Cloud migration: a 60% saving that shrank to 22%

The business case compared unit efficiency, not absolute consumption. Hyperscale facilities do deliver better power usage effectiveness and often better procurement, but absolute energy is efficiency per unit multiplied by the number of units. The migration changed both terms, and only one was modelled.

Name the mechanisms visible in the data:

• Rebound effect: capacity became cheap and self-serviceable, so 140 environments appeared that never existed on-premises. Consumption rose because friction fell.
• Placement: latency-tolerant analytics were put in a high-carbon region, so the same compute produces more emissions.
• Data growth: unreclaimed snapshots tripled storage, and storage carries energy and cost even when idle.

Rebuild the measurement on a defensible baseline. Fix the boundary - compute, storage, network, and supporting services - and use metered baseline data from a representative period, not vendor benchmarks. Record which provider allocation method produced the emissions figures and confirm it covers the services in scope.

Report both accounting views (location-based and market-based) from the outset, and state in the target which one applies. A target set on one basis and reported on the other is a restatement waiting to happen.

Track leading indicators monthly, not annually: vCPU and storage volumes, idle and orphaned resources, region carbon intensity by workload, and cost as a proxy for consumption. These move before the emissions figure does, which is what makes them useful for management.

Clarify where the emissions sit before arguing about the number. For a cloud customer, electricity consumed in the provider's facilities is the provider's scope 2 and the customer's scope 3, purchased services. The customer does not get to book the cloud's electricity as its own scope 2. That structural point settles many disputes about which figure to quote.

Distinguish the two methods precisely. Location-based reporting uses the average grid intensity where the equipment runs. Market-based reporting uses contractual instruments - certificates or power purchase agreements - to reflect procured renewable electricity. Both are legitimate; they answer different questions. Location-based reflects the physical burden on the grid; market-based reflects procurement.

Test the quality of the claim. Global annual matching means the provider bought certificates equal to its total consumption over the year, somewhere. It does not mean the specific region running Halden's analytics had clean electrons in the hours the workload ran. That gap is precisely where the three relocated analytics workloads cause real emissions.

Guard the target's integrity. If the 2026 target was set on location-based figures, switching to market-based to close the gap is a methodology change dressed as progress, and it will be read that way. Either hold the basis constant and explain the shortfall, or restate the target transparently with the reason and the effect quantified.

Treat procurement as a separate lever. Renewable purchasing can improve the market-based result, but it should be presented as procurement, not as demand reduction. Blurring the two is the most common route from a defensible position into a greenwashing allegation.

Attack demand first, because it is the only lever fully under Halden's control.

• Non-production: automated shutdown outside working hours, time-boxed environments with expiry dates, and an approval gate for new ones. The 140 extra environments are the largest single anomaly and the cheapest to fix.
• Storage: lifecycle policies, snapshot retention limits, and deletion of orphaned volumes. Tripled storage with no retention policy is a governance failure, not a technology problem.
• Rightsizing: match instance size to measured utilisation, and put cost and carbon on the same dashboard with the same owner, so the incentive points one way.
• Placement policy: route latency-tolerant batch work to low-carbon regions as a standing rule, not a one-off optimisation.
• Architecture: reduce data movement and duplication; inefficient data flows cost energy twice, once to move and once to store.

Then re-forecast honestly. Build a bottom-up estimate from current volumes and the committed actions, and present the CSO with a realistic 2026 position and the size of any residual gap. A target missed with a credible recovery plan is survivable; a target quietly restated in the final quarter is not.

Invest in the contract. Require provider emissions data at the granularity Halden reports at - service, region and period - with the methodology documented. You cannot manage a boundary you cannot see.

If the gap cannot close, disclose it with the cause, the corrective actions and a revised date. Restating a public target is a governance decision for the board, not a communications decision for the sustainability team.