AI agents are creating explosive demand for memory: Micron and SK Hynix stocks could grow 10 times.
Demand for high-speed memory for artificial intelligence tasks so far exceeds current production capacity that the stocks of leading manufacturers — Micron and SK Hynix — could see a tenfold increase from current levels. This is the conclusion of an analyst known under the pseudonym Zeitgeist, who suggests evaluating the sector's prospects not through the lens of historical highs, but through the arithmetic of real computing needs.
As a clear example, the expert cites an investment of $50,000 in Micron shares last September, which would have turned into approximately $489,000 today. According to him, some investors fear they are already too late, while others worry they will become "the very liquidity through which large players exit their positions." However, Zeitgeist proposes a fundamentally different approach to valuation — through the mathematics of memory demand.
Memory as the Bottleneck of AI Infrastructure
Every accelerator, whether the H100 or newer generations, is equipped with a fixed amount of high-bandwidth memory (HBM) that cannot be expanded. A standard H100 chip carries only 80 GB, newer models up to 192 GB, and the future B300 — 288 GB. It is this "ceiling" that determines how many requests a single accelerator can handle.
The main load falls not on the model weights, but on the so-called KV-cache — session memory that grows with each generated word. According to Zeitgeist's calculations, one session with a context of 128,000 tokens requires about 20 GB of memory. Just four such sessions completely exhaust the resources of a single H100. For advanced models like Claude Opus 4.8 or GPT-5.5, the need is even higher — from 40 to 100 GB per single long request.
That is why, the analyst emphasizes, every additional gigabyte of memory is worth its weight in gold, and manufacturers like Micron and SK Hynix physically cannot keep up with increasing output.
The AI Agent Effect and the Demand Gap
The key shift, according to Zeitgeist, is related to the transition from simple chats to AI agents. While a regular question barely loads memory, an agent that independently accesses tools and accumulates context easily reaches 100,000 tokens or more. One knowledge worker running ten such agents in parallel requires about 152 GB of memory.
There are approximately 250 million knowledge workers in the world. If you multiply their number by the number of simultaneous agent sessions, memory demand doesn't just grow — it "explodes." According to the analyst's estimate, with one hundred agent sessions per person per day, the world would need roughly 60 times more memory than will be produced in 2026.
Zeitgeist acknowledges that algorithms will eventually reduce memory consumption — new "attention methods" could cut the load by four to eight times. But demand is growing incomparably faster: agents are replacing simple chats, context windows are expanding from 128,000 to 10 million tokens, and each worker's AI usage is going from zero to hundreds of sessions.
In a world where language models are "woven into every aspect of daily life," memory becomes a critical resource. The companies that produce it will see unprecedented revenue.
Expert Opinion: This analysis is not just a speculative forecast, but a mathematically grounded scenario. The memory industry stands on the brink of a structural deficit that could last for years. Investors should pay close attention to Micron and SK Hynix, but with volatility in mind: a tenfold increase will not be linear, and corrections are inevitable. However, the long-term trend is clear.