DeepSeek V4.1 Flash cuts global KV cache to 890 bytes/token, but HBM demand may rise with agent swarms
teortaxesTex · x · 2026-09-12
An analysis of DeepSeek V4.1 Flash argues HBM capacity is less of a bottleneck: via CSA2 (cross-layer global KV reuse with FP4 KV caching and Top-K indices) plus SWA Bounded Replay, global KV drops from 3.5KB/tok to 890 bytes and persistent KV cache footprint falls to 1/8 of V4 Flash. But a rebuttal notes HBM demand isn't going away — at 4 agents, V4.1's HBM consumption/second exceeds V4's, and at 64 it exceeds V3.2. Jevons paradox strikes again.
Related event: DeepSeek V4.1 Flash Shrinks KV Cache 437x, Sparking HBM Debate(2 posts)→
More from Infra
- TSMC reportedly raising CoWoS packaging prices up to 20% next year on AI chip demand — Beth_Kindig · 2026-10-31
- effect-harness on Cloudflare: run agent loops, tools and credentials on separate VMs — samgoodwin89 · 2026-10-31
- Google's Project Suncatcher boots TPU in orbit and beams back first AI response — sundarpichai · 2026-10-31
- Together AI launches Post-Training platform for agentic RL with tool-use environments — togethercompute · 2026-10-31
- General Compute raises $100M for inference capacity as CoreWeave stock hits 12-month low — YvesMulkers · 2026-10-31
- Analyst warns of unstable AI compute feedback loop as Kioxia's NAND price warning triggers flash crash — bookwormengr · 2026-10-31