Explosive demand for AI memory: Micron and SK Hynix stocks could soar 10 times
Demand for high-speed memory for artificial intelligence is so far outstripping supply that shares of leading manufacturers—Micron, SK Hynix, and Samsung—have the potential to grow 10 times from current levels. This conclusion was reached by an analyst under the pseudonym Zeitgeist, who proposed assessing the sector's prospects not by historical highs, but by the real need for computing resources.
As a clear example, he cites a $50,000 investment in Micron shares last September: today that amount would have turned into approximately $489,000. However, according to him, many investors now fear that they are either too late or, on the contrary, will become "that very liquidity through which large players exit their positions." Zeitgeist suggests looking at the situation through the arithmetic of memory demand, and this arithmetic looks staggering.
Why memory has become a bottleneck
Every accelerator—whether it's the NVIDIA H100, H200, or the future B300—is equipped with a fixed amount of high-bandwidth memory (HBM) that cannot be expanded. The standard H100 has only 80 GB, newer generations have up to 192 GB, and the B300 has 288 GB. This "ceiling" directly determines how many requests a single chip can handle.
The main load falls not on the model itself, but on the so-called KV-cache—session memory that grows with each generated word. According to 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 one H100. And for advanced models like Claude Opus 4.8 or GPT-5.5, the need for one long request reaches 40–100 GB.
"Every additional gigabyte of memory is worth its weight in gold, and manufacturers like Micron and SK Hynix physically cannot keep up with increasing production," the analyst states.
The effect of AI agents and the demand gap
The key driver of explosive growth is the transition from simple chats to AI agents. A simple question barely loads 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 approximately 250 million knowledge workers in the world. If you multiply their number by the number of simultaneous agent sessions, the demand for memory doesn't just grow—it "explodes." According to Zeitgeist's estimate, with a hundred agent sessions per person per day, the world would need about 60 times more memory than will be produced in 2026.
The analyst admits that algorithms will eventually reduce memory consumption—new "attention methods" can 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 AI usage per worker 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 are destined for unprecedented revenue.
My expert commentary: The market is clearly underestimating the scale of the upcoming HBM shortage. Even with technological optimizations, the demand curve, fueled by agentic AI, will outpace production capacity at least until 2028. Investors who are currently looking at "expensive" Micron shares may be overlooking the fact that fundamental demand is just starting to accelerate.