The material cost of computation: a review of current accounting practice
Digital infrastructure has a physical footprint in energy, water, land and minerals. This report reviews how that footprint is currently reported, and why the reported figures tend to fall as the actual demand rises.
Digital systems are routinely described as immaterial, and are accounted for accordingly. This report reviews the accounting practices used to report the environmental cost of computing infrastructure and identifies where those practices systematically understate it.
The efficiency paradox in reporting
The dominant reported metric is efficiency: energy per unit of computation, water per unit of cooling, emissions per transaction. These figures have improved substantially and continue to improve. They are also close to useless as a measure of total impact, because efficiency gains in computing have historically funded expansion rather than reduction.
An organisation can report a genuine, audited, year-on-year improvement in efficiency while its absolute consumption rises. Both statements are true. Only one is usually published.
Four accounting choices that shift the total
- Boundary selection. Whether embodied cost in hardware manufacture falls inside the reporting boundary changes the total substantially, and practice varies without disclosure.
- Market-based instruments. Contractual attribution of renewable generation can reduce reported emissions without changing the physical demand placed on a grid at the time it is placed.
- Water accounting. Withdrawal and consumption are frequently reported interchangeably, and locational scarcity is rarely weighted at all.
- Attribution of shared infrastructure. Where cost is apportioned across tenants, the apportionment method is usually undisclosed and frequently flattering.
Recommendations
We recommend that absolute consumption be reported alongside every efficiency figure; that reporting boundaries be disclosed rather than assumed; that water be reported as consumption weighted by local scarcity; and that market-based and location-based emissions figures be published together rather than the more favourable one being selected.
None of this is technically difficult. All of it is already done by some operators. The obstacle is that no obligation requires it, and the resulting figures are less pleasant than the ones currently published.
Methods, the full review protocol and the reporting frameworks assessed are set out in the appendices. The underlying assessment matrix is available on request.
