Carbon-Aware Scheduling: Time-of-Day and Region-Based Workload Placement

This article frames the practice, the temporal and spatial lever sub-types, the trade-offs, and the integration into the MLOps platform.

The variability of the grid emission factor

Grid emission factors vary on multiple time scales. Diurnally: in a grid with significant solar generation, the emission factor drops during midday hours when solar is dominant and rises in evening hours when fossil generation must compensate for the loss of solar. Seasonally: in a grid with significant wind generation, the emission factor varies with the seasonal wind pattern — typically lower in spring and autumn than in summer. Across regions: a hydroelectric-heavy grid (Iceland, Quebec) has structural emission factors near zero; a nuclear-heavy grid (France) has emission factors in the tens of grams per kilowatt-hour; a wind-and-solar-heavy grid (Texas, the Nordic countries) varies dramatically with weather; a coal-heavy grid (parts of India, China, Poland, the U.S. Midwest) has structural emission factors in the hundreds of grams per kilowatt-hour.

The spread between the cleanest hour and the dirtiest hour of a single grid is typically 2-5x. The spread between the cleanest available grid and the dirtiest available grid (for a workload that can be globally placed) is typically 10-50x. These spreads are the headroom that carbon-aware scheduling exploits.

The data sources for real-time grid emission factors include the WattTime, ElectricityMaps, and equivalent services that publish per-region, per-hour grid emission factors as APIs. The cloud providers have begun to integrate carbon-intensity signals into their region-selection and workload-scheduling APIs, making carbon-aware scheduling increasingly accessible to developers.

Temporal shifting

Temporal shifting is the practice of running a workload at a cleaner hour of the day or week than the workload would run by default. The opportunity is largest for workloads that can tolerate hour- or day-scale latency — training runs (typically tolerate days of latency), periodic batch inference (typically tolerates hours), retraining pipelines (typically tolerates days). The opportunity is smallest for workloads that must serve real-time queries — interactive inference (must run when the user asks).

The implementation pattern is to register the workload with a carbon-aware scheduler that consumes the regional grid-emission-factor forecast and places the workload in the lowest-emission window that satisfies the workload’s deadline. The Green Software Foundation publishes reference architectures for the integration.¹

The McKinsey State of AI surveys have begun to identify carbon-aware scheduling as a practice that the most sustainability-mature enterprise AI programs are adopting, particularly for the large-scale training and retraining workloads that dominate their operational emissions.²

Spatial shifting

Spatial shifting is the practice of running a workload in a cleaner geographic region than the workload would run by default. The opportunity is largest for workloads that have no data-residency or latency constraints tying them to a specific region — typically training runs and some batch inference. The opportunity is constrained for workloads that must run in a specific region for regulatory or latency reasons.

The implementation pattern is to maintain a portfolio of compute regions, monitor the structural and instantaneous emission factors of each, and place the workload in the lowest-emission region that satisfies the workload’s constraints. For an organization running on hyperscaler infrastructure, the spatial-shifting decision is typically a region-selection decision — choosing to run training in the Iceland or Sweden region rather than the Virginia or Mumbai region, for example.

Spatial shifting introduces secondary considerations: data-transfer emissions and latency for moving training data into the cleaner region; data-residency compliance with the EU General Data Protection Regulation and equivalent regional regimes; cost differences across regions. The practitioner must balance these against the emission savings.

The combination

The largest savings come from combining temporal and spatial shifting — running the workload in the cleanest hour of the cleanest available region. For a flexible training workload, the combined savings can be 80%+ relative to the default of “wherever there is capacity, whenever the queue empties.”

Maturity Indicators

The COMPEL D19 maturity rubric does not name carbon-aware scheduling explicitly but the Level 4 (Advanced) indicator that “model efficiency optimization is standard practice” includes scheduling-level optimization in its scope.³ An organization at Level 4 has integrated carbon-aware scheduling into its MLOps platform — by default, training jobs are submitted to the carbon-aware scheduler and executed in the lowest-emission window. An organization at Level 5 (Transformational) is publishing the avoided-emissions figure attributable to carbon-aware scheduling as part of its sustainability disclosure and is contributing the practice back to the industry.

The Stanford Foundation Model Transparency Index (FMTI) compute-layer scoring increasingly recognizes the disclosure of carbon-aware scheduling practices as a transparency indicator, which is creating procurement-decision visibility.⁴

Practical Application

A foundational practitioner who is rolling out carbon-aware scheduling should produce four artifacts.

Artifact 1: the workload-flexibility classification. A classification, for every AI workload, of whether the workload is temporal-shiftable, spatial-shiftable, both, or neither. The classification is the input to the scheduling-policy decisions.

Artifact 2: the grid-emission-factor data integration. A data feed that, for every region the organization uses, provides real-time and forecast grid-emission-factor data from WattTime, ElectricityMaps, the cloud provider’s API, or equivalent. The data feed is the input to the scheduler.

Artifact 3: the scheduling policy. The policy that, for each workload class, defines the maximum acceptable latency for shifting, the regions the workload can be placed in, the constraints (data-residency, network-egress, cost), and the optimization objective (minimize emissions subject to constraints).

Artifact 4: the avoided-emissions disclosure. A disclosure that, periodically (typically annually), reports the avoided emissions attributable to carbon-aware scheduling — the difference between the emissions the workloads would have produced under default placement and the emissions they actually produced under carbon-aware placement. The disclosure is the input to the ESG report and to the customer-facing sustainability narrative.

The Greenhouse Gas Protocol Scope 2 Guidance permits market-based reporting that reflects carbon-aware scheduling outcomes, and the European Union Corporate Sustainability Reporting Directive (CSRD) ESRS E1 disclosure recognizes year-over-year emission-intensity improvements that include scheduling-driven reductions.⁵ The International Energy Agency Electricity 2024 report’s projections of data-center growth assume continued adoption of demand-side flexibility — including carbon-aware scheduling — as a grid-management practice.⁶ The Organisation for Economic Co-operation and Development (OECD) AI Principles’ lifecycle framing supports the practitioner’s expectation that scheduling decisions are integrated into the AI program.⁷

Summary

Carbon-aware scheduling exploits the temporal and spatial variability of grid emission factors to reduce operational emissions by 30-70% for flexible workloads, with no change to model architecture, hardware, or procurement. Temporal shifting places training and batch workloads in the lowest-emission hour of the day; spatial shifting places workloads in the lowest-emission region; the combination produces the largest savings. Real-time grid-emission-factor data feeds (WattTime, ElectricityMaps, cloud-provider APIs) make the practice operationally accessible. The COMPEL D19 maturity rubric requires the practice as part of standardized efficiency optimization at Level 4, with avoided-emissions disclosure at Level 5. The next article, M1.9Embodied Carbon: Lifecycle Assessment of AI Hardware, develops the embodied-carbon category that the operational-emissions optimization eventually surfaces as the next frontier.

© FlowRidge.io — COMPEL AI Transformation Methodology. All rights reserved.