Unlike traditional DRAM $Roundhill Memory ETF(DRAM.US), where DDR3 remained in use for about 15 years before DDR5 arrived, HBM upgrades much faster — roughly every two years per generation, with the pace even accelerating recently. HBM size and bandwidth continue to roughly double with each new generation, which matches the expected pattern.

Nvidia’s GPU bandwidth has been rising exponentially across generations: from 2 TB/s to 3.5 TB/s, then 4.8 TB/s, 8 TB/s, and now reaching 22 TB/s. Because inference token throughput scales linearly with HBM speed, the older generations quickly become uneconomical to use. Companies have a strong incentive to adopt the latest HBM, even if it costs more, because the gains in token throughput are significantly larger.

In the current “token factory” era, greater technical upgrades in HBM bandwidth directly translate into higher profits. This fast pace of improvement creates a situation similar to CPUs: older products lose value very quickly, making it less attractive to hold large inventories of previous-generation HBM. For example, HBM3 has already depreciated sharply and is rarely used in mainstream products today.

As a result, HBM manufacturers have rationally shifted their strategy. Instead of competing mainly on current production volume and fighting for market share through quantity, they now focus on technological leadership — improving stability, increasing speed, and securing qualification wins for the next generation on Nvidia platforms. This quality-focused competition helps them avoid the classic “prisoner’s dilemma” of traditional memory cycles, where companies hesitate to cut production during downturns for fear of losing share.

@Bridge Buzz SG