Memory has quietly become the most expensive part of an AI chip. According to Hacker News, high-bandwidth memory (HBM) now accounts for 63% of component costs across chips from Nvidia, AMD, Google, and Amazon, up from 52% just two years ago. What stands out here is that memory, not the logic die everyone obsesses over, is the real bottleneck driving AI hardware economics.
The analysis covered Q1 2024 through Q4 2025 and broke per-chip costs into four buckets: memory (HBM), logic dies, advanced packaging (CoWoS), and auxiliary components. Researchers multiplied per-chip costs by estimated quarterly production volumes, then calculated each category’s share of total component spending per quarter. The shift is dramatic.
The numbers
Here’s how the component mix changed over the period:
- Memory (HBM): 52% → 63%
- Packaging (CoWoS): 19% → 15%
- Auxiliary components: 15% → 9%
- Logic dies: ~13-14% (roughly flat)
Total component spending on AI chips jumped from around $22 billion in 2024 to $52 billion in 2025. HBM alone accounted for roughly $20 billion of that $30 billion increase. Memory isn’t just a line item anymore. It’s the line item.
Why this matters
This reshapes how to think about the AI supply chain. The conversation has been dominated by TSMC’s logic node capacity and CoWoS packaging constraints, but the data points somewhere else: SK Hynix, Samsung, and Micron, the three HBM suppliers, are the real chokepoint. SK Hynix in particular has captured most of the Nvidia business, and that translates into pricing power across the entire AI accelerator stack.
For practitioners building or buying AI infrastructure, a few practical takeaways:
- Memory capacity is the scarce resource. When evaluating accelerators, HBM stack size and bandwidth matter more for cost than transistor count.
- Inference economics are memory-bound. Larger models and longer context windows hit memory walls before they hit compute walls. Expect HBM3E and HBM4 supply to set the pace of price drops, not new GPU generations.
- Vendor concentration is a risk. With three HBM suppliers and one dominant winner, supply shocks ripple straight into accelerator pricing.
- Custom silicon doesn’t escape it. Google’s TPUs and Amazon’s Trainium chips show the same memory-heavy cost structure. Designing your own chip doesn’t help if you still need HBM.
The limitations
The figures are estimates built from public production volumes and component pricing, not audited financials from the chipmakers. Margins, yield assumptions, and bulk discounts can shift the exact percentages. The direction of the trend, however, is consistent across every vendor analyzed.
The takeaway for anyone tracking AI infrastructure: watch the memory makers as closely as you watch Nvidia. The cost curve of frontier AI runs through Icheon and Pyeongtaek, not just Hsinchu. More details available at the original Hacker News post.