When people buy a new phone or laptop, they are often told to get more memory. More memory helps a device run faster, handle more tasks, and store more data. It costs a little more, but most people understand why it matters.
The same idea applies to artificial intelligence, only at a much larger scale. AI requires far more memory than consumer devices and different kinds of memory altogether. Demand is rising quickly, and the world is struggling to produce enough advanced chips to keep up.
Every time AI answers a question, huge amounts of data must be retrieved, processed, and delivered back to the user in seconds. Software gets most of the attention, but the underlying hardware makes it possible. One of the most important parts of that hardware is the memory chip.
A simple way to think about it is the human brain. Some information stays front and center while we work on a task. Other information gets stored away for later. Memory chips work in a similar way.
DRAM is the short-term memory. NAND is the long-term memory. Every smartphone, laptop, server, and data center relies on both.
A traditional Google search uses very little memory. An AI answer uses much more. Agentic AI agents could push that demand even higher.
Today, most AI use is conversational. A person asks a question, gets an answer, and moves on. Agents are different. They can work for longer periods, remember more context, manage multiple steps, and coordinate with other AI systems. That means more data needs to be stored, accessed, and moved around constantly. It is like moving from one employee handling a single task to a team working continuously in the background.
|
|
Today: Chat-Based AI |
Tomorrow: AI Agents |
|
Inputs |
One prompt at a time |
Many tasks running at once |
|
Checkpoints |
Human checks each step |
Human sets the objective |
|
Duration |
Short sessions |
Long-duration workflows |
|
Storage Needs |
Limited context |
Expanding memory needs |
Forecasted Growth in AI Token Consumption By 20301
24x
Making more memory chips is not simple. New factories cost tens of billions of dollars, take years to build, and require deep technical expertise. That is one reason the DRAM market is still dominated by three companies—Samsung, SK hynix, and Micron—whose leadership was established over decades.
At the same time, demand is being driven by the largest technology companies in the world. In 2026, hyperscaler capex is expected to reach roughly $680 billion,2 led by Microsoft, Google, Meta, and Amazon. These companies are not just experimenting with AI infrastructure. They are committing enormous capital to it, and memory is a critical input.
AI Capex Is Just Getting Started, Thanks to the Hyperscalers
This shortage may not resolve quickly. New supply takes years to come online, and the current leaders cannot simply flip a switch to meet AI-driven demand. Samsung has warned of shortages through at least 2027, while SK hynix has suggested tight conditions could extend toward 2030.3
The next time a phone or laptop comes with a recommendation to get more memory, it is worth remembering that the same small component now sits at the center of one of the most important supply crunches in technology.
We believe memory capacity and the ability to move data quickly between chips, servers, and data centers are two critical bottlenecks in the AI infrastructure buildout. For investors looking to access the long-term structural opportunity in memory, the Tema Memory ETF (DISK) offers high-conviction, actively managed exposure to leading semiconductor companies across HBM, DRAM, and NAND, with strong secular growth potential.