GPU Cloud Carbon Intensity: Measuring gCO2/kWh for AI Workloads

Understanding the gCO2/kWh Metric

Grams of carbon dioxide equivalent per kilowatt-hour (gCO2/kWh) is the standard metric for quantifying the greenhouse gas emissions associated with electricity generation. It measures the total climate impact of producing one kilowatt-hour of electrical energy, expressed in CO2 equivalents. The metric accounts for direct emissions from fuel combustion at the generation source, as well as upstream emissions from fuel extraction, processing, and transportation. For renewable sources, it also includes lifecycle emissions from manufacturing solar panels, wind turbines, and associated infrastructure — though these are typically an order of magnitude lower than fossil fuel emissions.

The range of carbon intensity across global electricity grids is enormous. At the clean end, Iceland's geothermal and hydroelectric grid operates at approximately 10-15 gCO2/kWh. France's nuclear-heavy grid runs at roughly 50-60 gCO2/kWh. At the other extreme, grids dominated by coal — such as Poland's at approximately 700 gCO2/kWh or Australia's at roughly 600 gCO2/kWh — produce more than fifty times the emissions per kilowatt-hour. The global weighted average is approximately 450 gCO2/kWh, which is also the commonly cited industry average for data center operations worldwide. Against this backdrop, Harch Intelligence's measured carbon intensity of 47 gCO2/kWh — achieved through Morocco's 81.5% renewable grid combined with carbon-aware scheduling — represents an 89% reduction from the industry norm and places it in the same tier as the cleanest grids in Europe.

Scope 2 Emissions: Location-Based vs Market-Based Accounting

Under the GHG Protocol, data center carbon emissions fall under Scope 2 — indirect emissions from purchased electricity. The protocol defines two methods for calculating Scope 2 emissions. Location-based accounting uses the average emissions intensity of the local grid where the data center operates. This method reflects the physical reality of the electricity flowing into the facility. Market-based accounting uses emissions factors from specific contractual instruments — such as renewable energy certificates (RECs), power purchase agreements (PPAs), or supplier-specific emission rates. This method reflects the purchasing decisions of the data center operator. Both methods must be reported under the GHG Protocol, and the gap between them reveals whether a facility is truly powered by clean energy or merely purchasing certificates.

This distinction is critical for evaluating GPU cloud providers. A data center in Virginia running on a grid with 400 gCO2/kWh location-based intensity can claim 0 gCO2/kWh market-based intensity by purchasing unbundled RECs from a wind farm in Texas — even though no physical electrons from that wind farm reach the data center. Harch Intelligence's advantage is that its location-based and market-based intensities are closely aligned, because the Moroccan grid itself is 81.5% renewable. The 47 gCO2/kWh figure is a location-based measurement reflecting the actual carbon intensity of the electricity powering the GPUs — not a market construct enabled by certificate purchases from distant generators.

Tools and Data Sources for Carbon Intensity Measurement

Real-time carbon intensity data is available through several established APIs. ElectricityMap provides hourly carbon intensity data for over 100 grids worldwide, sourcing data from national grid operators and ENTSO-E. WattTime offers marginal emissions intensity data — the emissions associated with the next megawatt-hour of generation dispatched to meet load — which is more relevant for carbon-aware scheduling than average intensity. The Carbon Aware SDK, an open-source project under the Green Software Foundation, provides a standardized interface for querying carbon intensity across multiple data sources, enabling developers to build carbon-aware applications without coupling to a single provider.

HarchOS integrates all three data sources into its scheduling engine. The platform ingests real-time carbon intensity feeds from Morocco's Office National de l'Electricite et de l'Eau Potable (ONEE), cross-referenced with electricityMap's granular hourly data and WattTime's marginal emissions signals. On-site metering at each of the five hub locations provides a third data layer — actual electricity consumption measured at the busbar, enabling per-workload carbon accounting that goes beyond grid averages to reflect the specific energy mix powering each GPU at each moment.

Per-Workload Carbon Accounting with HarchOS

The most advanced form of carbon measurement is per-workload accounting — attributing specific emissions to individual AI training jobs or inference requests. HarchOS implements this by tracking GPU utilization, power draw, and grid carbon intensity at one-minute intervals for every job running on the 1,798-GPU fleet. When a model training job consumes 2.4 MWh of electricity over 72 hours at an average carbon intensity of 42 gCO2/kWh, HarchOS records a total emission of 100.8 grams of CO2 equivalent. This granular accounting enables customers to report actual emissions in their sustainability disclosures rather than relying on industry averages or estimates.

Per-workload carbon accounting also enables optimization. When HarchOS identifies that a deferred training job would have generated 35% less carbon if shifted 4 hours to align with peak solar generation, it records both the actual and the avoided emissions — providing a quantitative measure of the carbon-aware scheduling benefit. Over a quarter, these avoided emissions aggregate into a carbon savings report that customers can include in ESG filings, demonstrating not just low emissions but active emission reduction through intelligent scheduling.

Why Accurate Measurement Matters Now

Carbon measurement is no longer optional. The EU Corporate Sustainability Reporting Directive (CSRD), effective from 2024, requires large companies to report detailed Scope 2 emissions including both location-based and market-based figures. The SEC's climate disclosure rules, though contested, are moving toward similar requirements for US-listed companies. ESG investment criteria — used by funds managing over $40 trillion in assets — increasingly require verified emissions data rather than self-reported estimates. Companies running AI workloads on GPU clouds that cannot provide per-workload carbon reports face a growing compliance gap that will widen as regulations tighten.

Carbon Intensity by Region: A Comparative View

The following comparison illustrates the dramatic variation in grid carbon intensity that determines the environmental impact of GPU workloads. Iceland: 10-15 gCO2/kWh (geothermal and hydro). Morocco: 47 gCO2/kWh (81.5% renewable, Harch Intelligence measured). France: 55 gCO2/kWh (nuclear-dominant). United Kingdom: 230 gCO2/kWh (mixed gas and wind). United States (Virginia): 380 gCO2/kWh (mixed gas, nuclear, coal). Germany: 350 gCO2/kWh (coal and gas transitioning to renewables). India: 700 gCO2/kWh (coal-dominant). Poland: 720 gCO2/kWh (coal-dominant). The data makes clear that where a workload runs matters as much as how efficiently it runs — a training job producing 100 grams of CO2 in Morocco would produce over 1,500 grams in Poland for the same computational output.

Temporal Variation: Time of Day and Season

Grid carbon intensity is not constant — it fluctuates significantly based on the generation mix at any given moment. On Morocco's solar-heavy grid, carbon intensity drops to its lowest between 10:00 and 15:00 local time when solar farms are at peak output, and rises in the evening as solar generation declines and natural gas plants ramp up. Seasonal variation adds another dimension: summer months with longer solar days produce lower average carbon intensity than winter months with shorter days and higher heating-driven demand. Carbon-aware schedulers exploit these temporal patterns, deferring non-urgent GPU workloads to periods of low carbon intensity and running critical workloads during high-carbon periods only when necessary. HarchOS's carbon-aware scheduling reduces fleet-wide carbon intensity by an additional 15-20% beyond what Morocco's grid mix alone would deliver, achieving the 47 gCO2/kWh average through intelligent temporal placement.

Third-Party Auditing and Verification

Self-reported carbon data is increasingly scrutinized by regulators and investors. Harch Intelligence subjects its carbon intensity measurements to independent third-party auditing on a quarterly basis. Auditors verify the alignment between reported grid carbon intensity and data from electricityMap and ONEE, confirm that per-workload energy measurements match utility meter readings, and validate that carbon-aware scheduling decisions are executed as claimed. The audit results are published in Harch Intelligence's quarterly sustainability report, providing the transparency that regulators, investors, and enterprise customers require. In a market where greenwashing is a genuine concern — and where some GPU cloud providers report market-based figures that diverge dramatically from their location-based reality — independently verified measurements are the only credible foundation for sustainability claims.