GreenOps & Green IT

A workload's footprint depends directly on the carbon intensity of the electricity grid powering the data centre where it runs. The same processing can see its footprint vary by a wide factor depending on whether it runs on a largely decarbonised grid or on a mix heavy in fossil generation. Region selection therefore becomes a first-order lever, actionable without technical change wherever latency, sovereignty and data residency constraints allow it.

That intensity also varies over time, following wind and solar output and grid demand. Time-shiftable workloads, batch processing, retraining runs, backups and analytical processing peaks lend themselves to scheduling aligned with lower-intensity windows.

The contractual landscape deserves close reading: providers report through distinct methods, some valuing renewable energy attributes purchased on an annual basis, others targeting hourly matching between consumption and carbon-free generation. The difference between these two approaches substantially changes how a figure presented as neutral should be read.

Facility efficiency indicators complete the picture, overall data centre energy efficiency and associated water consumption, which determine the share of energy consumed beyond the useful computation itself.

Carbon accounting allocates emissions across three scopes. Scope 1 covers direct emissions from sources owned or controlled by the organisation, such as generators in an internal data centre. Scope 2 covers indirect emissions from purchased energy, mainly electricity consumed by owned facilities. Scope 3 covers all other indirect emissions across the value chain.

For most organisations, IT sits overwhelmingly in scope 3: Cloud services, SaaS, end user devices and hosting services all fall under upstream emissions. That positioning explains the difficulty of measurement, scope 3 depending on data produced by third parties, with heterogeneous methods and variable publication timelines.

Scope 2 admits two accounting methods. The location-based approach uses the average carbon intensity of the grid supplying the site. The market-based approach uses the contractual electricity attributes acquired by the organisation. Both figures legitimately coexist in an inventory and tell different stories: the first describes the physical footprint, the second the contractual effort committed. Propagation between the two scopes is the decisive point: your Cloud provider's scope 2, the electricity consumed by its data centres, constitutes your scope 3 category 1. You therefore inherit its accounting method before you even inherit its figure.

IT data usable in carbon accounting therefore requires knowing which scope it belongs to, which method produced it, and which share of the boundary it effectively covers.

Generative AI shifts the IT footprint onto three distinct items. Training a model concentrates intense consumption over a short period, with a footprint largely determined by placement and compute duration. Inference spreads lower consumption per request across volumes that grow structurally, making it the dominant item over the lifetime of a production service. Accelerated infrastructure itself carries a significant embodied footprint, amortised over an operating life that is often short given the pace of hardware generation renewal.

Reduction levers largely coincide with those of cost control. Routing a task to the lightest model reaching the required quality level reduces both the bill and the consumption. Effective caching avoids recomputing what has already been produced. Bounding agentic loops removes compute cycles without value. Optimised document retrieval limits the volume processed on each call.

At infrastructure level, accelerated capacity reserved and then left idle is waste that costs twice over, financially and environmentally. Tracking its real utilisation rate ranks among the most directly actionable indicators.

Region selection keeps its full weight here: a training workload scheduled on a decarbonised grid produces a substantially lower footprint than the same workload run elsewhere, for an identical technical result.