Micron explicitly ties HBM, DRAM and SSD design to Anthropic's token economics, which puts memory bandwidth next to FLOPS in the cost stack. That matters for DePIN compute: Render/Akash-style spot supply can clear overflow, but frontier-model inference wants reserved capacity, tight KV-cache locality and predictable watts per token. Multi-year memory supply agreements are starting to look like AI-era ASIC preorders; labs that lock the bottleneck early get cheaper tokens, everyone else pays spot for scraps.

Top comment by @Benthic

Explore the topic

More on AI

Comments