Explosive demand for AI memory: shares of manufacturers could grow 10 times
The semiconductor memory market is on the verge of a tectonic shift. An analyst under the pseudonym Zeitgeist has presented a compelling argument: current demand for AI memory so far exceeds production capabilities that shares of leading manufacturers could skyrocket tenfold from today's levels.
The key thesis is that traditional valuation methods—based on historical highs—no longer work. It is necessary to look at the real need for computing power and, critically, memory. Zeitgeist provides a clear example: a $50,000 investment in Micron shares last September would have turned into nearly $489,000 today. This is not an anomaly but a signal of a new trend.
Why Memory Has Become a Bottleneck
Every modern accelerator, such as the H100, is equipped with a fixed amount of high-bandwidth memory (HBM) that cannot be expanded. The H100 has only 80 GB, new generations have up to 192 GB, and the future B300 will have 288 GB. This "ceiling" directly determines how many tasks a single chip can handle.
The main load comes not from model weights but from 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 query. This is why every additional gigabyte of memory becomes 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
The key driver is the shift from simple chats to AI agents. A simple query barely strains memory, but 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 about 250 million knowledge workers worldwide. If you multiply their number by the volume of simultaneous agent sessions, memory demand doesn't just grow—it "explodes." According to Zeitgeist's estimate, with a hundred agent sessions per person per day, the world would need 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 disproportionately faster: agents are replacing simple chats, context windows are expanding from 128,000 to 10 million tokens, and AI usage per worker is going from zero to hundreds of sessions.
My analysis: 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. A tenfold increase is not a fantasy but a matter of time and production scaling. Investors should closely watch Micron, SK Hynix, and Samsung—they will be the beneficiaries of this "memory rush."