Carmack Suggests Using NAND Flash Over HBM for AI Inference

John Carmack and analysts suggest using cheaper NAND flash instead of expensive HBM to store model weights. Since AI inference relies on deterministic memory access, this shift could drastically reduce hardware costs.

2026-07-07 ~ 2026-07-08 · 2 related posts

• Carmack：用 NAND 闪存替代 HBM 降低推理成本 — ID_AA_Carmack · 2026-07-07
• Replacing HBM with Cheap Flash for Model Weights — BenBajarin · 2026-07-08