High-bandwidth memory (HBM) grew to represent 63% of total AI chip component spending by the fourth quarter of 2025, up from 52% in early 2024, according to research published by Epoch AI analyst Venkat Somala. Absolute spending on HBM across chips designed by Nvidia, AMD, Google, and Amazon expanded from approximately $12 billion in 2024 to $32 billion in 2025.
Shifting Breakdown of Component Costs
Epoch AI analyzed per-chip bills of materials alongside quarterly production volumes for custom and commercial AI accelerators. Total component spending across the four leading designers surged from $22 billion in 2024 to $52 billion in 2025, with HBM accounting for roughly $20 billion of that $30 billion annual increase.
As memory captured a larger share of the overall bill of materials, other component cost shares shifted between Q1 2024 and Q4 2025:
- Memory (HBM): Increased from 52% to 63% of component spending.
- Logic Dies: Remained essentially flat, shifting slightly from 14% to 13%.
- Advanced Packaging (CoWoS): Decreased from 19% to 15% of spending.
- Auxiliary Components: Fell from 15% to 10% of total component spend.
Limitations in Per-Chip Cost Modeling
Epoch AI notes that individual component costs carry inherent uncertainty due to private contract variations, supplier timing, and production volume estimates. Component prices were evaluated using a 90% confidence interval, meaning HBM's share of total component spend in Q4 2025 ranged between 60% and 67% under standard variations, or 54% to 73% when evaluating all components at statistical extremes.
Why it matters
As memory bandwidth requirements outpace logic scaling, HBM availability and price volatility have become significant cost drivers in AI infrastructure deployment. For enterprise chip buyers and cloud customers, elevated memory prices could keep hardware expenditures high even as advanced logic manufacturing nodes mature.