What happens to AI operations when the grid cannot carry everyone?
Picture a call centre where every agent phones in sick on the same morning with no one to pick up. That is what a major power interruption to a key data centre can do to a business that hands its frontline to AI. The software needs compute, and compute needs electricity. Cut the power at the wrong place and the operation is absent until it comes back.
In South Africa, that risk is fast moving from a theoretical scenario to a looming infrastructure challenge.
New data centres planned for Cape Town will draw more than a third (580 MW) of the city’s entire peak electricity demand (1 676 MW), all of it aimed at buildings full of servers. But Eskom’s own planning shows roughly 5 gigawatts of coal capacity retiring from 2029. The grid is set to tighten again exactly as artificial intelligence (AI) infrastructure demand arrives to compete for what is left.
Now we are in a phase where many organisations are looking at bringing their AI compute closer to home. The move is motivated by seeking things like better cost predictability and control. The “AI shortage” South Africa currently watches relates to the microchips and graphic processing units (GPUs) AI infrastructure is built on. Wait times on the newest Nvidia H100 GPUs, for instance, run past a year, and banks and insurers are racing each other to lock in orders. But bringing a GPU to a warehouse does nothing – what turns silicon into a working AI model is power, and power is the input this country has very little room to spare on. In that sense, a less obvious binding constraint compared to ‘how much’ compute an organisation can buy, is actually ‘where’ it can afford to run it – and what happens when it cannot.
Unpredictable cost hits first, since every cloud provider meters AI differently. Finance teams that spent a decade learning to forecast cloud spend relating to hyperscalers are discovering that AI throws the hard-won discipline out. The more capable the AI model, the more it gets used, increasing costs on a curve that hasn’t been drawn in the budget or even really plotted. These days a workload can double in cost between board meetings, and a CFO who cannot predict a number that large cannot control it either.
Continuity and cost are ultimately the same problem. If the grid tightens and power has to be (hypothetically) auctioned to whoever pays a premium, keeping an AI call centre or trading desk online quickly goes from being a technical hurdle to a financial one. A business relying on constant compute from a local facility will find that simply staying operational during a power squeeze can then suddenly destroy its budget, even though the reasoning was that the budget had become more stable.
The standard answer would seem to be diesel backup at facilities like that of Teraco, but the capacities in place at most operator sites aren’t readily disclosed. Running diesel-backups for such giant facilities in a single city with requirements of hundreds of megawatts is an enormously difficult task. To illustrate, bear in mind that even Eskom’s utility-scale open-cycle (diesel) turbines that augment the national grid supply have a collective capacity of less than 3 GW.
There will be terms and conditions that develop to cater to these looming (possible) scenarios, and that will add more layers of complexity to the business continuity and corporate recovery plans that motivated for hedging against global risk in the first place.
None of this is to say that local compute isn’t a solution. The lesson in this fluid, multifaceted and dynamic space comes down to what can essentially be summarised as “not putting all your eggs in one basket” – if you can help it. The thing is, many people may have not realised that the national grid constraints South Africa could face again even represented a single basket for their AI repatriation efforts. Many baskets are invisible for a long time, it turns out.
The ideal design moves the work to wherever it is best positioned to deliver at optimal ROI. That may mean agility and flexibility with regards to sovereignty, or moving to where power and price are cheapest at the time, rather than hauling data across the country to reach the compute. It’s this ‘rolling with what the world throws at you’ approach that forms the core of a future proof strategy.
Keep data where it lives; send the processing to it. Bring in compute capacity to your data from one cloud environment or hyperscaler this month and move to a different one next month to handle spikes during peak usage in a more cost effective way! Fail over from a provider that goes dark to one that is up, without shifting a terabyte to do it. That flexibility turns power from a fixed liability into a variable an organisation can manage.
