No RAM for Sam

Sam Altman recently called it “quite terrible” that personal computers hold so much RAM while using it perhaps five percent of the time, arguing that memory would be better pooled in data centers where it stays busy. The prolbem is that Altman himself helped cause the shortage. In October 2025, OpenAI signed letters of intent with Samsung and SK Hynix projecting demand for 900,000 DRAM wafers a month, up to 40% of global capacity. That spooked other buyers into locking in supply, and conventional DRAM contract prices roughly tripled in six months. Consumers are paying for it through much pricier PC memory, and new production is years away. SRAM is built for speed, DRAM for capacity and HBM for bandwidth. HBM, the stacked memory that powers AI GPUs, uses about three times the wafer area of conventional DRAM for the same capacity but sells for about five times the price per bit, so manufacturers prefer to make it. Even Apple is affected despite not using HBM, because its conventional low-power DRAM competes for the same wafers, and its prices have risen as a result. However, Apple’s unified memory lets the CPU and GPU share one pool, making the Mac well suited to running large AI models on the device.