Physix FrontierPhysix Frontier · News Briefing Card (Leiphone · Oct 8, 2026)
Huawei, Nvidia Push KV Cache Storage; SSD Specs Unsettled
KEY FACTS
- Huawei unveiled the OceanStor M900 at its 2026 Connect conference, targeting external KV Cache storage.
- Nvidia has already launched CMX Context Memory Storage, also aimed at AI inference caching.
- Raising KV Cache hit rates from 90% to 95% can cut recomputation compute overhead in half.
- Industry estimates put a unified SSD product consensus 6 to 12 months away, with development and validation taking about another year.
- Some North American customers have already increased SSD purchases because of KV Cache, while China remains at the solution-design stage.
KEY DATA
90%→95%Hit rate improvement
50%Recomputation compute overhead reduction
3 DWPD→9 DWPDSSD write endurance requirement change
6-12 monthsEstimated time to unified standard
PHYSIX OBSERVATION
Externalizing KV Cache is moving from an in-server optimization to an independent storage tier, and Huawei and Nvidia have already staked out positions. But dedicated SSD specs remain undefined, customers are still buying ordinary enterprise drives, and the market is in a hundred-flowers-blooming phase. For storage vendors, this is both an incremental opportunity and a standards battle; for users, the math of trading storage for compute must account clearly for capacity, endurance, and data movement costs, not just cheaper storage.
Source: Leiphone report