Memory Tax

A 32GB kit of DDR5 now costs over seven hundred dollars. Six months ago it was under two hundred.

High-bandwidth memory for AI is eating the production lines. Every bit of HBM manufactured costs three bits of conventional DRAM. Data centers take first pick. Consumers get the scraps. The PC building community is in chaos.

Something strange is happening. On Amazon Japan, the best-selling motherboard is a last-gen DDR4 board. In 2026. ASUS has ramped up DDR4 motherboard production. DDR5 is too expensive, so people are hunting DDR4 stock and secondhand. That DDR4 is disappearing from the market too.

Then Apple quietly removed the 512GB option from the Mac Studio. March 5th. The 256GB configuration survived, but the upgrade price went up by four hundred dollars. The DRAM shortage reached unified memory. I thought that was further out. It wasn't.

The irony is that unified memory fits today's AI workloads well. LLM inference is memory-bandwidth-bound. It reads model data from memory constantly. On a normal PC, the model has to fit in GPU VRAM, which is expensive and limited. The RTX 5090 has 32GB. If the model doesn't fit, it doesn't run.

Macs don't have that constraint. CPU and GPU share the same memory pool. The M5 Max has 614GB/s of memory bandwidth. Four times faster than a typical DDR5 system's 100–150GB/s. Load a 70B-parameter model into 128GB of unified memory and run inference. No separate GPU needed.

So now, people who want to run local LLMs are buying up Macs loaded with memory. A Mac costing several thousand dollars is "cheap for an AI machine." For the price of a single NVIDIA A100, you can buy a few of them.

Come to think of it, the IDC rack I picked up recently came with a pile of DDR4 ECC Registered modules. Worth a small fortune right now.

Shameful for a sysadmin, but I'm fighting the temptation to flip disaster-recovery spares for profit. Every day.