The Energy Bill Nobody Modelled: AI Workloads and Data Centres

For most of the past two decades, energy was a line in the facilities budget that nobody outside operations examined. The economics of compute-intensive workloads have changed that.

In this region the effect is amplified. Cooling load is a larger share of total consumption than in temperate climates, and it peaks precisely when the grid is most stressed.

The cost model moved and the budget did not

Cloud pricing made compute feel like a variable cost with no infrastructure consequence. For conventional workloads that held reasonably well.

Sustained high-intensity workloads break the assumption. Organisations are receiving cloud bills with a shape they did not forecast, and those running their own facilities are finding power and cooling capacity, rather than floor space, is the binding constraint.

Efficiency work has moved up the agenda

When compute was cheap relative to engineering time, optimisation was rarely worth doing. That trade has shifted.

Model selection, batching, caching and the discipline of asking whether a task requires the largest available model now have a measurable cost consequence. In several cases we have seen, careful engineering reduced spend more than any commercial negotiation would have.

Location decisions have a new input

Where workloads run is now a function of power availability and cost as much as latency and data residency. Regional capacity is expanding, and availability is not uniform.

Organisations planning significant capacity should be treating power as a procurement category with its own strategy, rather than an assumption embedded in a facilities contract.

Reporting obligations are catching up

Energy consumption and the associated emissions increasingly need to be reported, and for many organisations the compute footprint is a growing share of the total.

Where that consumption sits with a cloud provider it is not absent from your footprint, and the data required to report it is not always easy to obtain. Establishing what your provider can give you is worth doing before you are asked for it.

What to put in place

Start by attributing compute cost to the teams and products that generate it. Most organisations pool it centrally, which removes any incentive to be efficient.

Then set an expectation that new workloads come with an estimated cost and consumption figure. Neither is difficult. Together they turn energy from something discovered in arrears into something considered at the point the decision is made.