How We Achieved 47 gCO2/kWh: The Carbon-Aware Scheduling Algorithm

The global data center industry operates at an average carbon intensity of approximately 450 gCO2/kWh. The best hyperscale facilities, powered by renewable energy credits and power purchase agreements, achieve 150 to 200 gCO2/kWh on paper — though actual performance varies dramatically with time of day, season, and grid composition. Harch Intelligence's production compute infrastructure operates at 47 gCO2/kWh. Not on paper. Not on annual average. In real time, measured at the meter, validated by third-party auditors. This is the story of the algorithm that made it possible.

The carbon-aware scheduling algorithm operates on three optimization layers. The first layer, temporal shifting, exploits the fact that renewable generation follows predictable diurnal and seasonal patterns. Solar production in Morocco peaks between 10:00 and 15:00 local time; wind generation in the Sahel corridor peaks in the afternoon and evening. Jobs that can tolerate scheduling delays — batch training runs, data preprocessing pipelines, model evaluation sweeps — are automatically deferred to high-renewable windows. The algorithm maintains a priority queue ordered by job flexibility, carbon intensity forecast, and deadline constraints. In production, 34% of all compute workloads are temporally shifted with zero impact on delivery timelines.

The second layer, spatial routing, distributes workloads across Harch Intelligence's geographically distributed GPU clusters based on real-time carbon intensity at each location. The algorithm ingests grid carbon data from electricityMap and proprietary sensors every 30 seconds, overlays on-site renewable generation telemetry from Harch Energy's SCADA systems, and computes marginal carbon intensity for each cluster. When a job is submitted, it is routed to the cluster with the lowest marginal emissions — provided the latency and data sovereignty constraints are satisfied. Cross-cluster checkpointing occurs over dedicated fiber links with sub-200ms transfer times, making spatial routing transparent to the application layer.

The third layer, predictive procurement, uses machine learning models trained on two years of grid and weather data to forecast carbon intensity 24 to 72 hours ahead. When the model predicts a low-carbon window, the scheduler preemptively queues deferred workloads and pre-warms GPU clusters. When a high-carbon period is forecast, non-critical jobs are throttled or migrated. This predictive capability reduces the fraction of compute that occurs during carbon peaks from 40% to under 8%. The model achieves a carbon intensity forecast accuracy of 94% at the 24-hour horizon and 87% at 72 hours.

The combined effect is dramatic. Without carbon-aware scheduling, Harch Intelligence's clusters would operate at approximately 140 gCO2/kWh — already well below the industry average, thanks to Morocco's renewable-heavy grid. With the algorithm active, intensity drops to 47 gCO2/kWh. That 93 gCO2/kWh reduction translates to 12,400 tonnes of CO2 avoided annually across current operations. When the Dakhla 500MW facility reaches full capacity, the algorithm will avoid over 120,000 tonnes per year — equivalent to taking 26,000 cars off the road.

"We did not achieve 47 gCO2/kWh by buying offsets or purchasing renewable energy credits," stated Amine Harch El Korane, Founder and CEO of Harch Corp. "We achieved it by building infrastructure on top of the world's cheapest renewables and then writing software that extracts every gram of carbon advantage from that position. This is not greenwashing. This is engineering. And the result speaks for itself."