For decades, DRAM shortages followed a predictable script: overshoot, crash, recover, repeat. Demand spikes would hit, manufacturers would add capacity, and the market would flood with chips again. This time is different.
At first glance, the global DRAM shortage may look like a typical semiconductor cycle, but it is actually one of the first pain points that we are seeing as AI collides with physical resource limits. Consumers, enterprises, and governments are already on the rollercoaster.
More important than the demand and supply imbalance, there are other factors like strategic capacity control, long term AI demand (that is just ramping up), and geopolitical constraints all converging at once. And it’s the first time AI is materially crowding out everyone else.
When DRAM tightens, PCs, laptops, phones, and tablets get more expensive. Automotive electronics systems face supply risk and I think we all remember what that looked like post-COVID. DRAM access becomes more unequal because priority (and manufacturing nodes) go to hyperscalers, leaving everyone else to fight over the remaining capacity. This means that consumer prices on all electronics will rise because the companies that make them are paying 2x, 3x, 4x more for the memory chips that allow them to work.
Samsung, SK Hynix, and Micron control ~95% of the DRAM market and they optimize profitability to focus on high margin node migration. A disproportionate number of manufacturing lines go to the newest technology that is used for AI, so the manufacturing capacity that is left is not enough to supply everyone else who needs DRAM for their products (which is basically everyone making any technology hardware product). So prices rise dramatically as companies fight over the available chips. What is different about this cycle of shortages is that it’s not just the lagging technology nodes that are not enough to supply demand, the companies buying up DRAM for expand their AI capabilities can’t get nearly enough to support their build plans. This was highlighted by the recent firing of a top tech executive who failed to sign multi-year contracts with memory vendors to lock in supply. These chips will affect the winners and losers in the AI race because without them, all progress stops.
Another interesting aspect, which is typical of every cycle, is that bringing up a new node technology has low early yield (for every 100 chips manufactured maybe only 75 of them work). As the technology matures, these numbers improve but that takes months to years to get them where the manufacturing capacity reaches an optimal level. So the more the technology transitions, the lower the overall yield of the factory. And the other contributing factor is the cost of a new FAB. Building additional manufacturing capacity takes 7-10 years and around $50B. These are strategically planned and advertised for years before they become reality, so there are no fast response options to demand spikes. And the final consideration here is the EUV technology required to manufacture the advanced nodes that is prohibitively expensive, in extremely high demand, and also controlled by government policy and export controls limiting other potential suppliers from entering the market. Here, Micron has the key home court advantage.
The demand curve is also changing, likely for the long-term, not just a short term bump. GPU’s requiring high bandwidth memory (HBM), DDR5 RDIMM density required for training clusters combined with long qualification cycles locks supply to only a handful of customers. HBM is particularly destabilizing because it requires leading edge DRAM wafers, advanced packaging capacity and in-demand manufacturing tools (like EUV).
In conditions like these, memory suppliers stop selling memory and they start allocating supply. This is also not new and not specific to this AI-driven spike, but what is different is the allocation hitting the largest customers so hard that they are firing executives over it. This is a new level of scarcity.
Memory has always been the market leading commodity indicator for the semiconductor industry. I am not totally convinced that this is changing but I do think there is a significant and fundamental shift that has already started because competing for memory is the first battle ground in the AI fight for resources. The next will likely be water and energy but the global implications of the fight for memory may give us an estimate of scale of what is to come.
