While the recent and unprecedented increase in memory prices has been hard to miss, it is still largely misunderstood. Yes, it is driven by the speed and scale of the AI infrastructure buildout, crossed with how slow it is to scale up memory manufacture. But there is more to it than that.
First, however, a graph (data courtesy of Epoch):
So, some less discussed drivers:
- HBM suppliers are only slowly adding capacity, wary of prior boom-bust cycles.
- The bottleneck is High Bandwidth Memory (HBM) specifically, and it is controlled by three vendors (SK Hynix, Samsung, Micron). That's an oligopoly layered on top of cyclicality.
- This is leading to shortages, long-term purchase agreements at peak prices, price floor deals, over-ordering, and more.
2. Long-term purchase agreements (LTAs) are making things worse.
- LTAs are reshaping the market. Hyperscalers locking in multi-year supply at peak prices, making the market tighter. This drives up prices for everyone without an LTA.
- This is a feedback loop that will get worse and then crack catastrophically, leaving hyperscalers sitting on mispriced LTAs and surplus inventory simultaneously.