
KV cache offload from HBM to DRAM/SSD boosts MFU and cuts $/token, making memory stocks AI-cycle leaders.
Before reading this article, you may want to review our earlier reports on memory and storage.
GOOG’s Optical Edge: OCS, Gemini 3, Model Progress, and the Coming Storage Crunch
Mid-Training Revolution: Why Every AI Lab is Making 3x Bets
SNDK 3Q25: Further Margin Upside Expected
COHR at inflection point, shift in TPU competition landscape, Enterprise AI demand, MU
At CES 2026, NVIDIA CEO Jensen Huang captured the data center’s central tension in a single line, shifting the bottleneck from “insufficient compute” to “insufficient context”: “Context is the new bottleneck — storage must be rearchitected.” As models grow larger, context windows extend, multi-turn interactions increase, and concurrent sessions proliferate, the key determinant of system throughput and cost is no longer just FLOPS. It is how “context data” is retained, moved, and reused.
This report is available to subscribers. Sign in or subscribe to read the full analysis.