AI Expert Warns South Africa Against Pursuing Large Data Centres

South Africa is being urged to host large AI data centres, but AI implementation specialist Dr Francois du Plessis argues the country should not chase this.

Du Plessis frames his objection around three inter‑related concerns.

Why Du Plessis Opposes a Data‑Centre Rush

First, the resource load of a hyperscale facility would place heavy demand on water and electricity, resources that South Africa already finds scarce.

He emphasizes that “My view is that South Africa should not chase this,” underscoring the urgency of the resource argument.

Second, he notes that “The computing infrastructure can be situated elsewhere in the world, yet the response can still reach the user within seconds,” suggesting latency is not a binding constraint.

Third, he warns that “The cost of hosting these facilities could fall heavily on agriculture,” making the decision both economic and environmental because farming communities depend directly on natural resources.

His overall recommendation is that South Africa “use AI extensively, rely on established international data centres where appropriate, and concentrate national efforts on developing AI‑augmented industries.”

Industry Momentum and Counterarguments

Despite Du Plessis’s cautions, Cape Town approved two hyperscale facilities totalling around 174 MW in July.

Teraco is bringing its 30 MW JB5 site online to host its first AI workloads at scale.

Equinix is scaling Johannesburg 1 toward 24 MW, while MTN has named South Africa a priority market for its data‑centre platform.

Microsoft has pledged R5.4 billion by the end of 2027 on top of roughly R20.4 billion invested over the preceding four years.

The prevailing narrative is that regions with reliable energy are competing aggressively for AI infrastructure investment, treating energy security as an economic development lever.

Proponents argue that countries failing to address power constraints will miss the next wave of digital investment, and that new facilities generate demand across generation, construction, telecommunications, security and professional services.

Du Plessis counters that latency is not the binding constraint most people assume, and that the resource cost is borne locally while the compute benefit is available regardless of location.

Water, Power and Heat – The Resource Equation

Teraco’s recent data provides the first hard African data‑centre water figure reported by iAfrica, showing water usage effectiveness declining from 0.10 litres per IT kWh in 2023 to 0.04 litres in 2024 – a 60 % reduction in two years.

The company attributes the drop to closed‑loop cooling systems that do not depend on continual water replenishment.

Modern closed‑loop facilities use far less water than older evaporatively cooled sites, though water remains required for sanitation, cleaning, maintenance and landscaping.

A Schneider Electric and Microsoft life‑cycle assessment found that shifting from air to advanced liquid cooling can cut blue‑water consumption by up to 82 %.

Facilities that rely on evaporative cooling continuously lose water to the atmosphere, whereas closed‑loop, air‑cooled chillers or direct liquid cooling do not.

South Africa remains water‑scarce, with the Vaal and Orange rivers supporting urban, industrial and irrigated agricultural demand simultaneously.

Any large new allocation for a data centre would have to be assessed against existing downstream users of these rivers.

Beyond water, servers run continuously and produce substantial thermal output that must be removed without interruption.

Du Plessis cautions that waste heat and the potential formation of localized heat islands could affect surrounding communities and agricultural activity, with severity depending on the facility’s cooling design.

Heat‑island effects could further strain water resources needed for irrigation, creating a feedback loop between energy use and agricultural productivity.

These intertwined resource pressures form the core of Du Plessis’s argument against a domestic AI data‑centre rush.

Why This Matters