The RAM shortage could last years

Original Article ↗ Hacker News Discussion ↗

Global demand for DRAM and HBM is surging as AI companies lock in massive long-term wafer and memory contracts, squeezing supply for consumer and general-purpose hardware and driving prices up. Commenters debate whether manufacturers’ reluctance to rapidly expand capacity reflects rational caution after decades of boom–bust cycles or cartel-like behavior that shifts costs onto consumers, and whether Chinese memory makers will eventually fill the gap. Many expect the crunch to last into the late 2020s unless the AI investment boom deflates, in which case overcapacity could suddenly flip into a glut and crash prices.

Drivers of the RAM crunch

  • AI datacenters, especially HBM-based accelerators, are soaking up capacity; major vendors are prioritizing HBM over commodity DRAM.
  • Large pre-purchase wafer deals by big AI companies suddenly tightened supply and triggered panic.
  • There’s concern that part of the demand is strategic hoarding to starve competitors, not just genuine usage.

Memory industry behavior & risks

  • Manufacturers are reluctant to expand aggressively due to decades of boom‑bust “pork cycles” where overbuilding led to price collapses and bankruptcies.
  • Some expect them to enjoy high margins now and accept slower growth to avoid another crash.
  • Others warn they could miscalculate and be “left holding the bag” if AI demand collapses or key customers default on massive orders.

Role of China and geopolitics

  • Chinese DRAM/NAND makers are ramping, but are estimated to lag leading firms by ~3 years in process nodes and yields; unlikely to fix shortages before ~2028–2029.
  • If incumbents underserve non‑AI markets, commenters expect Chinese memory to gain a foothold that might be hard to dislodge.
  • Broader geopolitical risks (e.g., Taiwan/TSMC, stressed power grids like in the Netherlands) are seen as amplifying fragility.

AI advances and the Jevons effect

  • Techniques like TurboQuant and other KV‑cache quantization schemes, plus new attention/SSM architectures, can cut memory use per token substantially.
  • Implementations so far often trade speed or quality, and are “good but not magic.”
  • Many argue savings will just be reinvested into longer contexts and more usage (Jevons paradox), not lower total demand.

Impact on consumers & hardware

  • Consumer RAM and GPUs have risen in price; some people find new prebuilts cheaper than self‑built systems with equivalent parts.
  • Older DDR3/DDR4 systems and second‑hand RAM are gaining value and being repurposed.
  • Some foresee eventual overshoot: fabs expand for AI, AI bubble pops, and consumers later enjoy ultra‑cheap, high‑capacity RAM.

Software efficiency debate

  • Some hope high prices will punish bloated software and reduce reliance on heavyweight stacks like Electron.
  • Others note optimization time is expensive, CPU–RAM tradeoffs are subtle, and most organizations lack incentives to rewrite for efficiency.

Economics, regulation, and uncertainty

  • One camp appeals to supply‑and‑demand: high prices will eventually attract capacity and then crash.
  • Another emphasizes oligopoly/cartel behavior, past price‑fixing, and weak antitrust.
  • Proposals include stricter, market‑share‑based regulation or tax schemes to discourage extreme concentration.
  • Overall, commenters see both a years‑long shortage scenario and an AI‑driven bust as plausible.