Separate a forecast from your facility
Lawrence Berkeley National Laboratory estimates that U.S. data centers used about 176 TWh in 2023, around 4.4% of national electricity. Its 2025 update presents a scenario in which the share could reach 11.8% by 2030, with a range of 9.5% to 15.3%. These are national estimates, not a forecast for an individual building. A purchasing team should start with metered power, rack capacity, cooling headroom and the actual workload profile at its own site.
Compare useful work per constrained resource
For servers, record measured throughput and latency at the intended model size, typical and peak power draw, accelerator utilization, memory capacity and serviceability. For storage, compare usable capacity after protection overhead, sustained bandwidth, latency, endurance, recovery time and expansion limits. An inexpensive unit can be costly if it forces additional racks, cooling or downtime. Evaluate complete configurations under the same workload rather than comparing isolated component specifications.
Distinguish training, inference and ordinary data processing
An overnight training batch, employee knowledge search and real-time customer-service inference place different demands on infrastructure. In a planning example rather than a customer measurement, knowledge search can depend on document processing, vector queries and model responses. More accelerator compute alone may not shorten waiting time. Business owners should define users, concurrency, input length, output quality, response time and update frequency before engineers identify compute, memory, storage and network constraints.
Test low demand and peaks, not just ideal full utilization. Identify work that can queue, work that must remain available and work existing resources can absorb. When demand is uncertain, an expandable pilot provides evidence before committing to the most optimistic forecast.
Run comparable tests on every candidate
Fix model or software version, dataset, precision, concurrency, warm-up and duration across candidates, and retain failed requests and logs. If a supplier changes quantization or shortens inputs, disclose the quality effect rather than compare throughput alone. Consider business-relevant slow requests and peak behavior, not only average latency, and agree how statistics are calculated.
The test lead retains parameters, raw results and environment details; business staff assess output quality; infrastructure staff assess resource use and stability. Use authorized, appropriately processed data, not sensitive production information in an unapproved demonstration environment. Unreproducible demo results are leads, not procurement commitments.
Measure power and useful output within consistent boundaries
State whether power covers an accelerator, a server, a rack or a facility including cooling. Do not divide figures from inconsistent boundaries. Record idle, typical and peak behavior, measurement periods and meter locations. Energy per successful request or completed task should exclude failed output and hold quality constant, rather than reward lower quality with apparent efficiency.
Facilities staff should verify electrical capacity, redundancy, rack loading, airflow or liquid-cooling requirements and service space. Rated equipment power is not an annual average, and a short test average cannot replace a capacity safety assessment. Qualified specialists should design and accept electrical or cooling changes; server specifications alone cannot establish facility readiness.
Plan storage from ingestion through disposal
Estimate raw data, training or indexing intermediates, models, logs, snapshots and backups separately. Deduct protection and reserve overhead from advertised capacity; validate compression and deduplication with representative data. Examine small files, random queries and sequential reads separately instead of treating one large-file copy as representative.
Test service behavior during a node failure, rebuild or backup and record effects on normal tasks. Data owners confirm retention and deletion requirements; storage owners explain expansion and compatibility limits. Procurement should distinguish immediately usable capacity from capacity requiring later licenses, hardware or downtime, exposing gaps in the quoted scope before launch.
Use complete costs and explicit expansion triggers
Include networking, racks, electrical and cooling changes, licenses, support, spares, staff and migration downtime alongside hardware. Compare in-house and external services using the same demand and period, and state assumptions for transfers, idle resources and exit costs. Estimates support decisions; they are not guaranteed savings.
Expansion triggers might include sustained queuing, declining capacity headroom or insufficient maintenance windows, with thresholds based on local evidence. For normal and high growth, document lead times, installation requirements and fallback options. Confirm power and delivery constraints before capacity runs out, and do not count unapproved facility expansion as available resources.
Accept an operable workload
Acceptance should retain evidence for performance, quality, power, stability and recovery, and verify handover of monitoring, alerts, escalation and maintenance instructions. Business owners accept results, technical owners accept operating conditions, and facilities owners confirm environmental boundaries. Assign owners and deadlines to unfinished conditions.
National reports explain demand and power constraints; they cannot select equipment or guarantee returns for an enterprise. These comparison methods are JOTOglobal recommendations. '2025 Update' is the report's version name; its official page lists publication in June 2026. Keep the version year distinct from its release date.
Design the expansion path before purchase
List available electrical feeds, power distribution, network ports, floor space, cooling and lead times. Test at least a normal-demand and a high-growth scenario; document which assumptions would trigger another purchase or a site upgrade. Include monitoring and a retirement plan so power and capacity figures can be checked against actual operation. The global or national reports explain why the constraint matters; local measurements decide which architecture is suitable.
Sources
United States Data Center Energy Usage Report: 2025 Update
https://datacenters.lbl.gov/publications/united-states-data-center-energy-2025
2024 United States Data Center Energy Usage Report
https://seta.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report
