1 editor ' s note
RACA Journal: ISSN 1812-772X
+ 27( 0) 82 560 8718
RACA Journal Publication www. refrigerationandaircon. co. za eamonn @ interactmedia. co. za
AI’ S INFRASTRUCTURE REALITY CHECK – AN ELUSIVE ROI
Artificial intelligence( AI) has become the defining technology narrative of the decade.
Hyperscalers are spending at levels previously associated with national infrastructure programmes, governments are racing to secure domestic compute capacity, and data centres are increasingly described as critical infrastructure. Beneath the enthusiasm, however, a more uncomfortable set of questions is emerging – questions with direct implications for the energy, HVAC and mechanical services sectors.
At the heart of the issue is a growing mismatch between investment, operational complexity and measurable outcomes. While AI adoption is accelerating, evidence that it delivers enterprise-level productivity gains remains limited. At the same time, the physical infrastructure required to support AI – particularly power generation and cooling – has become one of the fastest-growing sources of strain on electricity grids and industrial HVAC systems.
Despite the hype surrounding artificial intelligence, many organisations are struggling to translate that promise into meaningful productivity gains at scale. According to a Gartner ¹ survey of finance and business leaders, AI investments – across both traditional machine learning and generative AI – have produced only modest improvements in workforce output. Just 37 % of teams using traditional AI tools reported significant productivity gains, while the figure fell to 34 % for generative AI. Analysts have described this as an‘ AI productivity paradox’, where technology that appears powerful in isolated tasks fails to deliver broad organisational benefits that materially improve enterprise performance. As a result, CFOs are increasingly urging a reset in expectations around how and when AI delivers efficiency gains, alongside a greater focus on workflow redesign and process integration.
This view is reinforced by additional research. A large MIT-linked study ² found that the vast majority of generative AI pilot projects fail to deliver measurable improvements in key business outcomes such as revenue growth or cost reduction, highlighting a persistent gap between adoption and financial impact.
Nowhere is the tension between ambition and reality more visible than in electricity markets. AI data centres are extraordinarily energyintensive, with individual facilities drawing power comparable to small cities. In the US alone, projected data centre demand over the next five years is expected to exceed currently available generation capacity by a wide margin.
While multiple solutions are being proposed – including new generation, grid upgrades and alternative energy sources – most operate on long timelines and face regulatory, environmental and community resistance. In the interim, grid congestion, connection delays and rising energy prices are becoming increasingly common.
For data centre operators, this creates a fundamental paradox. Compute capacity can be built relatively quickly, but power availability cannot. Without guaranteed electricity supply, data halls
RACA Journal I April 2026 risk becoming stranded assets. For HVAC designers and mechanical contractors, this uncertainty complicates system sizing, redundancy planning and long-term service commitments.
As power density rises, cooling becomes not just an efficiency issue but a feasibility constraint. AI workloads generate sustained, high-intensity heat loads that exceed the assumptions underpinning many legacy facilities. Traditional air-cooled systems are increasingly inadequate, accelerating the shift toward liquid cooling, direct-to-chip solutions and hybrid thermal architectures.
This transition presents both opportunity and risk for the HVAC sector. Demand for advanced thermal management solutions has never been stronger, but rapid technological change increases the risk of mis-specification, stranded equipment and shortened asset lifecycles. If AI investment slows or becomes more selective, projects currently in planning may be deferred, resized or cancelled, leaving contractors and suppliers exposed to sudden shifts in demand.
The most likely outcome is not a collapse in data centre growth, but a rebalancing. Even if enterprise AI returns remain underwhelming, data centres will continue to expand due to cloud computing, digital services and data sovereignty requirements. However, the pace and composition of that growth are likely to change.
For the energy sector, this points toward greater emphasis on resilient and flexible power solutions, including on-site generation, microgrids, waste-heat recovery and load management. Reliability and efficiency will matter more than raw capacity as grid constraints intensify.
For HVAC and mechanical services, the opportunity lies in efficiency, adaptability and lifecycle performance. Clients are increasingly focused on total cost of ownership rather than headline capacity, favouring systems that reduce energy intensity, enable modular expansion and integrate with advanced monitoring platforms. Retrofitting existing facilities is also becoming a critical growth area, as operators seek to extract more compute from existing footprints rather than building entirely new sites.
The AI boom has brought unprecedented attention— and capital— to data centre infrastructure. The next phase, however, is likely to be more disciplined, shaped by physical constraints rather than technological ambition alone. Power availability, cooling efficiency and demonstrable operational returns will increasingly determine which projects proceed and which do not.
For professionals in energy and HVAC, this is not a reason for pessimism. It signals a shift from hype-driven expansion to engineering-led optimisation. In that context, some of the most important AI breakthroughs of the coming years may not come from software, but from the systems that keep the servers running. RACA
REFERENCES:
1. https:// www. gartner. com / en / newsroom / press-releases / 2025-03-25-gartnersays-cfos-should-reset-expectations-about-ais-impact-on-workforceproductivity-and-headcount?
2. https:// www. tomshardware. com / tech-industry / artificial-intelligence / 95-percentof-generative-ai-implementations-in-enterprise-have-no-measurable-impact-onp-and-l-says-mit-flawed-integration-key-reason-why-ai-projects-underperform?
www. refrigerationandaircon. co. za
1 editor ' s note