explorer），该工具通过财务披露、供应商备案文件和分析师报告构建了芯片级的物料清单（BOM）。我们追踪了四类组件：

• Memory: HBM stacks (HBM3, HBM3e).
• 内存： HBM 堆栈（HBM3, HBM3e）。
• Logic: advanced-node logic dies (3 - 5nm).
• 逻辑： 先进节点逻辑芯片（3-5nm）。
• Packaging: TSMC CoWoS advanced packaging.
• 封装： 台积电 CoWoS 先进封装。
• Auxiliary: substrate, power delivery, and other non-logic, non-memory inputs.
• 辅助： 基板、供电系统以及其他非逻辑、非内存输入组件。

Analysis

分析

Each unit of a component has some cost uncertainty, including the price of an HBM stack, a logic die, or a CoWoS package. We model each chip’s per-component cost with a 90% confidence interval and the share is a ratio of this component’s cost over the total, so both numerator and denominator are uncertain. 每个组件单元都存在一定的成本不确定性，包括 HBM 堆栈、逻辑芯片或 CoWoS 封装的价格。我们使用 90% 的置信区间对每款芯片的各组件成本进行建模，由于占比是该组件成本与总成本的比率，因此分子和分母均存在不确定性。

• Range from this component’s cost: the share when this component’s cost lands at its 5th or 95th percentile and the other three components are held at their medians.
• 基于该组件成本的范围： 当该组件成本处于其 5% 或 95% 分位数，而其他三个组件保持其中位数时的占比。
• Range when all components vary at extremes: the share when this component sits at one extreme of its CI and every other component lands simultaneously at the opposite extreme of theirs.
• 所有组件处于极端变化时的范围： 当该组件处于其置信区间（CI）的一端，而所有其他组件同时处于其置信区间另一端时的占比。