Kioxia: Not a Flash in the Pan

Historically, the dominant narrative in AI infrastructure investment has been GPU-centric — more compute unlocks better models with higher benchmark scores. However, the focus has now shifted from training scaling laws to agent scaling laws, where the relevant input is not training compute.

Rather, agent performance on complex multi-step tasks scales with the amount of context and memory available to each agent, the number of parallel agents that can collaborate or verify each other’s work, and the number of reasoning steps the agent can take before returning a result.

Agent scaling is bottlenecked by the infrastructure surrounding the GPU — you need the agent’s working memory to be accessible fast enough, cheaply enough, and at large enough scale to support long sessions. Infrastructure layers that were historically secondary — NAND storage, DRAM, CPU orchestration, and high-bandwidth interconnects — become critical determinants of system performance and cost.

In our GTC preview, we observed that “NAND should not be considered a secondary topic in GTC discussions” because it is becoming a “core architectural variable”. In an even earlier post, we noted that the investment logic for DRAM and SSD increasingly resembles “AI infrastructure growth” rather than merely “cyclical beneficiaries of price upcycles.”

In this note on Kioxia, we expand on these observations.

The demand for Agentic AI inference infrastructure will continue to support Kioxia’s negotiating position.