WattTime’s 25% Carbon Opportunity Reshapes AI Operations

WattTime’s March 2026 update to its North America grid emissions models quietly changed the economics of sustainable AI operations. The update improves the accuracy of its CO2 Marginal Operating Emissions Rate (MOER) signals and, according to the company, creates roughly 25% more opportunity for CO2 reduction across the region for enterprises running carbon-aware workloads. For AI infrastructure teams already contending with surging energy bills, that is a material number.

Phase 1: Establish Baseline Measurement and Monitoring

Effective energy cost management starts with precise measurement. Without a clear picture of current consumption and emissions, optimisation efforts tend to miss their targets. Several tools now give enterprises the visibility they need at both the workload and grid level.

Granular carbon tracking with CodeCarbon. To understand the environmental impact of individual training runs and inference jobs, development teams are integrating CodeCarbon, an open-source Python library that tracks CO2 emissions generated by computing processes directly within application code. Wrapping training scripts or inference pipelines with CodeCarbon lets teams record energy consumption and CO2e emissions, establishing a baseline for specific models and workloads. The tool visualises outputs and provides real-world equivalents for context. A long-running GPU training job, for instance, may reveal a carbon footprint that was previously invisible and prompt a re-evaluation of whether the run is necessary or where it should execute.

Real-time grid emissions data via WattTime. Beyond internal workload tracking, understanding the carbon intensity of the electricity grid powering those workloads is essential. WattTime provides real-time and forecasted MOER data for electricity grids across more than 200 countries and territories. These signals show the change in emissions that would result from an incremental increase or decrease in electricity demand at any given moment, and they vary by time of day and region depending on the energy mix currently active. If a grid is predominantly powered by renewables during certain hours, running AI tasks in those windows can materially reduce carbon output compared to peak periods when fossil fuel plants carry more of the load. WattTime’s 2026 North America model update improved the accuracy of these signals, with the company projecting a roughly 25% increase in potential CO2 reduction for users in the region.

Microsoft Azure offers Azure Carbon Optimization, a feature designed to help IT professionals understand and reduce emissions from their Azure usage through actionable recommendations. These dashboards provide a macro view of an organisation’s cloud carbon footprint and help identify high-impact areas across regions and services.

Phase 2: Optimise Infrastructure for Green AI Operations

Hardware choices and data centre strategy set the floor for energy efficiency. Getting these decisions right compounds through every workload that runs on top of them.

Energy-efficient AI hardware. The choice of processing unit, whether GPU, CPU, or a specialised AI accelerator such as a TPU or AWS Trainium, directly drives energy consumption. Newer hardware generations deliver more computational work per watt. AWS recently announced the general availability of Graviton5, designed for agentic AI workloads and positioned around performance per watt. Pinterest has committed $4 billion to AWS through 2031, planning to use AWS Trainium and Graviton to train and run its AI models, citing the efficiency advantages of custom silicon. When procuring or provisioning compute, comparing power consumption against performance benchmarks such as TFLOPS per watt across different accelerators is a practical starting point. Inference workloads, which typically require less power per operation than training but scale massively in production, are driving particular interest in purpose-built accelerators.

Google Cloud has noted that selecting its lowest-carbon locations can lead to up to a 90% reduction in a workload’s gross carbon footprint compared to running the same job in a grid with a high fossil fuel mix, according to the company.

Advanced cooling technologies. Cooling high-density AI hardware is one of the largest energy draws in a modern data centre. AWS has highlighted custom liquid cooling systems developed in 11 months to meet the demands of denser AI chips. Amazon announced on June 11, 2026, that its data centres are more than seven times more water-efficient than the industry average, achieved primarily through air cooling for most of the year and expanded use of recycled-water cooling. Immersion cooling, where servers are submerged in dielectric liquid, is being evaluated by some enterprises for superior heat dissipation and reduced energy use compared to traditional air cooling.

Phase 3: Optimise AI Software and Models for Efficiency

Standardised methodologies for measuring AI’s climate impact are emerging, reinforcing the case for these software-centric approaches.

Smaller, more efficient models. Larger models carry a steep energy cost. Quantization reduces the precision of model weights, from 32-bit to 8-bit integers for example, with minimal accuracy loss, producing smaller model sizes, faster inference and lower energy consumption. Pruning removes redundant connections from a neural network, cutting computational requirements. Distillation transfers knowledge from a large teacher model to a smaller student model, achieving comparable performance with substantially less compute. Google’s “4Ms” best practices for AI sustainability recommend selecting efficient model architectures, noting that sparse models can improve quality and reduce computation by 3 to 10 times compared to dense equivalents, according to Google.

Training algorithms and processes. Early stopping halts training once validation performance plateaus, preventing unnecessary computation. Dynamic batching adjusts batch size during training to maximise GPU utilisation and reduce idle time. Distributed training, when managed well, can reduce energy use by completing runs faster, though poorly designed distributed systems can introduce overhead that erases those gains. Research into more energy-efficient optimisers continues to advance.

Inference efficiency and edge deployment. Inference typically accounts for a large share of an AI system’s lifetime energy consumption, given how frequently it executes at scale. Specialised inference engines such as NVIDIA TensorRT and OpenVINO convert models into hardware-optimised formats that run significantly faster and at lower power. Deploying inference closer to the data source or end user, edge AI, reduces data transfer overhead and uses lower-power devices. AWS has positioned its custom inference chips and smart routing systems as central to making AI inference more affordable at scale. For enterprises scaling AI in production, the shift in focus from peak training performance to full-lifecycle efficiency, including inference, is where the largest energy savings are now available. This connects directly to broader questions about scaling enterprise AI infrastructure sustainably.

Phase 4: Implement Carbon-Aware Workload Scheduling

With measurement in place and infrastructure and software optimised, the next step is scheduling AI workloads to coincide with periods of low carbon intensity on the electricity grid. This is where WattTime’s March 2026 update has the most direct operational impact.

Integrating WattTime MOER signals into schedulers. The practical foundation of carbon-aware scheduling is feeding WattTime’s real-time and forecasted MOER data directly into workload management systems and cloud schedulers. WattTime’s API provides granular data by grid region, allowing automated systems to decide when and where to run compute-intensive AI tasks. An MLOps pipeline, for example, could be configured to pause non-critical model training or large-scale data processing during high-carbon grid windows and resume during periods of strong wind or solar generation. The March 2026 WattTime update improved the accuracy of these signals across North America, giving scheduling systems a more reliable emissions signal to act on.

Geographic and temporal load shifting. Enterprises running geographically distributed data centres or multi-region cloud deployments can use MOER data to shift workloads both in time and in location. A training job that could run in a Pacific Northwest cloud region, often supplied by hydropower, or in a coal-heavy Midwest region presents a clear carbon optimisation choice. For global operations, carbon-aware scheduling means routing AI tasks toward regions where renewable energy is available, compounding the benefit of the roughly 25% increased CO2 reduction opportunity from WattTime’s updated North America models, per the company. WattTime has noted that its MOER data can be used by flexible, internet-connected infrastructure, including data centres and AI training jobs, for load-shifting to reduce emissions.

Automated carbon-aware orchestration. Building custom integrations is complex, and some enterprises are developing orchestration platforms that treat carbon intensity as a first-class scheduling variable alongside cost and latency. These platforms adjust batch job queues, scale cloud resources and, in some cases, migrate containerised AI workloads based on real-time MOER data and predefined carbon reduction targets. Fully automated systems of this kind are still maturing, but the availability of precise, API-accessible emissions data from providers like WattTime is accelerating development. The Green Software Foundation’s expanded scope, covering energy, water and waste, reinforces the direction of travel toward holistic, automated environmental accounting across the technology stack. For teams thinking about AI security posture alongside sustainability, the same agent-level governance questions apply, as explored in recent work on governing web-connected AI agents.

Pulling It Together

Managing the energy cost of AI workloads is now a hard operational and financial problem, not a sustainability add-on. The tools to address it exist at every layer: CodeCarbon for per-job emissions tracking, WattTime’s updated MOER signals for grid-level carbon-aware scheduling, and native carbon reporting from AWS and Google Cloud for macro visibility. The enterprises making progress are treating these as integrated system-level decisions rather than isolated hardware procurement choices. WattTime’s March 2026 update gives North American operations a more accurate emissions signal to schedule against. The question for most enterprises is whether their MLOps pipelines are instrumented to act on it. For more analysis on enterprise AI strategy, visit our Enterprise AI section.