The Memory Tax

Here's something that won't make headlines but matters more than most things that do: we're running out of RAM.

Not in the "oh no my browser tabs" sense. In the "global manufacturing capacity is being reallocated and you're not the priority" sense.

The Reallocation

Samsung, SK Hynix, and Micron — the three companies that make basically all the world's memory — have a choice to make every quarter: what do we build?

Regular DDR5 for servers and PCs? Or HBM for AI accelerators?

HBM sells for 5-10x the price per bit. The machines that make it are the same machines that could be making your RAM. Every wafer that becomes HBM is a wafer that doesn't become DDR5.

Nvidia's H100 and H200 need HBM. Lots of it. And Nvidia is shipping as fast as they can make them, which means the fabs are cranking out HBM as fast as they can make it.

The Squeeze

This creates a tax on everyone else.

Building a new database server? RAM prices are up 20-30% from last year. Planning a PC upgrade? DDR5 kits aren't getting cheaper like they usually do. Running a hosting company? Your hardware costs just went up and your customers don't care why.

The memory shortage of 2020-2023 was about supply chains breaking. This one is about supply chains choosing. The capacity exists. It's just pointed somewhere else.

The Play

If you're planning hardware purchases, front-load them. Prices aren't coming down until HBM demand saturates — and that's not happening in 2026.

If you're in the hosting business, you already know this. The smart operators have been pre-buying for six months.

Everyone else is about to learn what a memory tax feels like.