AI data centers are driving a global memory chip shortage in 2026

In short: Instead of using less memory, AI data centers are on track to use so much that they could cause a major global shortage of memory chips in 2026.

What's going on

Recent research challenges the idea that AI data centers will need less memory than investors expected. It points the other way, saying AI systems are using far more memory than standard computer servers.

Memory chips are the parts that hold working data while a computer is running, like a desk where you keep papers you are actively using (not long-term storage). AI servers often use 10 to 20 times more memory than traditional servers because they rely on GPUs and other AI chips to train and run AI models.

A big pressure point is High-Bandwidth Memory, or HBM, which is a very fast type of memory placed close to a GPU so it can be fed data quickly. Some top-end GPUs can require up to 1TB of HBM each. On top of that, AI servers also need large amounts of DDR5 and conventional DRAM for the main processors that support these systems.

Supply is tight because three companies, Samsung, SK Hynix, and Micron, make about 90% of the world’s memory chips. The research says they have shifted production away from older DRAM toward higher-profit HBM and enterprise DDR5 for AI, which can reduce availability for other devices. It also notes that new chip factories take about 3 to 5 years to build, so output cannot ramp up quickly.

Projections cited in the research suggest AI data centers could use about 70% of all DRAM production in 2026. That could mean higher prices and longer wait times for servers, laptops, PCs, and consumer electronics.

What to watch

If large buyers like Microsoft, Google, Meta, Amazon, and OpenAI keep ordering and stockpiling memory, shortages could worsen through 2026, with relief not expected until around 2028.

Source: Financial Times