AI Memory Shortage: Analyst Predicts Tenfold Growth in Micron and SK Hynix Stocks
The semiconductor memory market is on the verge of a historic rally. An analyst under the pseudonym Zeitgeist has presented compelling calculations demonstrating that current demand for artificial intelligence memory far exceeds production capacity. According to their estimate, shares of leading memory manufacturers — Micron, SK Hynix, and Samsung — could increase tenfold from current levels if valued not by historical highs, but by the real need for computing resources.
As a clear example, Zeitgeist cites an investment case: a $50,000 investment in Micron shares last September would now be worth approximately $489,000. However, according to them, many investors are gripped by fear: some worry they are already too late, others fear becoming "the liquidity through which large players exit their positions." The analyst proposes a fundamentally different approach — evaluating prospects through the arithmetic of memory demand.
Why Memory Has Become a Bottleneck
Each AI accelerator is equipped with a fixed amount of high-bandwidth memory (HBM) that cannot be expanded. The standard H100 chip carries only 80 GB, newer generations up to 192 GB, and the future B300 will have 288 GB. This "ceiling" directly determines how many requests a single accelerator can handle. The key load falls not on the model weights themselves, but on the KV-cache — session memory that grows with each generated word. 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 requirement is even higher — from 40 to 100 GB per single long request. That is why, Zeitgeist emphasizes, every additional gigabyte of memory is worth its weight in gold, and manufacturers like Micron and SK Hynix physically cannot ramp up production fast enough.
The AI Agent Effect and the Demand Gap
According to the analyst, the key shift 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 worldwide. Multiplying their number by the volume of simultaneous agent sessions, memory demand does not just grow — it "explodes."
According to Zeitgeist's calculations, with one hundred agent sessions per person per day, the world would require roughly 60 times more memory than will be produced in 2026. The analyst 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.
Expert commentary: This analysis highlights a fundamental shift in the structure of semiconductor demand. Unlike the cyclical DRAM and NAND markets, where growth was driven by consumer electronics, the current driver is infrastructure investments by hyperscalers, which are price-insensitive. Until HBM production catches up with demand, shares of Micron and SK Hynix will continue to be revalued with a scarcity premium. Investors should closely monitor these companies' quarterly reports — the gap between production capacity and orders will continue to widen.