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

Original Article ↗ Hacker News Discussion ↗

DRAM and Storage Price Spike

  • Many commenters report 3–6x price jumps for DDR4/DDR5 RAM and SSDs within 1–2 years, sometimes making used RAM worth more than the machines it came in.
  • Even old DDR3 has risen sharply as people downgrade to older platforms as a “substitute good.”
  • Some see this as normal DRAM boom–bust cycling, but at an unusually extreme level.

Causes and Supply Constraints

  • Consensus: core bottleneck is fab capacity, not raw silicon. DRAM fabs are highly specialized and slow/expensive to expand.
  • Memory makers have shifted capacity toward high‑margin HBM and AI/datacenter products, starving consumer DDR and LPDDR.
  • Several comments cite the historic cyclicality of DRAM: overbuild → crash → bankruptcies. This history makes firms cautious about aggressive expansion now.
  • Some claim large AI labs have pre‑purchased huge shares of Samsung/SK Hynix capacity, triggering a “RAM starvation crisis,” though details are debated and somewhat unclear.

Market Structure, Collusion, and Geopolitics

  • Only three major DRAM vendors dominate; past price‑fixing scandals make posters suspicious of tacit or explicit collusion.
  • Others argue that fear of the next bust, not cartel behavior, explains slow capacity growth.
  • China’s CXMT is emerging as a DRAM supplier; some expect it to quickly take over low‑end consumer RAM, especially if Western firms abandon that segment.
  • Export controls on EUV/DUV tools are seen as limiting China’s ability to fully relieve shortages.

Impact on PCs, Gaming, and Devices

  • Home‑built PCs and gaming builds are becoming prohibitively expensive; some fear long‑term damage to the consumer PC ecosystem and component vendors.
  • Used servers and old platforms are being snapped up as relatively cheap ways to get lots of RAM.
  • Rising RAM costs also hit phones and cheap devices, with concern about pricing out poorer consumers.

AI Demand, Efficiency, and Future Scenarios

  • Debate over whether algorithmic advances and smaller models will reduce RAM demand, or just be reinvested into larger contexts and models (Jevons paradox).
  • Some expect an AI/datacenter overbuild and eventual crash leading to cheap used hardware; others think sustained 20–25% annual DRAM capacity growth still won’t be enough in the near term.
  • Cloud gaming/“PC as a subscription” is seen by some as the likely consumer future; others push back on latency, economics, and loss of ownership.