HBM Is Now 63% of AI Chip Component Costs — Up From 52% Just 18 Months Ago
High-bandwidth memory has become the dominant cost in building AI chips — and the gap is widening. Epoch AI’s latest analysis of AI chip component spending across Nvidia, AMD, Google, and Amazon finds that HBM’s share of total component costs rose from 52% in Q1 2024 to 63% in Q4 2025.
In absolute terms the shift is larger still. Total AI chip component spend grew from roughly $22 billion in 2024 to $52 billion in 2025. HBM alone accounted for about $20 billion of that $30 billion increase — going from ~$12B to ~$32B annually. Every other component shrank as a share: advanced packaging (CoWoS) fell from 19% to 15%, auxiliary components from 15% to 9%, and logic dies held roughly flat near 13–14%.
What Drove the Shift
The analysis covers all AI chips produced by the four largest designers, weighted by production volume. Component costs are built from financial disclosures, supplier filings, and analyst reports.
Two things drove the HBM surge:
Volume: More AI chips were built. HBM is baked into every unit of every H100, B100, Instinct MI300, TPUv5, and Trainium 2.
Price: HBM supply stayed tight through 2025. Micron sold out its 2026 HBM allocation before the year started. SK Hynix and Samsung command pricing power that does not exist in commodity DRAM.
Epoch expects the trend to continue in 2026 as memory supply remains constrained and demand accelerates with Blackwell ramp-up and new inference infrastructure.
Capex Guidance Is Already Absorbing the Number
The HBM squeeze is already showing up in hyperscaler guidance. Microsoft’s $190 billion FY2026 capex outlook explicitly includes about $25 billion attributed to higher component prices. Meta raised its 2026 capex range by $10 billion, citing the same cause. Neither company specified memory by name, but HBM price increases are the primary component-level driver at this scale.
Why It Matters
The logic die is the part of an AI chip that does the compute. HBM is the part that moves data to and from that compute. That HBM now costs 4.8x more than logic per chip — and growing — reflects a structural reality: AI workloads are memory-bandwidth-bound more than they are compute-bound.
Architectures that reduce memory access (sparse attention, MoE with fewer active experts, quantisation) lower cost more directly than raw FLOP improvements. That explains why techniques like DeepSeek’s multi-head latent attention, which slashes KV cache by 90% at 1M tokens, land as genuine competitive moats rather than incremental engineering. They attack the expensive part of the bill of materials directly.
Key Numbers
| Component | Q1 2024 | Q4 2025 |
|---|---|---|
| Memory (HBM) | 52% | 63% |
| Packaging (CoWoS) | 19% | 15% |
| Auxiliary | 15% | 10% |
| Logic | 14% | 13% |
| Metric | 2024 | 2025 |
|---|---|---|
| Total component spend | ~$22B | ~$52B |
| HBM spend | ~$12B | ~$32B |