Transparens / Opacitet

Energy measurement: feasibility and regulation

What can already be measured, what standards are missing, and how regulatory demands fit in.

Measurement already exists; what's missing is auditable disclosure

The argument that measuring consumption per query is technically impossible no longer holds. The available evidence shows sufficient tools, precedents, and regulatory pathways to demand transparency, even though the industry continues using operational complexity as a narrative shield.

Audited methods

5

From direct hardware telemetry to external black-box auditing.

Real production precedent

0.24 Wh

Google already published a median per query and broke the technical impossibility argument.

Plausible regulatory models

4

Intervention doesn't depend on inventing new science, but on setting comparable obligations.

Closed providers continue to claim that multi-tenant architecture, batching, and IP protection prevent publishing energy per query in production. That narrative no longer fits the available technical evidence.

The real question is different: who decides to expose the telemetry, at what granularity, and under what audit. Without that transparency layer, climate regulation and commercial comparison continue to operate on indirect approximations.

AISHA: the critical barrier is no longer physical. It is a combination of product willingness, reputational risk, and the absence of truly granular mandatory standards.

Existing standards and initiatives (2026)

The regulatory ecosystem is fragmented. Laws coexist with voluntary frameworks and governance certifications that rarely require publishing energy per query. The question is not whether standards exist, but whether they truly force usable transparency.

Legal framework

EU AI Act (art. 53.1.e)

Scope: Europe

Opens the door to documenting resources used by GPAI, but remains ambiguous about whether this requires publishing energy per inference in a comparable way.

Corporate reporting

CSRD / ESRS E1

Scope: Europe

Strengthens the obligation to report corporate emissions, although it still treats data centers as an aggregated mass rather than as disaggregated AI services.

Legal framework

California SB 253 / SB 261

Scope: California

Increases pressure on climate disclosure for large companies, but still does not require an operational energy figure per call or per agent session.

Open benchmark

Hugging Face AI Energy Score

Scope: Global

It is the most visible attempt to standardize comparative efficiency among open models, although the bulk of dominant closed platforms remain outside the benchmark.

Methodological framework

SCI for AI

Scope: Global

Provides a solid methodology for software carbon intensity, but remains too technical and insufficiently adopted by the executive and regulatory layer.

Certification

ISO/IEC 42001

Scope: Global

Organizes AI governance, risk, and processes, but its environmental component remains declarative and does not force the publication of granular product telemetry.

AISHA reading

Today there is far more declarative governance than operational transparency. Most frameworks accept aggregated company or data center footprints but do not force a useful label at the service level.

This explains why the market can simultaneously claim that sustainability matters and that publishing energy per query would be unfeasible. Current regulations are not yet designed to close that contradiction.

Mandatory requirements still operate at the corporate level, not at the API level.
Voluntary standards help compare open source, but do not discipline closed providers.
The critical gap remains in agents, commercial video, and recurring inference.

Plausible regulatory models for 2026–2028

If measurement is already feasible, the next discussion is political: what regulatory design can force useful transparency without turning the system into a blind bureaucracy or a greenwashing compliance exercise.

Multidimensional evaluation of regulatory models

Relative comparison across technical feasibility, political feasibility, implementation speed, and potential impact on real reduction.

Model 1

Inference energy label

Advantage: easy to communicate

Audited laboratory benchmark with a visible A–G classification on the public product or model listing.

Good entry point for clients and regulators.
Risk: oversimplifying very different workloads.

Model 2

API transparency mandate

Advantage: real granularity

Requires returning estimated energy and carbon per call in the service's own technical response.

It is the most useful approach for operations, procurement, and auditing.
Requires the provider to convert internal telemetry into a stable product.

Model 3

Corporate reporting extension

Advantage: fits with ESG reporting

Expands CSRD and equivalent frameworks to separate AI compute from general cloud and require breakdown by service family.

Can advance quickly because it reuses existing compliance infrastructure.
Still falls short for comparing specific modalities or prompts.

Model 4

Carbon tax per compute

Advantage: corrects incentives

Progressive levy tied to energy cost or effective computation per token, query, or automated session.

Would be powerful for driving efficiency, but is the hardest to legislate today.
Without robust prior measurement, the tax becomes politically unfeasible.

What to prioritize first

The most realistic sequence does not start with the tax. It starts by standardizing evidence: an audited label, a reproducible benchmark, and a technical disclosure obligation at the product level for the most intensive services.

Once a comparable metric exists, the discussion changes entirely. Without mandatory measurement, every subsequent policy inherits the original opacity.

Källor

Samma kategori

Energy measurement: feasibility and regulation

Measurement already exists; what's missing is auditable disclosure

Technical feasibility assessment of measurement

Methods analyzed

Comparison: technical precision vs. commercial scalability

Existing standards and initiatives (2026)

EU AI Act (art. 53.1.e)

CSRD / ESRS E1

California SB 253 / SB 261

Hugging Face AI Energy Score

SCI for AI

ISO/IEC 42001

AISHA reading

Dismantling opacity: real obstacles vs. corporate excuses

Frictions for measurement

Analysis of measurement friction

Plausible regulatory models for 2026–2028

Multidimensional evaluation of regulatory models

Inference energy label

API transparency mandate

Corporate reporting extension

Carbon tax per compute

What to prioritize first

Källor

Fortsätt utforska transparens / opacitet

Energy measurement: feasibility and regulation

Measurement already exists; what's missing is auditable disclosure

Technical feasibility assessment of measurement

Methods analyzed

Comparison: technical precision vs. commercial scalability

Existing standards and initiatives (2026)

EU AI Act (art. 53.1.e)

CSRD / ESRS E1

California SB 253 / SB 261

Hugging Face AI Energy Score

SCI for AI

ISO/IEC 42001

AISHA reading

Dismantling opacity: real obstacles vs. corporate excuses

Frictions for measurement

Analysis of measurement friction

Plausible regulatory models for 2026–2028

Multidimensional evaluation of regulatory models

Inference energy label

API transparency mandate

Corporate reporting extension

Carbon tax per compute

What to prioritize first

Källor

Fortsätt utforska transparens / opacitet

Forensiskt inventarium över energiopacitet

Vem gynnas av energiopaciteten

Impacto hipotético de la transparencia energética