UBS Predicts 1 ZB KV Cache by 2030, Cuts Rubin Standard VRAM to 8.3TB
zephyr_z9 · x · 2026-08-12
A UBS report projects that KV cache could approach 1 ZB by 2030. Assuming 30 million accelerators are sold that year, the attach rate would be nearly 33TB per device.
Additionally, Nvidia's Rubin architecture was initially introduced with 16 TB of VRAM per GPU, but the standard configuration has been halved to 8.3 TB.
More from Infra
- Nvidia RTX 50 Series GPUs See Massive Price Hikes Globally Amid VRAM Shortages — sujingshen · 2026-08-12
- From CUDA to Multi-GPU: A Curated GitHub Repo for GPU Engineering — tom_doerr · 2026-08-12
- Realistic Expectations for Running Qwen 3.6 27B on a Single 3090: Speeds, Quants, and Context Lengths — oldschooldaw · 2026-08-12
- OpenAI Exec Highlights AI Creating High-Quality, Unionized Blue-Collar Jobs — jasonkwon · 2026-08-12
- Recycling Wasted Logits from Speculative Decoding for On-Policy Training — ycombinator · 2026-08-12
- Network Capacity for Two AI Data Centers Could Double a Decade of Global Backbone: Analyst — zephyr_z9 · 2026-08-12