AI data-centre power demand: what compute buyers should ask
Power and delivery constraints deserve a place beside GPU specifications in a capacity decision.
· Arvica Cloud · Analysis & buying guide
The observed trend
The IEA reports that global data-centre electricity demand grew 17% in 2025. This is a global sector measure, not a forecast of GPU rental prices in any particular country. It provides context for why energy and infrastructure readiness deserve attention when planning AI capacity. [1]
Arvica analysis: capacity is more than a chip
For a buyer, a server specification and a usable deployment date answer different questions. Our procurement approach is to ask whether the proposed site can deliver the complete configuration within the required window, including network and storage. Do not infer ready-to-use capacity from a hardware announcement or a general statement about a region. Ask for the operational readiness and acceptance milestones for the quoted order.
Questions for a reserved deployment
Specify your earliest useful start date and latest acceptable delivery date. Clarify what happens if only part of a cluster is ready. Ask whether maintenance windows, network dependencies or location substitutions affect the workload. If energy or sustainability criteria are part of your procurement policy, request site-specific evidence and its reporting period rather than relying on a broad marketing label.
Build options without losing requirements
Separate mandatory constraints from preferences. Data residency may be mandatory, while a preferred city might be negotiable after a network test. Prepare a smaller initial deployment and a scale-up plan if your application supports it. Alternative regions are only useful when the data, latency and contractual requirements still hold. Arvica can review those constraints in a capacity enquiry; no regional price increase or guaranteed supply is predicted in this analysis.
Sources
Reviewed: 2026-09-11
Apply this to your workload
Further reading
How much GPU memory does a 70B model really need? →
Your GPU is busy waiting: investigate the pipeline before renting more →
Eight GPUs are not automatically eight times faster →
Before the long reservation: a GPU server acceptance checklist →
AI inference economics: why the GPU-hour is only the starting point →
B300 and Rubin: plan the workload before chasing the roadmap →
Reserved vs on-demand GPUs: calculate the utilisation break-even →