Everyone’s talking about artificial intelligence. And I mean everyone. The hype reminds me of the fanfare around 5G’s launch – all promise, waiting to see delivery. We built and enabled those towers everywhere, and now? Well, let’s just say the revolution was somewhat overstated.
Which brings us to the current question: are we about to make the same mistake with AI data centres?
Around the world, we’re witnessing explosive growth of AI data centres, which are fundamentally different beasts to conventional data centres. Traditional facilities are essentially sophisticated warehouses – racks, power, air conditioning, fire suppression. Nothing revolutionary.
AI data centres, however, are dramatically more power hungry. Nvidia GPUs consume staggering amounts of electricity and generate tremendous heat, requiring advanced liquid cooling alongside air conditioning. This means dependency not just on power, but on water resources, too.
So, here’s South Africa’s conundrum: we desperately need local AI infrastructure for time-critical applications, but we’re woefully short of the two things these centres demand most: reliable electricity and water. And rather than solving that fundamental problem, we’re racing to build more capacity that might be chasing hype rather than genuine demand.
Let’s examine where we stand. In South Africa, and Africa generally, we’re dramatically behind the rest of the world in data centre capacity. The US has 5 500 data centres, with 30% being AI-capable. There are only 12 000 to 13 000 data centres globally, giving you an idea of America’s dominance.
Expanding
In South Africa, we have 56 data centres with combined capacity of 350MW. Compare this to Stargate UAE’s planned 5GW AI data centre, or Colossus 2 in Memphis at similar scale. Only five local data centres are AI-capable, meaning we hold approximately 1% of global AI data centre capacity. We’re a tiny player.
Yet the market is certainly expanding. The AI data centre market in South Africa is projected to grow from US$78.9-million in 2025 to $483.5-million by 2030 – an extraordinary 43.7% annual growth rate. Africa’s per-capita data centre electricity consumption remains below 1kWh/person today, but is expected to double by decade’s end as cloud adoption and sovereign AI projects scale up.
Which begs the question: we’re obviously planning to build more, but do we actually need to? And critically – can we even power it all?
Read: South Africa’s next crisis? Being offline in an AI-driven world
Yes, we need local AI infrastructure. Many South African companies using AI would benefit enormously from local resources. We’re also serving large global companies with local presence, like Microsoft, Google and Amazon.
Financial institutions are at the forefront of AI-driven network transformations, with South African banks leveraging AI for fraud detection, customer validation and automated advisory services. PwC estimates AI could add up to R2.5-trillion to the South African economy by 2030, equivalent to nearly 10% of GDP.
For time-critical applications – financial trading where milliseconds matter, healthcare diagnostics requiring immediate analysis, autonomous systems needing split-second decisions – we genuinely cannot leverage global capacity. The latency would be unacceptable. So, we do need local infrastructure. That’s not the question.
The question is: are we building because there’s genuine demand, or simply because we can? Which companies are actually going to rent those racks?
Here’s an analogy: developers keep building giant shopping malls not because thousands of shoppers desperately need places to shop but because there’s a market for the property itself. It’s the “build it and they will come” philosophy – sometimes it works, sometimes you end up with empty retail spaces.
When building large AI data centres, operators need to make them extraordinarily attractive. Their primary “hunting ground” comprises companies already housed in other data centres, which suggests we may risk overbuilding. We could end up cannibalising existing capacity rather than serving genuinely new demand.
But here’s the real problem: these new AI data centres are voraciously power hungry. It would take three Koeberg nuclear power plants to fire up Stargate UAE. Yet both water and electricity are extremely scarce in South Africa – a reality we cannot ignore.
Consider this: Medupi power station is 4.8GW and took 14 years to build. Kusile took 17 years – the most expensive power station ever constructed globally. In its second month of operation, a large unit exploded, losing 800MW of capacity. So, if someone approached government wanting to build a Colossus 2-sized data centre, it would conceivably take Eskom decades to build sufficient capacity just to power it.
Profoundly different
There’s considerable talk about the grid being “sorted” due to recent absence of load shedding. But demand has dropped significantly because customers adopted backup power and solar generation. The grid hasn’t improved as much as demand has decreased.
The largest data centres in South Africa belong to Teraco. Its new JB4 facility has 50MW of capacity – still a baby by global standards but enormous for us. Teraco has been using grid power with serious backup, but it’s building a 120MW solar farm in the Free State to wean itself off the grid.
That’s the reality: any significant growth must be filled by renewables. There’s simply no alternative.
Read: Your Wi-Fi router is about to start watching you
So, here’s where we are: South Africa is profoundly different to Europe and the US, yet we’re pouring money into AI data centre projects while using the “shopping mall” model of stealing customers from existing facilities. That’s not necessarily catastrophic – many companies will migrate from conventional data centres to AI facilities to improve their capabilities. The market will find equilibrium eventually.
But we need to think carefully about our path forward. That’s before we factor in thermal and noise pollution these centres generate, and the fossil fuel pollution if running on Eskom power – which holds the dubious distinction of being the world’s number one sulphur dioxide producer.
We will ultimately need local AI data centres to provide low latency for critical applications. It’s a puzzle entire industries must solve collectively – balancing genuine demand against infrastructure constraints, environmental concerns against economic opportunities, present limitations against future potential.
The answer isn’t to avoid building. It’s to build strategically, retrofit intelligently and power sustainably. Otherwise we risk creating expensive monuments to hype rather than functional infrastructure for genuine innovation.
We’ve already seen how that worked out for cellular carriers and their overhyped 5G towers. Let’s not build another set of expensive monuments to promises that never quite materialised. This time, let’s match ambition with infrastructure reality – or we’ll end up with world-class data centres we literally can’t afford to switch on.
Get breaking news from TechCentral on WhatsApp. Sign up here.
- Paul Colmer is an executive member at the Wireless Access Providers’ Association, focusing on digital infrastructure and telecommunications policy in South Africa
