RIGHT-SIZING AI INFRASTRUCTURE
RACA Journal I July 2026 www. refrigerationandaircon. co. za 21
DATA CENTRE FEATURE SERIES
All images by © RACA Journal
RIGHT-SIZING AI INFRASTRUCTURE
By Eamonn Ryan
Why‘ pay-as-you-grow’ is becoming the data centre design philosophy.
As organisations across Africa explore artificial intelligence
( AI) deployments, one of the biggest challenges facing the data centre sector is determining how much infrastructure to build – and when. Overbuilding risks massive capital expenditure on unused capacity, while underbuilding can leave operators unable to meet rapidly growing demand.
These issues were explored in depth during a panel discussion at the seminar AI Unlocked: What‘ AI-ready’ means for 2026 decisions, featuring:
• Marc Matthews, director of engineering and head of PMO at Open Access Data Centres( OADC)
• Wojtek Piorko, managing director for Africa at Vertiv
• Kevin Subramani, director of industry for mission critical facilities and aviation at Atana
The discussion highlighted the growing importance of scalable infrastructure, flexible design and realistic planning horizons in an AI-driven data economy.
THE DANGER OF BUILDING TOO BIG According to Matthews, one of the most common mistakes is building infrastructure far larger than the market requires. He illustrated this with a case study from Nigeria. OADC initially planned a 24MW data centre based partly on competitor activity and expectations of future demand.
“ On paper the project looked fantastic. But once we started procurement and the numbers came back, our eyes started watering,” Matthews said.
( Left to right): Moderator Dion Chang, Marc Matthews, director of engineering and head of PMO at Open Access Data Centres( OADC); Kevin Subramani, director of industry for mission critical facilities and aviation at Atana; Wojtek Piorko, managing director for Africa at Vertiv.
When the full cost profile became clear, the company realised the project had been significantly over-capitalised. After consulting with customers and partners, OADC adopted a much different strategy.“ We spoke to one of our tier-one clients and asked where they thought the Nigerian market was going. Their view was that they wouldn’ t go beyond one megawatt at this stage,” Matthews explained.
The result was a shift to a‘ pay-as-you-grow’ model, in which capacity is expanded in phases as demand materialises rather than being built all at once.
DESIGNING FOR CHANGE, NOT CERTAINTY Subramani argued that the rapid evolution of AI technologies makes traditional long-term infrastructure planning increasingly difficult.“ The risk with overbuilding or underbuilding is that you’ re planning for the wrong future,” he said.
AI workloads, chip technologies and compute requirements are changing at such speed that designing a facility for a fixed future state is becoming unrealistic.
Instead of attempting to‘ future-proof’ data centres, Subramani suggested designers should focus on‘ future-enabling’ them.“ This means building facilities that can adapt as workloads change, as technologies evolve, and as new requirements emerge. You invest what you need now and scale later.”
This typically involves modular designs, flexible power infrastructure and cooling systems capable of supporting higher rack densities over time.
WHERE DATA CENTRES ARE OFTEN MISDESIGNED Piorko added that early-stage designs frequently include unnecessary levels of redundancy and capacity.“ Power infrastructure is often overdesigned. You’ ll see requests for additional UPS systems, more generators or higher tiers of redundancy than the workload actually requires,” he said.
At the same time, other critical elements are sometimes under-designed – connectivity is one of them.“ With the volume of data and compute today, connectivity becomes absolutely critical. But it’ s sometimes overlooked in the early planning stages,” Piorko said.
Cooling infrastructure is another area where designs can fall short. As compute densities increase – particularly with AI workloads – traditional aircooling systems may struggle to keep up.“ Liquid cooling is coming whether we like it or not. That means there will be water in the white space, and facilities must be designed for that from the beginning,” Piorko said.
Data artificial centres intelligence
RACA Journal I July 2026 www. refrigerationandaircon. co. za 21