Memory Now 63% of AI Chip Cost: HBM Spend Hits $32B

Epoch AI's latest analysis reveals that high-bandwidth memory (HBM) now accounts for nearly two-thirds of AI chip component costs. Between Q1 2024 and Q4 2025, HBM's share rose from 52% to 63% across chips designed by Nvidia, AMD, Google, and Amazon, weighted by production volume. Total component spending on AI chips grew from $22 billion in 2024 to $52 billion in 2025, with HBM alone contributing $20 billion of that increase.

Key Numbers

• Memory (HBM): 52% → 63% share. Absolute spend grew from ~$12B (2024) to $32B (2025).
• Logic dies: Roughly flat at 13–14% share.
• Advanced packaging (CoWoS): Fell from 19% to 15% share.
• Auxiliary components: Fell from 15% to 9% share.

Why It Matters

Memory is now the dominant cost driver in AI accelerators. Epoch models per-chip costs from financial disclosures, supplier filings, and analyst reports across four categories: HBM stacks (HBM3, HBM3e), advanced-node logic dies (3–5nm), TSMC CoWoS packaging, and auxiliary components (substrate, power delivery). The shift is driven by tight memory supply and rising HBM prices. Hyperscalers are already incorporating this into capex guidance: Microsoft's $190B FY2026 capex outlook includes ~$25B from higher component prices; Meta raised its 2026 capex range by $10B citing the same factor.

Uncertainty Ranges

Epoch provides 90% confidence intervals and two bound metrics:

• Range from memory's cost alone: 60–67% in Q4 2025.
• Range with all components at extremes: 54–73% in Q4 2025.

The trend suggests memory's share will grow further in 2026 as HBM supply remains constrained.