The Hidden Cost of AI Infrastructure: Security, Redundancy and Resilience

The economics of AI data centres extend far beyond land, chips and electricity. Investors must also price the systems that keep compute available when power, cooling, networks or physical security fail. 

AI infrastructure is physical infrastructure

 
The investment case for artificial-intelligence infrastructure is usually presented through a familiar set of variables: demand for computing capacity, access to advanced chips, contracted customers, power prices and utilisation. These factors matter, but they describe the asset in normal conditions. They do not fully capture what it costs to keep the asset operating when normal conditions disappear.
 
An AI data centre is not merely a property filled with servers. It is a tightly coupled system of electricity, cooling, water, telecommunications, cybersecurity, specialised staff, replacement parts and legal permissions. Failure in any one layer can interrupt the revenue stream even when the computing hardware itself remains intact.
 
The International Energy Agency reported in 2025 that data centres consumed about 415 terawatt-hours of electricity in 2024 and projected demand to more than double to approximately 945 TWh by 2030. It also estimated that around 20% of planned projects could face delays unless grid risks are addressed. For investors, power availability is therefore not a technical footnote; it is a development, financing and valuation risk.
 
Source: International Energy Agency, Energy and AI, 2025
 

The UAE case brings resilience into the investment model

 
OpenAI announced Stargate UAE on 22 May 2025 as a one-gigawatt cluster in Abu Dhabi, with 200 megawatts expected to become operational in 2026. The project sits within a broader UAE ambition to build five gigawatts of AI capacity and illustrates the scale at which compute, power and sovereign strategy are converging.
 
Source: OpenAI, Introducing Stargate UAE, 22 May 2025

On 11 September 2026, Reuters reported that the UAE was reconsidering the design of the wider campus following Iranian attacks on technology infrastructure in the Gulf. Options under review reportedly included distributing facilities across several sites, underground construction, blast-resistant materials and stronger air defence. Reuters could not determine the final configuration, cost or timetable. These remain reported options, not a confirmed redesign.

Source: Reuters, 11 September 2026
 
The significance is broader than one project. Security is moving from a perimeter-control expense to a design variable. If capacity must be dispersed, hardened and supported by multiple energy and network routes, resilience can alter capital expenditure, operating costs, construction sequencing and the date at which revenue begins.
 

Where the hidden costs sit

Power that is available—not merely contracted

 
A connection agreement does not guarantee timely energisation or uninterrupted supply. The investment model should distinguish grid capacity, firm generation, backup generation, fuel availability, storage and the ability to operate during peak demand. Transformers, turbines and transmission lines can have multi-year lead times. A nominally low electricity tariff may be less valuable than dependable power with credible expansion capacity.
 

Cooling and water

 
Higher-density computing produces more heat and raises the importance of cooling design. In hot, water-scarce markets, the chosen system affects capital expenditure, energy use, water consumption and operating resilience. The World Bank noted in January 2026 that more than one-third of data-centre infrastructure is already located in water-scarce areas and that warmer climates can make cooling more water-intensive. Water rights, recycled-water access and heat rejection should therefore be underwritten alongside electricity.
 
Source: World Bank, When the Cloud Meets a Thirsty World, 12 January 2026
 

Physical and cyber security

 
AI facilities may host commercially sensitive, government or strategic workloads. Protection can require controlled setbacks, surveillance, access management, hardened structures, protected fuel and cooling systems, secure supply chains and specialised response teams. Cybersecurity must cover both information technology and operational technology: an attack on building controls, power management or cooling can stop compute without breaching a customer database.
 

Redundancy and geographic failover

 
Duplicate generators, network routes and cooling systems improve availability only when they do not share the same point of failure. Two fibre links routed through one exchange, or two sites supplied by one substation, create the appearance of redundancy without independence. Geographic distribution can reduce concentration risk, but it adds land, connectivity, replication, synchronisation and operating costs.
 

Insurance, contracts and recovery capacity

 
Insurance may exclude war, cyber events, grid failure or consequential loss, while business-interruption cover may not match the time required to replace specialised equipment. Customer contracts can impose service credits or termination rights after outages. The economic model should include deductibles, coverage limits, restoration periods, spare-parts access and the possibility that an event is operationally damaging but contractually uninsured.
 

Redundancy can create its own risk

 
More resilience is not automatically better. Every additional system requires capital, maintenance, testing and people able to operate it. Software-based failover can improve availability while increasing architectural complexity. Uptime Institute's 2025 outage analysis found that power remained the leading cause of impactful outages, while distributed resilience tools could also introduce new risks and blur responsibility for failures. Nearly 40% of surveyed organisations had experienced a major human-error-related outage over the preceding three years.
 
Source: Uptime Institute, Annual Outage Analysis 2025, 6 May 2025
 
The objective is not maximum duplication. It is proportionate resilience: identify critical workloads, determine the maximum tolerable interruption and design independent recovery capacity for those functions. A facility that promises exceptional uptime but cannot demonstrate live failover tests may be less resilient than a simpler system with clear procedures and accountable operators.
 

What investors should ask before committing capital

 
·        What proportion of projected capacity has secured land, power, permits, chips and contracted demand?
 
·        Which power, cooling, network and security systems represent single points of failure?
 
·        Are backup systems genuinely independent, and when were they last tested under load?
 
·        How do heat, water scarcity, dust and local grid conditions affect performance and cost?
 
·        What additional capital is required for hardening, geographic dispersion and recovery sites?
 
·        Who bears cost overruns, energisation delays and equipment-replacement risk?
 
·        What events are excluded from insurance, and how long does business-interruption cover last?
 
·        Do customer contracts protect revenue during force majeure, or permit credits and termination?
 
·        How quickly could operations recover if staff, telecommunications or imported parts were unavailable?
 
·        Does the return case remain credible after including resilience capital and realistic utilisation ramp-up?
 

Portfolio implications extend beyond data-centre equity

 
The resilience question reaches every layer of the capital structure. Equity absorbs delays and overruns. Lenders depend on completion tests, contracted revenues and insurance proceeds. Property investors face questions about specialised buildings and alternative use. Utilities must fund generation and grid reinforcement before demand is certain. Private funds may carry concentration through several portfolio companies that depend on the same operator, power zone or cloud platform.
 
Family offices should therefore aggregate AI-infrastructure exposure across listed technology holdings, private funds, real estate, utilities, credit and operating businesses. Investments that appear diversified by legal form can remain connected to the same physical bottlenecks.
 

Hauberk View

 
AI infrastructure can support long-duration economic growth, sovereign capability and new investment opportunities across the Gulf. But demand for compute does not by itself guarantee attractive returns. The decisive question is whether the project can convert enormous fixed investment into dependable, insurable and contractually protected cash flow.
 
Security, redundancy and resilience should be priced before capital is committed. If they are added after a threat emerges, they can alter the design, delay revenue and weaken returns. If they are omitted, the asset may carry a level of interruption risk that its headline valuation never reflected.
 
The best AI-infrastructure investment may not be the project with the largest announced capacity. It may be the one whose power, cooling, connectivity, governance and recovery arrangements have been tested as rigorously as its technology.
 
About Hauberk Capital: Hauberk Capital provides strategic perspectives focused on capital preservation, long-term wealth management and the institutionalisation of family wealth.
 
This article is provided for informational and educational purposes only. It does not constitute investment, legal, tax, insurance, engineering or cybersecurity advice, or a recommendation to undertake any transaction.