AI Is Creating 'Insatiable' Demand for Memory, Says Alger CEO on the Semiconductor Boom

In an interview with Bloomberg, Dan Chung, CEO of investment firm Alger, called demand for memory 'insatiable' and tied it directly to the explosive growth of artificial intelligence. Chung says this is redefining the entire tech investment landscape and forcing investors to rethink where the profits in the AI boom will come from.

Why AI needs so much memory

Modern large language models need petabytes of data for training. Every model parameter takes up a chunk of memory. A single training run across hundreds of thousands of GPUs requires coordination across petabytes of DRAM and fast high-bandwidth memory (HBM). To serve millions of users at once, data centers use tens of TB of VRAM.

The Transformer architecture drives linear growth in memory consumption with every doubling of model size, a scaling law that nobody is going to repeal. Data centers running inference 24/7 consume memory like a black hole, absorbing it in enormous quantities.

Who feeds the machine

Demand extends beyond GPUs such as the NVIDIA H100 to DRAM, HBM (High Bandwidth Memory), and long-term storage. Companies like Micron, Samsung, and SK Hynix are seeing record orders for memory chips, with contracts booked 2-3 years ahead.

Memory makers are experiencing their strongest demand growth in two decades. Fabs are running at full capacity as they try to keep up.

• Model training: petabytes of memory for every update
• Inference at scale: many copies of a model distributed across servers
• Vector databases: embeddings for RAG eat up memory exponentially
• Fine-tuning: custom models for every enterprise customer

The investment trend

Dan Chung stresses that the value in the AI stack lies not in the processing chips themselves but in the memory ecosystem around them. Investors who used to watch only NVIDIA and new AI startups are now paying closer attention to memory makers, controllers, and cooling solutions.

This is a long demand cycle. It is not a fad or speculation but an architectural necessity built into the very nature of large language models.

"AI cannot advance without memory.

It's not an option, it's a constraint," as the interview puts it.

What it means

Semiconductor memory is becoming the new 'railroad': the basic infrastructure on which every other technology is built. Investment is shifting away from the flashiest AI startups and toward the companies supplying the picks and shovels for the AI gold rush. That means long-term demand, protection from speculation, and revenue visibility for many years ahead.