AI

3 analyses tagged with "AI"

Gemma 4 MTP on GB10 — Inference 1.83x faster on 26B, 3.52x on 31B Dense

Gemma 4 MTP on GB10 — Inference 1.83x faster on 26B, 3.52x on 31B Dense

Independent measurement of Google's Gemma 4 MTP drafter on NVIDIA GB10 (GX10). 1.83x speedup on 26B-A4B and 3.52x on 31B Dense. No public independent benchmark of 31B Dense + 1 GPU + MTP could be found, making this the first such measurement. Reproduces and exceeds Google's 'up to 3x' claim on real hardware.

Read →
Decomposing the AI Value Chain in the Agentic Era

Decomposing the AI Value Chain in the Agentic Era

Building on SemiAnalysis's 'AI Value Capture', this analysis traces how value in the Agentic AI era propagates upstream — from model labs and NVIDIA to HBM, advanced packaging, equipment, and power/cooling — using detailed causal subgraphs from an internal investment-theme model. Source attribution is kept explicit throughout.

Read →
How DeepSeek V4's 90% KV Cache Reduction Reshapes HBM Demand — From Capacity to Bandwidth, Packaging, and Thermal Control

How DeepSeek V4's 90% KV Cache Reduction Reshapes HBM Demand — From Capacity to Bandwidth, Packaging, and Thermal Control

DeepSeek V4 claims a 90% KV cache reduction at 1M tokens. This analysis argues HBM demand is not destroyed — its center of value shifts from raw capacity to bandwidth, packaging, and thermal control. Examined across technology, market, and supply chain layers.

Read →