A researcher writing under the pseudonym Zeitgeist said that demand for memory for artificial intelligence exceeds what the world is currently able to produce by orders of magnitude. In his view, shares of memory manufacturers could rise tenfold from current levels if they are valued not by historical highs, but by the real need for compute.
Zeitgeist gave a striking example: a $50,000 investment in Micron shares in September of last year would be worth around $489,000 today. According to him, some investors fear they have already missed the move, while others fear becoming “the liquidity that major players use to exit their positions.” The analyst suggests looking at the opportunity differently, through the arithmetic of memory demand.
Why Memory Has Become The Bottleneck
Each accelerator is equipped with a fixed amount of high-speed memory, known as HBM, which cannot be expanded. According to the analyst, a standard H100 chip carries only 80 GB, newer generations offer up to 192 GB, and the future B300 will have 288 GB. That ceiling determines how many requests a single accelerator can handle.
The main load does not come from the model weights themselves, but from the so-called KV cache: the session memory that grows with every generated word. According to Zeitgeist’s calculations, one session with a 128,000-token context requires around 20 GB of memory. Just four such sessions would fully exhaust the resources of a single H100.
Memory usage per session depending on the context window size, and the number of sessions per H100 accelerator. Source: Zeitgeist
For advanced models such as Claude Opus 4.8 or GPT-5.5, the requirement is even higher: from 40 GB to 100 GB for a single long request. According to the analyst, this is why every additional gigabyte of memory is worth its weight in gold, while manufacturers such as Micron and SK Hynix physically cannot scale production fast enough.
The AI Agent Effect And The Demand Gap
According to Zeitgeist, the key shift is the move from simple chatbots to AI agents. While a normal question places almost no burden on memory, an agent that independently calls tools and accumulates context can easily reach 100,000 tokens or more. A single knowledge worker running ten such agents in parallel would require around 152 GB of memory.
Peak memory usage per knowledge worker during parallel chats and agentic sessions. Source: Zeitgeist
The analyst noted that there are around 250 million knowledge workers worldwide. If that number is multiplied by the number of simultaneous agentic sessions, memory demand does not merely grow, it “explodes.” By his estimate, with 100 agentic sessions per person per day, the world would need roughly 60 times more memory than will be produced in 2026.
Zeitgeist acknowledged that algorithms will reduce memory usage over time, with new “attention methods” capable of cutting the load by four to eight times. But in his view, demand is growing far faster: agents are replacing simple chats, context windows are expanding from 128,000 to 10 million tokens, and AI usage by each worker is moving from zero toward hundreds of sessions.
According to the analyst, in a world where language models are “woven into every aspect of everyday life,” memory becomes a critical resource. Under his forecast, the companies that produce it will generate unprecedented revenue.
SK Hynix Overtakes Samsung As AI Memory Demand Surges
The scale of the AI memory boom is already reshaping the chip market. South Korean memory manufacturer SK Hynix has overtaken Samsung as the country’s most valuable listed company, driven by its strong position in high-bandwidth memory chips used for artificial intelligence workloads.
For Zeitgeist, this shift supports the broader thesis that memory producers are becoming some of the biggest beneficiaries of the AI infrastructure race. As demand for HBM grows faster than supply, companies capable of producing advanced memory chips could see revenue and valuations rise sharply.
