# MENA AI cooling: why one efficiency score is not enough

An AI facility can have a good power-efficiency score and still be a poor fit for its location. Cooling may save electricity while consuming scarce water. A clean-energy claim may describe the national grid rather than the supply available to a particular site.

The World Economic Forum and SRMG Think put these trade-offs at the centre of [a report published on 14 October](https://www.weforum.org/publications/powering-ai-at-scale-optimizing-hyperscale-data-centre-infrastructure-in-mena/). Its useful contribution is to connect AI infrastructure decisions to electricity, cooling and water planning. For Saudi and UAE buyers, that suggests a more demanding comparison than choosing the lowest advertised power usage effectiveness, or PUE.

The following is a proposed procurement approach, not a ranking of facilities or a claim that any operator has passed these checks.

*Cover: original SultanByte artwork showing electricity, heat rejection and water as separate parts of an AI facility's operating boundary.*

## A cooling headline needs a denominator

The report's cooling discussion illustrates why buyers should follow a number back to its source. Its executive summary says more than 40% of electricity consumption at regional data centres is dedicated to cooling. The cited [US Department of Energy announcement](https://www.energy.gov/articles/doe-announces-40-million-more-efficient-cooling-data-centers), published in May 2023, instead says cooling can account for **up to 40%** of data-centre energy usage overall. It does not establish a measured MENA-wide average.

Those formulations are not equivalent. Neither tells a buyer how a proposed Riyadh or Abu Dhabi hall will perform.

The [IEA's Energy and AI analysis](https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai) makes the variation clearer: its equipment breakdown puts cooling at about 7% for efficient hyperscale facilities and above 30% for less-efficient enterprise facilities. These are different facility categories, not a Saudi-versus-UAE comparison. The useful lesson is to specify the building, operating conditions and measurement boundary before comparing percentages.

Request cooling electricity separately from IT electricity, and ask what the reported period includes. A design simulation, a mild-weather test and a full operating year answer different questions. Keep all three labelled if the supplier provides them.

## PUE measures overhead, not useful AI output

PUE compares total facility energy with IT equipment energy. It helps expose overhead, but it does not say whether the servers are doing useful work, whether the electricity is low-carbon, or how much water the facility consumes.

[Google's measurement explanation](https://datacenters.google/efficiency/) is useful here because it describes a trailing twelve-month approach, all seasons and included overhead sources. It also distinguishes mature campuses with enough operating data from newer facilities. That is a reporting model worth examining, not a target every Gulf site should copy from Google's global fleet.

For a proposed Saudi or UAE deployment, ask the supplier to state which meters define its PUE calculation. Then request the matching IT load, reporting period and seasonal behaviour. If one bidder quotes a design value and another quotes measured operation at partial occupancy, do not put those values in an unqualified league table.

Application teams need another measure alongside facility efficiency: energy for an agreed amount of useful work. Define the model, hardware, input and output lengths, latency constraint and quality threshold. Changing those conditions can make an apparent improvement meaningless. A faster system that fails the Arabic task is not a successful efficiency result.

## Follow the heat beyond the chip

“Liquid cooled” does not describe the complete cooling system. The chip-to-fluid loop is only part of the route by which heat leaves a facility. Ask what ultimately rejects that heat and which parts consume water.

The [full WEF–SRMG report](https://reports.weforum.org/docs/WEF_Powering_AI_at_Scale_Optimizing_Hyperscale_Data_Centre_Infrastructure_in_MENA_2026.pdf) recommends location- and workload-specific cooling, including hybrid arrangements, closed loops and treated wastewater where appropriate. Treat these as options to investigate, not a universal recipe. Its recommendations are a planning framework, not evidence that a named installation achieves the promised result.

Our proposed comparison should include annual water consumption, peak-day demand, water source and the assumptions behind both. Ask whether a “closed-loop” claim covers only the internal circuit or the entire heat-rejection arrangement. Request the electricity penalty, if any, when the system switches to a lower-water operating mode.

For a wastewater proposal, require an actual supply arrangement: delivery route, quality specification, minimum availability and the response to an interruption. A nearby treatment plant on a map does not settle those questions.

![AI cooling comparison: define the facility boundary; separate PUE, water and useful-work measures; test summer and interruption conditions; and verify the site's own Saudi or UAE power and water arrangements.](https://cdn.hashnode.com/uploads/covers/60ecf4a0fc37a15ec15655e8/806b5589-1591-4627-9b1d-c36d88436821.png align="center")

*Original infographic: SultanByte. Proposed comparison framework informed by WEF–SRMG Think, IEA, DOE and Google's measurement guidance. October 2026; no measured facility results are shown.*

## Saudi Arabia and the UAE need different evidence

Saudi Arabia's renewable build-out provides useful context, but a national project cannot be treated as a private power entitlement. [ACWA Power's Shuaibah project record](https://acwapower.com/en/what-we-do/projects/shuaibah-pv-ipp/) identifies two solar plants in Makkah Province and names Saudi Power Procurement Company as the offtaker.

For a Saudi facility, ask which supply agreement actually serves the site, what backs it when solar output is unavailable and which connection or expansion milestones remain outstanding. Do not turn a solar plant's capacity or contract into a claim that a particular AI hall has continuous renewable supply. The cited project record does not establish that relationship.

The UAE has a different operating reference. [ENEC's Barakah overview](https://www.enec.ae/barakah-plant/) describes an operating nuclear fleet supplying the country's electricity system. That supports the existence of a substantial nuclear contribution, not a conclusion that every UAE data centre receives exclusively nuclear electricity.

For a UAE bid, separate national generation context from the facility's contractual energy claim, physical connection and backup operation. Ask how the supplier accounts for the electricity it consumes and what evidence supports any hourly or annual clean-energy statement. Use the same water questions in both countries, but obtain site-specific answers rather than assuming identical utility arrangements.

## Buy a tested operating range

Procurement can turn this into a short, enforceable evidence schedule. Require the supplier to name its measurement boundary, distinguish designed from measured values and disclose the conditions under which each commitment applies. Put water-source changes and cooling-mode changes into the reporting requirements, not merely the sales presentation.

Engineering should define an acceptance exercise with the operator. Include representative AI demand, sustained high load, partial load and the facility's stated summer design conditions. Ask for an approved way to validate interruption behaviour without endangering live equipment. Record who decides to reduce computing load when power or cooling headroom falls.

An application team can also identify work that can move in time. Batch evaluation or a deferrable training run may tolerate scheduling changes; an interactive customer service usually has a different latency obligation. Agree those limits explicitly before treating workload flexibility as available capacity.

The October report makes a useful case for planning energy and computing together. Buyers can act on that without accepting every regional headline number. Compare facilities using matched boundaries and operating conditions, then contract for the evidence and response procedures that matter at the chosen site.