Memory is becoming the "new oil" of AI: analyst predicts tenfold growth in manufacturer stocks
The demand for high-speed memory for artificial intelligence is not just growing—it is literally exploding, creating a colossal imbalance between market needs and production capacity. Based on my estimates, derived from a deep analysis of market data, the stocks of leading memory manufacturers such as Micron and SK Hynix could increase tenfold from current levels. This is not a speculative forecast, but the cold arithmetic of demand.
Let's break down the numbers. Each accelerator, whether it's the H100 or newer generations, is equipped with a fixed amount of high-speed HBM memory that cannot be expanded. The standard H100 has only 80 GB, newer models have up to 192 GB, and the future B300 will have 288 GB. This "ceiling" directly determines how many requests a single chip can handle.
The key load comes not from the model's weights themselves, but from the so-called 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.
AI Agents: A New Catalyst for Demand
The real turning point is related to the transition from simple chats to AI agents. While a regular question barely strains memory, an agent that independently accesses tools and accumulates context can easily reach 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 quantity of simultaneous agent sessions, the demand for memory doesn't just grow—it explodes. According to my calculations, with one hundred agent sessions per person per day, the world would need roughly 60 times more memory than will be produced in 2026.
Yes, algorithms will reduce memory consumption over time—new "attention methods" can 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 each worker's AI usage is going from zero to hundreds of sessions.
My expert assessment: In a world where language models are woven into every aspect of daily life, memory becomes a critical resource, and the companies that produce it will see unprecedented revenue. Investors who are now afraid that they are too late or that major players are exiting these securities are looking in the wrong direction. The real perspective lies through the arithmetic of demand, not through historical highs. Every additional gigabyte of memory is worth its weight in gold, and manufacturers physically cannot ramp up production fast enough.