Dali and Carbon-Aware Inference: Grounding AI Load in Real Grid Data

GammaLex ships Dali as a control plane so inference is not an unmanaged tap on the grid: every request is classified, gated on budget and policy, and routed with observability. That design choice matters because third-party evidence shows AI infrastructure is becoming a first-class electricity story—not a footnote in sustainability decks.

What the IEA measures: load, growth, and emissions

In its 2025 special report Energy and AI, the International Energy Agency (IEA) estimates global data centres accounted for about 1.5% of world electricity consumption in 2024—roughly 415 terawatt-hours (TWh)—with consumption having grown about 12% per year since 2017, more than four times the rate of total electricity demand growth. The same analysis projects data-centre electricity use more than doubling to around 945 TWh by 2030 (in the IEA’s central outlook), with the United States accounting for the largest national share of both current use and projected growth.

On emissions, the IEA reports that carbon dioxide from data-centre electricity use grows from about 180 million tonnes today toward roughly 300 million tonnes by 2035 in its Base Case, and up to about 500 million tonnes in a higher-growth “Lift-Off” sensitivity—still a modest share of total energy-sector emissions, but among the fastest-growing sources. The report is explicit that AI is not a climate strategy by itself: efficiency gains and clean procurement must be real, measured, and sustained—not assumed from vendor marketing.

Why “green cloud” labels are not enough

Average annual renewable share or a corporate PPA does not fully describe the carbon intensity of the next megawatt-hour at a specific time and place. For U.S. planning and disclosure, the Environmental Protection Agency’s Emissions & Generation Resource Integrated Database (eGRID) publishes regional and subregional CO₂ emission rates for grid electricity—practical reference values when teams model inference deployed in different Independent System Operator / Regional Transmission Organization footprints.

The IEA also stresses grid integration risks: connection queues, transmission lead times, and concentration of new capacity in a handful of clusters—meaning “move compute to a cheap region” can collide with local constraints unless routing decisions are coordinated with reliability and compliance requirements, not only headline price.

How Dali operationalizes energy and carbon context

Dali is built to pair governance gates with routing and measurement: cost validation, energy and carbon assessment as signals allow, security and compliance checks, then real-time provider and region selection when policy permits. Failed requests are blocked before downstream infrastructure is touched; allowed requests carry an audit trail that ties decisions to the policies and inputs that were true at execution time—so sustainability, finance, and security teams are not reconciling three different spreadsheets after the fact.

The objective matches what public data imply: when data-centre load scales at the pace the IEA describes, inference must be treated as flexible, reportable infrastructure load—routed with the same seriousness as other large industrial electricity users, and evidenced with references regulators and counterparties can inspect.

Sources (open, citable)

International Energy Agency (2025). Energy and AI—Executive Summary. https://www.iea.org/reports/energy-and-ai/executive-summary — global data-centre electricity and emissions outlook, grid integration, and sensitivity cases cited above.

U.S. Environmental Protection Agency. Emissions & Generation Resource Integrated Database (eGRID). https://www.epa.gov/egrid — subregion-level CO₂ emission rates for U.S. grid electricity, updated on a published cadence for inventory and analysis work.

U.S. Energy Information Administration. Electricity explained & state electricity profiles. https://www.eia.gov/electricity/ — contextual generation mix and demand trends by state and region (useful alongside eGRID for narrative and planning).