# HBM demand — X 热门讨论 (2026-09-22 03:19 UTC)
## @LIWEI_TWCapital (LIWEI_TW Capital) · 09-21 12:01 · ♥36 ↻4 💬4 Intel: An AI Infrastructure All-in-One Play https://x.com/LIWEI_TWCapital/status/2102005309741130044
## @HyperTechInvest (Daniel Romero) · 09-21 16:34 · ♥34 ↻2 💬3 Agents Are A Blessing For NAND Prices
The market expected NAND prices to peak in Q3 2026
> Instead, NAND prices have taken a very different direction
The reason? Hyperscalers' enterprise SSD demand is nearly double earlier forecasts
Hyperscaler NAND inventories stood at about 14 weeks in July, below the normal level of 15 weeks, with inventories expected to decline further in Q4 2026
What's driving this demand?
1. Agents
AI agents perform repeated retrieval operations, access large databases, and generate substantial amounts of intermediate data
Unlike traditional chatbots, agents can execute multiple tasks autonomously, repeatedly accessing and processing information. This increases the amount of data that needs to be stored and retrieved
Cloud providers are consequently expanding their use of high-capacity QLC enterprise SSDs for vector databases, retrieval, and caching
QLC NAND allows them to store more data at a lower cost per bit, making it particularly attractive for these workloads
2. KV-cache offloading
AI models need to remember the context of a conversation or task to avoid processing the same information repeatedly. This information is stored in what is known as the KV cache
Traditionally, this data is kept in expensive HBM or DRAM. However, not all of it needs to remain in high-speed memory at all times
By moving less frequently accessed KV-cache data to SSDs, cloud providers can reduce memory costs while maintaining access to much larger amounts of context
TrendForce specifically identifies DeepSeek-related architectures in China as a driver of growing demand for high-capacity QLC enterprise SSDs for KV-cache offloading https://x.com/HyperTechInvest/status/2102073879531491584