FIELD NOTE
Demand is growing in more than one direction
The IEA reports that data-centre electricity use rose 17% in 2025, with AI-focused centres growing faster. It expects global data-centre electricity consumption to double by 2030, while AI-focused consumption could triple.
The same report points to tight supply chains and bottlenecks in turbines, transformers, chips, IT components, and approvals. More compute is therefore a planning problem involving hardware, grid connections, construction, and time.
FIELD NOTE
Efficiency is a system metric
A smaller model, better retrieval, shorter context, or fewer repeated calls can reduce cost and energy without reducing the quality of a business outcome. The right measure is not only tokens or benchmark performance; it is useful work per unit of compute.
- Choose the smallest model that meets the task
- Cache stable outputs
- Route simple work to efficient models
- Measure retries and wasted calls
- Connect energy and cost to business outcomes
FIELD NOTE
The operating layer is the advantage
As AI workloads become more varied, teams need to know which tasks are running, what capacity they consume, and whether the result is worth the load. That view can sit across product, finance, infrastructure, and sustainability teams.
The future is not less AI. It is better-directed AI: more useful work, fewer invisible loops, and infrastructure decisions that remain connected to the people and operations they serve.
DIRECT ANSWERS
Questions operators ask
Does AI efficiency mainly mean choosing a smaller model?+
Model choice helps, but prompt design, retrieval, caching, routing, retries, data quality, and workflow design often determine the larger waste.
What should a company measure?+
Measure cost, latency, energy proxies, retries, human rework, and the business outcome for each meaningful AI workflow.
SOURCE LEDGER
Primary sources
Official material is linked directly. Claims are paraphrased and checked against the source status available on 2026-08-13.FRICTION EDITORIAL CONTROL
Original analysis. Visible limitations. No invented certainty.Prepared by Friction Research and reviewed against primary sources. This material is general information, not legal, tax, financial, or regulatory advice. Requirements can change; verify material decisions with the relevant authority or a qualified adviser.Read our editorial policyREADER EXCHANGE
Add to the conversation.
Name and email are required. Suspicious or abusive comments are held before publication.
Loading conversation.