Climate Impact:
Decentralized Sustainable AI Computing

Model Dashboard

Core metrics at a glance. Forecast year 2035 unless noted.

* Avoided emissions shown assume 1% market capture rate.

Projecting to Market Scale

At 647 TWh of global data center energy consumption (2035 forecast) and a unit impact of 0.031 Mt CO2e per TWh, at just 1% market capture, the avoided emissions would total approximately 203 thousand tonnes (kt) CO2e per year. The data center market is one of the fastest-growing energy consumers globally, driven by AI training and inference workloads.

Data center energy consumption is projected to grow 79% from 361 TWh (2025) to 647 TWh (2035), driven primarily by AI training and inference workloads. The unit impact declines modestly over the period (from 0.037 to 0.031 Mt/TWh) as baseline data center operations improve through broader industry efficiency gains.

The platform implements a distributed edge-cloud architecture with workload-aware ML schedulers, energy-proportional runtimes, and hardware power-management optimizations. By performing inference and aggregation closer to data sources, it reduces network transfers, increases average server utilization, and enables temporal shifting of compute toward lower-carbon electricity periods.

Key Findings

1. 1

   Incremental efficiency in the fastest-growing energy sector

   A 9.5% per-unit emissions reduction is modest, but data center energy consumption is projected to nearly double by 2035. Even incremental efficiency gains in a market growing this fast compound into meaningful absolute reductions - approximately 203 kt CO2e at 1% market capture.

2. 2

   AI itself is the growth driver and the solution

   The irony is not lost: AI workloads are the primary driver of data center energy growth, and this technology uses AI (workload-aware ML schedulers) to reduce the energy impact of those same workloads. The net effect depends on whether efficiency gains outpace the rebound effect of making compute cheaper and more available.

3. 3

   Edge computing reduces network and cooling overhead

   By performing inference closer to data sources, the platform reduces energy-intensive network transfers and can leverage smaller, more efficiently cooled facilities. Data-locality strategies and model partitioning allow workloads to run where energy is cleanest and cooling is cheapest.

4. 4

   Solution intensity carries lower confidence

   While baseline data and market sizing are fully validated, the solution intensity is an AI-assisted initial estimate (AI0) pending expert review, and market capture is only partially validated. This means the 9.5% reduction figure should be treated as preliminary until further validation is complete.

Methodology & Data Provenance

Data Quality Assessment

References & Resources

• Koi Data & Methodology Overview
• Koi Avoided Emissions: Terms & Concepts
• IEA - Data Centres and Data Transmission Networks
• Full Model Datasheet (Koi platform)

Published by Rho Impact. Data sourced from the Koi Data Lake. Last updated March 2026.