Overview
- OpenAI’s CEO Sam Altman publicly argued on Oct. 29 that running models on personal computers wastes scarce DRAM and urged routing inference through large cloud data centers to boost memory utilization.
- Reporting shows OpenAI has struck large supply arrangements that could let it buy as much as 900,000 HBM wafers per month, a volume that analysts say would equal roughly 40 percent of global output.
- Anthropic’s Oct. 26 IPO filing flagged a possible shortfall in high‑bandwidth memory from key suppliers, signaling that advanced AI labs now treat HBM as a material operational risk.
- Major memory makers say they are cautious about greenlighting new fabrication plants because past cycles of overbuilding produced deep losses, and building new HBM capacity typically takes years and massive capital.
- Analysts and policymakers warn that the memory squeeze is contributing to ‘chipflation’—pushing up prices for PCs, consoles, and autos—and have proposed tools such as priority allocation or faster domestic fab support to limit economic spillovers.