Server DRAM prices surge 50% as AI-induced memory shortage hits hyperscalers

Original Article ↗ Hacker News Discussion ↗

Scope of the DRAM Price Spike

  • Commenters report large increases across the board:
    • Desktop DDR5 nearly doubling in ~2 months; multiple anecdotes of 25–100% jumps vs late 2023 / early 2024.
    • DDR4 also rising as demand spills over; server RDIMM sticks that were ~$90 now seen at ~$430.
    • Even used ECC and desktop RAM on eBay has roughly doubled compared to year‑old posts.
  • Some say RAM had become “ridiculously cheap” pre‑spike; others strongly reject the idea that higher prices are “more reasonable.”

Regional Differences and Tracking

  • PCPartPicker trends are confirmed to be US‑centric; price rises there are clear.
  • UK and Japan users also report recent spikes using Amazon/camelcamelcamel and local price trackers.
  • Southern Europe data appears flatter to some; others insist prices are up ~40% across Europe, suggesting delays or low turnover in local channels.
  • PCPartPicker adds EUR‑grouped trends during the thread in response to these questions.

Causes: AI Demand, Hoarding, and Supply Constraints

  • Links cite:
    • OpenAI’s Stargate plans potentially consuming a large fraction of global DRAM output.
    • SK Hynix sold out of production for next year; Adata saying AI datacenters are “gobbling up” DRAM, SSDs, HDDs.
  • Hyperscalers reportedly hoard GPUs that can’t even be powered yet, indirectly hoarding attached RAM.
  • Some speculate on bulk buying and speculative reselling; others note that previous attempts to flip DDR4 weren’t highly profitable.

Manufacturer Strategy and Market Power

  • Several comments argue manufacturers learned from past oversupply crashes and now deliberately underproduce rather than risk low prices; collusion is hinted at but not proven.
  • Others counter that shortages are dangerous for vendors and that maximizing output to meet demand is still most profitable.
  • Another view: fear, inertia, and technical limits (e.g., HBM vs commodity DRAM, long fab lead times) explain the slow response more than conspiracy.

Impact on Consumers and Builders

  • Many regret “just missing” the cheap era when building PCs, NAS boxes, or high‑RAM workstations.
  • DDR4 systems (e.g., AM4) are touted as a relative safe harbor.
  • Some liken the situation to prior GPU booms where high‑end demand cascaded down and even “junk” parts became valuable.

AI Trajectory and Efficiency Debate

  • Some hope the DRAM crunch will force smaller, more efficient models (quantization, MoE, distillation).
  • Others respond that intense work on inference efficiency has been ongoing from day one, with many architectures and hardware startups already chasing lower costs.
  • One faction hopes the “AI craze” crashes to normalize prices; another argues AI demand will persist and is needed to fund advanced fabs.