En
← All articles

AI climate pledges face a test of measurement

The Antalya pledge matters if companies can turn environmental claims into evidence that customers can compare and use.

ColumnAI FrontiersAuthorOpen Market NotesTypeArticle
A data center beside a power pylon and electricity and water gauges with red needles.

On November 16, the COP31 presidency and the International Telecommunication Union launched the Antalya AI Pledge, with progress tracking planned through 2030. ITU announcement

Its practical importance lies in a management question: what can a customer verify when an AI service is described as green? A device, a model and a data centre have different reporting boundaries. An attractive sustainability report may not answer a procurement team's question. The following is Open Market Notes analysis, not an audit of any signatory.

What a voluntary pledge can accomplish

ITU describes the initiative as a voluntary framework. Its launch does not itself impose one new legal obligation on every AI company worldwide. Initiative description

That distinction does not make participation commercially irrelevant. Buyers can put information requests into supplier questionnaires or future contracts; lenders can ask for supporting evidence when assessing projects. Those would be subsequent decisions, not legal changes automatically created by the announcement. A participation statement therefore needs follow-up: what was promised, which operations are covered, who checks it and how shortfalls are explained.

A pilot project and an entire company are different boundaries. A successful pilot can be valuable without representing a whole platform. If that distinction is blurred, the supplier receives a reputational benefit while customers inherit the work of interpreting it. Useful disclosure lets an outsider see the difference.

An electricity bill is not a complete answer

ITU-T L.1801 supplies a life-cycle methodology for AI systems; national and global aggregation are outside its scope. Standard summary System assessment and macroeconomic forecasting consequently answer different questions.

Consider a hypothetical customer-support service comparing two models. Model A uses less electricity per request, but model B completes the job with fewer retries. A comparison per request could favour A, while a comparison per completed task of equivalent quality could reverse the ranking. This is an illustration, not a benchmark. It explains why the unit of comparison must be defined before the result is presented. Response length, accuracy, latency and human review can all change what completing a task means.

Hardware replacement poses a similar question. Greater operating efficiency does not establish that immediately retiring existing equipment is better across its full life. Utilisation, manufacturing impacts, useful life and the old equipment's next use all affect that judgment. A single ratio cannot replace those assumptions.

Facility metrics and service metrics are not interchangeable

EU Regulation 2024/1364 establishes data-centre reporting and indicators covering energy, water and energy reuse. It is a jurisdiction-specific legal framework, distinct from Antalya's voluntary initiative. Regulation

The useful comparison concerns the object being measured. Facility performance helps assess infrastructure operations, but cannot alone show the resources needed by one model to process one document. Conversely, a model benchmark cannot represent every tenant of a facility.

Buyers can therefore assemble evidence from both directions: infrastructure data with a stated location, period and boundary, and records of completed work and quality within their own service. Only an explained allocation method can connect the two. Where observations are missing, a labelled estimate is more useful than unsupported decimal precision.

Efficiency and total demand require separate evidence

The IEA's 2026 update projects global data-centre electricity use rising from about 485 TWh in 2025 to 950 TWh in 2030. These are data-centre totals, not AI-only consumption. IEA report

For managers, the implication is to track consumption per task alongside activity. Halving resource use per task while tripling tasks still increases the total. This arithmetic example predicts no particular company's future; it shows why efficiency improvement and a lower absolute footprint need separate proof.

Saved capacity can reduce costs or support additional features. The latter may be worthwhile, but its value should be explained rather than counted automatically as an absolute environmental saving. Procurement can also test adjustable service configurations, reserving intensive processing for requests that genuinely need it.

The next test is the quality of disclosure

Counting signatories will not establish the initiative's impact. Baselines, boundaries, repeatable methods and subsequent updates will be more informative. Participants may choose different goals, but changes should be explained rather than silently moving the comparison point.

The announcement does not prove that AI delivers a net climate benefit or that participants have completed an operational transition. It creates an opportunity to ask better questions. For customers, its immediate value could be a more useful procurement requirement: identify where information originates, what was measured, what was estimated and which actual decision the evidence supports.

Information only. No investment, legal, tax, or financial advice.