Hardware Guides

Cloud vs on-premise in South Africa: the break-even nobody can source

PDSThe PDS Team·2026-08-19·9 min read
An on-premise server standing in a lamp-lit South African office at dusk, with distant city lights beyond the window

Every article quoting you a break-even for owning a server versus renting cloud was published by somebody selling one of the two. Here is what the law, the price lists and the local grid actually say.

Ask a firm where its next server should live and you will get a confident answer in seconds. Ask where that answer came from and the trail goes cold. Every article quoting a break-even point — own it after three years, rent it and save thirty per cent — was published by somebody selling one of the two options. That is not a conspiracy. It is what happens when the only people motivated to do the arithmetic are the people with something to sell.

This article makes the case that for some firms a local server is the better long-term choice — honestly, which means naming where it falls apart and recommending the opposite where the opposite is right.

Dollar figures are converted at R16.1022 to the dollar, the rate on 20 August 2026 (Trading Economics). The original figures stay in brackets so you can check every one against its source.

TL;DR: There is no publicly sourceable break-even between owning a server and renting cloud capacity — every published figure traces back to a company selling one side. What is documented: SARS expects electronic tax records to be kept in South Africa unless a senior official authorises otherwise, POPIA does not require that, and AWS charges about R2.48 per GB ($0.154) to move data out of its Cape Town region to the internet, for the first 10TB a month.

Why nobody can tell you the cloud vs on-premise break-even point

The crossover between owning and renting is not publicly knowable. Not disputed — unavailable. Search for it and you get figures from hardware vendors, cloud vendors, and consultancies that sell migrations. No neutral party publishes the model underneath.

Take the input that matters most: how long a server lasts. Amazon changed its own answer twice in two years. Effective 1 January 2024 it lengthened server life from five years to six. Effective 1 January 2025 it shortened a subset of servers and networking equipment back to five, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning” (Amazon FY2025 Form 10-K). The reversal cut Amazon’s own reported profit by about R16.1 billion ($1.0 billion).

Microsoft, filing the same kind of document, books servers and network equipment at two to six years (Microsoft FY2026 Form 10-K). Two of the largest hardware buyers on earth, both audited, and they disagree. If the refresh horizon is unsettled at that scale, a confident five-year payback figure for a firm in Gqeberha is arithmetic performed on a guess.

What “it depends on utilisation” is hiding

The stock answer is that owning wins once the machine is busy enough. True, and useless without a threshold. The Uptime Institute puts the principle plainly — self-build economics depend on keeping capacity reasonably full (Global Data Center Survey 2025). What nobody independent publishes is the number where the lines cross. Anyone quoting you one either modelled it privately or took it from a vendor.

What South African law actually requires

Here the usual pitch has it backwards. POPIA does not require your data to stay in South Africa. Section 72 expressly permits transferring personal information across the border where one of five conditions is met (Protection of Personal Information Act 4 of 2013). The government’s own National Policy on Data and Cloud goes further, stating that POPIA governs those transfers and that restricting data flows harms business (Gazette No. 50741). If somebody sells you a local server “for POPIA compliance”, they have not read it.

There is a real residency rule, and it comes from SARS. Records kept in electronic form are to be kept at a place physically located in South Africa (Government Notice 787), and SARS still operates that requirement through its published retention policy (GEN-GEN-23-POL02).

It is a default, not a ban

This gets flattened constantly. You are not forbidden from using an overseas server. A senior SARS official may authorise records kept outside South Africa where the system stays accessible from your physical address here and the location does not impede access. Offshore hosting is a permission you apply for, not a closed door. Firms that never apply are simply taking the default.

Retention stretches the decision out. FICA obliges accountable institutions to keep records for at least five years (Financial Intelligence Centre Act 38 of 2001) — which can outlast the server holding them, on either hyperscaler’s own accounting above. Our guide to FICA record-keeping obligations covers the practicalities, and the POPIA guide does the same for legal practices.

One thing neither option changes: the duty stays yours. POPIA places the security obligation on your firm regardless of where data sits, and where an operator processes on your behalf it requires a written contract binding them to the same measures (sections 19 and 21). Breach notifications to the Information Regulator reached 2 374 in 2024/25, with the monthly rate rising a further 40% in 2025/26 (ITWeb). On-premises does not make you compliant. It shortens the chain you have to explain.

The costs that are knowable

You can’t get a break-even. You can get real inputs, and three are published.

Egress. Moving data out of AWS’s Cape Town region to the internet costs about R2.48 per GB ($0.154) for the first 10TB a month, per AWS’s own machine-readable price list (AWS pricing API, af-south-1). Data that never leaves the building doesn’t cost anything to move.

Power. A single top-end professional workstation GPU is rated to draw up to 600W (NVIDIA) — a running cost you own outright, and a rising one. NERSA approved an 8.76% average increase for Eskom’s direct customers from April 2026 (Engineering News).

Refresh risk. Amazon’s reversal is the cautionary input. Whatever life you assume, assume it might shorten.

The one repatriation figure worth quoting, and why to distrust it

37signals reported cutting roughly R32 million ($2 million) a year from its cloud bill after spending about R11.3 million ($700,000) on hardware (The Register). Treat it carefully. The numbers are self-reported by a company that has built a public identity around leaving the cloud, and its workload looks nothing like a professional firm’s. It proves the move can pay for somebody. It does not prove it pays for you.

Where the cloud genuinely wins

The market is not fleeing. Enterprise cloud infrastructure spending reached about R2.3 trillion ($143 billion) in a single quarter, growing 43% year on year (Synergy Research Group). Any article implying everyone is walking away is selling something.

The sharpest argument against owning is one the providers make themselves. AWS discounts sustained, committed usage by up to 72% against on-demand pricing (AWS Savings Plans). The naive case for owning — our workload runs constantly, so renting is wasteful — assumes you would pay the on-demand rate. You would not. Cloud is already priced for the steady workload, and that removes most of the easy arithmetic. Lock-in has loosened too: AWS now waives data-transfer-out charges for customers moving off the platform, on request and subject to approval (AWS News Blog).

So rent when demand is spiky, when the horizon is short, when you are still working out what the workload is, or when you want current hardware without a capital purchase. For those cases our hosting and cloud services are the right answer, and we will say so.

The Eastern Cape reality

The standard objection to on-premises hardware here is power, and it has genuinely weakened. By 4 August 2026 the country had gone 441 consecutive days without load shedding, per figures Eskom reported through the state news agency (SAnews). For how fast that shifted: 749GWh was shed in the first half of 2025, itself down 82% year on year (Engineering News, citing CSIR). Both figures originate with Eskom, which benefits from them being believed.

But national grid stability is not local supply. NERSA’s inspection of Nelson Mandela Bay’s electricity network found systemic infrastructure failure (Daily Maverick), and by May 2026 the metro was battling more than 1 100 faults across its low-voltage and overhead networks (The Herald). A server in a Gqeberha office doesn’t care that the national grid is holding. It cares about the feeder on its street.

Connectivity cuts both ways. The 2Africa subsea cable landed at a Port Elizabeth station, so the region is no longer at the end of the line (Submarine Networks). Against that, South Africa’s fixed broadband ranking fell from 102nd to 115th of 152 countries in a year (ICASA). And nothing is truly local either way: Azure’s South African capacity sits in Johannesburg (Microsoft), as does Google Cloud’s only African region (Google Cloud).

Distance costs milliseconds. Latency to centralised cloud AI services starts around 48 milliseconds, and a local model can answer faster to first token — though the same researchers are explicit that local inference is not a universal answer (Jang and Morabito, 2025). Speed is also not the same question as usefulness — where AI genuinely helps legal research is worth settling before you buy hardware to run it faster. Worth knowing before you chase it: in one study, answers returned after two seconds were rated less thoughtful than ones that took nine or twenty (Tan et al., 2026).

How to decide without a break-even

Since the number does not exist, decide on structure. Five questions settle most cases.

Does a regulator dictate where the records live? For an accounting practice holding electronic tax records, the SARS default points onshore and offshore needs an application. That is the clearest signal on this page, and it has nothing to do with cost.

Is the workload steady or spiky? Spiky rents well. Steady is worth pricing both ways — against committed-use discounts, not on-demand.

How much data leaves each month? Multiply it by the egress rate above. For most firms the answer is small; for firms shifting large document sets, it is not.

What happens at 02:00 when it fails? Renting buys somebody else’s night shift — the same trade-off we set out in managed IT versus break-fix. Owning means the machine is yours and so is the fault — which is why we supply, install and then keep running the servers installed and maintained on your own premises, rather than dropping a box and leaving.

What is your honest refresh horizon? If two hyperscalers cannot agree, treat your own assumption as a range and check the decision still holds at the pessimistic end.

Frequently asked questions

Does POPIA require my data to stay in South Africa?

No. Section 72 expressly permits transferring personal information outside the Republic where one of five conditions is met (POPIA). The security duty stays with your firm either way, so the question is whether the arrangement is documented and defensible.

Can I keep tax records on an overseas server?

The default is that electronic records are kept at a place physically located in South Africa, but a senior SARS official may authorise an offshore location where the system remains accessible from your South African address (SARS GEN-GEN-23-POL02). It is an application, not a prohibition.

Is load shedding still a reason to avoid on-premises hardware?

Nationally the argument has weakened — 441 consecutive days without load shedding to August 2026 (SAnews). Locally it has not: Nelson Mandela Bay was carrying more than 1 100 network faults in May 2026 (The Herald). Budget for local supply, not national headlines.

Is owning always cheaper over the long term?

No, and anyone claiming otherwise is guessing. Owning depends on keeping the hardware reasonably full (Uptime Institute), and providers already discount sustained committed usage by up to 72% (AWS). The steady-workload argument is weaker than it sounds.

The short version

Buy a local server for what it does for custody, latency and control of your own refresh cycle — and because a regulator may point that way for your records. Do not buy one because somebody showed you a payback chart, because the chart is not sourced and the person holding it is selling. Rent when the work is unpredictable or the horizon is short. Both are legitimate, and a firm that understands why it chose one will handle the next five years better than a firm that was sold a number. If you want that worked through against your actual workload and building, talk to us — we will tell you when renting is the better answer.

Let’s talk about your IT needs.

Connect with a principal architect to discuss your firm’s technology, security, and compliance — in the context of the Eastern Cape’s regulatory landscape.

No obligation. A response from a principal analyst, usually within one business day.