Ethan Mercer

The value of an X account can be measured by what doors it opens for you in real life.

"Memory is cyclical, everyone knows that, and the recent run up in memory names is an obvious bubble." That's the easy, reflexive view. But I think the people who hold it are missing the simple scale of what AI is doing to memory demand. The first clue that there might be more to the memory story came in January of this year when it came out that NVDA's next gen Rubin platform would require 16 TB of NAND per GPU, or 1152 TB per rack, and that required HBM bandwidth for the system would be 70% higher than what had been previously reported. That was the first time it became obvious to outside observers that memory would need to scale exponentially to keep up with already-known GPU demand. One under-appreciated fact is that while GPU compute has largely scaled with Moore's Law (doubling in compute ~every 2 years), memory density and speed hasn't. As GPU compute continues to scale, existing memory manufacturers must produce exponentially more chips. These chips will also need to be faster than ever, which introduces an incredible technical challenge: how can memory manufacturers find the required speed improvements that have eluded them for decades? When you combine this added technical complexity with an exponentially expanding demand for the product, memory starts to look less like the "commodity" everyone knows it to be, and much more like a high-margin proprietary chip. This hasn't even touched on memory's role in inference (compute needed for inference is expanding exponentially as well, and is highly memory-dependent), long context, etc. Agentic AI requires agents to pull massive amounts of data into their context, which increases the number of tokens per "turn" and also the amount of memory required to run them. True agentic systems will require both dramatically higher context, and also many more "turns" or iterations of each task (as they improve an output over and over until it reaches a target quality level). Longer context = more memory per workload, and more "turns" = more workload per output. To put a specific number on that, Micron SVP Jeremy Werner said recently on The Circuit that agentic AI is causing context length to grow 30x a year. Michael Dell recently framed the problem in extremely simple terms: H100 had 80GB of HBM; by 2028, accelerators could carry ~2TB. That is 25x more memory per accelerator. Over the same period, he expects roughly 25x more accelerators deployed. That's 25 x 25 = 625x more accelerator memory demand by 2028. Everyone knows memory stocks are cyclical, and they always look cheap right before the bubble bursts. But what if there are structural changes happening in the memory markets that could prove the consensus wrong? Does anyone remember another traditionally cyclical company that has rerated to a growth story due to the demand from AI? Hint: It's now the most valuable company in the world. Reminder: this is not a recommendation to buy or sell any securities. It's a framework for thinking about how the AI buildout may be changing the memory market.