# Neuralwatt links prompt cache misses to higher inference energy use

_Published Friday, October 16, 2026 at 5:55 PM EDT · Infrastructure, AI · Latest · Tier 2 — Notable_

![Neuralwatt links prompt cache misses to higher inference energy use — Primary](https://cdn.prod.website-files.com/69451759a4952bcf865fbb2d/6acec5062fef8892df4256e2_6acec50428ae0c788db4d04a_harness-energy-v4-hero-harness-energy.png)

Neuralwatt says mostly uncached requests used roughly twice the energy of cached requests with comparable prompt and output sizes on its cloud; GLM-5.3 used about 2.5 times as much. The company compared requests from October 1 to 7 and separately tested how coding agents construct prompts.

In laboratory tests, every reproduced cache break came from an agent changing previously submitted content, forcing the server to recompute from the first difference. Production measurements also include cache eviction under memory pressure, which Neuralwatt cannot always distinguish from prompt edits.

## Sources

- [neuralwatt.com](https://neuralwatt.com/news/why-harnesses-matter-for-energy-efficiency)

---
Canonical: https://techandbusiness.org/newswire/WZi4BqXac3_0L6JI64KYMG
Published: 2026-10-16T21:55:13.635Z
Story chronology: 2026-10-16T21:01:23.000Z
Retrieved: 2026-10-17T00:53:53.347Z
Publisher: Tech & Business (techandbusiness.org)