AI carbon governance · United Kingdom · Pilot programme
Audit-grade carbon accounting for enterprise AI.
CSRD and ESRS E1 ask enterprises to disclose the emissions of their value chain — and AI usage is becoming a material, unmetered line in it. ACERM meters AI workloads, prices them against the live grid, and writes the evidence down.
● Pilot stage — seeking design partnersTwo UK patent applications filedLondon
Disclosure rules assume you can measure. AI usage breaks that assumption.
Under CSRD, ESRS E1 asks for energy and emissions across the value chain, stated in a form an assurance provider can stand behind. Most enterprise AI consumption — API calls to foundation models, internal copilots, agent workloads — produces no meter reading at all. What exists instead is vendor marketing, screenshots of dashboards, and estimates built on other estimates.
That is the gap ACERM is built for: not another sustainability pledge, but a measurement layer that turns AI usage into ledger entries a compliance team can actually file.
Meter. Price. Ledger.
ACERM sits in the request path of your AI usage and treats every call as a metered event. Three stages, each designed to be checkable rather than taken on trust:
MMetering
Each request is converted to energy via a calibrated compute-usage estimate (CUE), validated against wall-plug measurements on real hardware rather than assumed from vendor figures.
PLive pricing
Energy is priced into CO₂e against live UK grid intensity from the NESO API — the same feed running at the top of this page — with declared registry conventions for workloads that can't be geographically attributed.
LLedger
Every priced event is written to an append-only audit ledger: timestamped, source-attributed, and designed so that a third party can re-derive each entry from its inputs.
What we have measured so far — labelled as exactly that.
| Result | Measurement | Context |
|---|---|---|
| 76.2% | energy reduction, 1.58-bit ternary model vs frontier-class baseline | Internal benchmark. Falcon3-1B-Instruct (1.58-bit), wall-plug metered on Apple M1. |
| ±13% | CUE-to-energy linearity across a 16× workload range | Metered validation on Apple M1 and NVIDIA A100 (cloud). |
| 16 / 16 | unit tests passing on the calibration and pricing codebase | Python reference implementation. |
How to read this table. These are internal benchmarks: measured by us, repeatable from our method notes, and not yet third-party assured. We publish them with their conditions attached because that is the standard we are asking the industry to meet. No deployment-scale figures appear on this site until a deployment produces them.
Seven days. Zero integration. Your AI estate, measured.
Shadow Mode is a read-only audit: we take a sample of your organisation's AI usage patterns and produce an evidence pack — estimated energy, CO₂e priced against the live grid, and a worked example of what your ESRS E1 AI line could look like with real measurement behind it.
Nothing touches your infrastructure. No SDK, no proxy change, no procurement cycle. It is a free sample of the ledger, built to be shown to your auditors.
Where this stands, plainly.
ACERM is at pilot stage. Two United Kingdom patent applications are filed — GB2610865.4 and GB2612614.4 — covering the metering and governance architecture. The reference implementation is under active development, and we are selecting a small number of design partners for the first Shadow Mode audits.
We would rather tell you exactly where the edges are than let a website imply a scale we haven't reached. If measured AI carbon evidence matters to your disclosure obligations, we would like to talk.