Skip to content
Blog

The Ananse Reports

Measurements of what edge AI actually costs to run, taken on a bench in Accra and published with the rig, the firmware and the raw telemetry attached. Each report answers a question the last one raised.

3 published · fortnightly · agoo-ai.com/blog/ananse-reports
Start here Introduction

The introduction to the series: what is being measured, what the numbering means, and what a reader is meant to be able to check.

Published · 3 min · Outside the numbering

The reports

—
IntroductionIntroduction · Outside the numbering
0.5 The Boards and the Measurement Rig

The bench itself: what the meters are, how they are sampled, and what the idle floor costs before a single token is generated.

Companion · no data of its own 5 min 1 I TinyML Reality Check, Part I: The Floor of the Curve

Three MCU boards, three TinyML workloads, and a measured answer to what AI inference costs in energy on real hardware.

Integer · Part I of three · has data 8 min 1 II TinyML Reality Check, Part II: The Evidence

The full TinyML matrix: energy, latency, accuracy, variance, and the one result that needed a second look.

Part II of No. 1 · same number 10 min
1 III TinyML Reality Check, Part III: What It Means Part III of No. 1 · same number Early Sep
1.5 The 43 Milliwatt Problem Companion Late Sep