Cloud spending is going up. According to ‘PwC’s 2025 Africa Cloud Business Survey’, 88% of organisations on the continent plan to increase their cloud budgets this year, yet many still can’t tell you exactly where that money is going. That’s the paradox of modern cloud economics – the more you spend, the harder it gets to see what you’re actually paying for. At the same time, Africa’s cloud maturity has accelerated. The PwC survey found that over 86% of organisations reported medium or high cloud maturity in 2025, up from 61% just two years earlier. But what this progress reveals is a new set of problems. The challenge is no longer getting into the cloud; it’s figuring out how to stop haemorrhaging money once you’re in. And while the usual cloud spend suspects – idle resources, overprovisioning, shadow IT – are well-documented, the real waste isn’t always the obvious waste. It’s quieter, more structural and often a lot harder to fix.
When you rent compute from a hyperscaler, you’re not getting a bespoke environment. You’re choosing from a menu: small, medium, large. The problem is that most real-world workloads don’t map cleanly onto those options. “If you need more than a small but don’t quite fill a medium, you have to take the medium,” says Marco Vieira, solutions architect at Nutanix. “You’re paying for resources you’re never going to use.” At a small scale, that gap is tolerable. But once you’re running 50, 80 or 100 virtual machines, that wasted capacity on every single instance starts adding up. This is when “micro waste” becomes a serious cost burden.
If infrastructure is permanently provisioned for those peaks, inefficiency becomes embedded.
Ian Engelbrecht, Veeam
Micro waste is a new kind of problem for organisations. Deleting orphaned resources and switching off idle instances are both important, but they’re reactive fixes. Micro waste, says Vieira, is embedded in the architecture from the outset. It’s something you don’t see on a dashboard; it’s just quietly built into every invoice. The answer, he says, is to right-size the total resource pool rather than trying to fit individual workloads into predefined instance types. “We can eliminate that micro waste by matching the actual resource requirement, not just picking the closest available size,” he says, adding that this is a more precise approach to provisioning, and one that pays off most at scale.
There’s a related cloud cost problem that’s equally common and equally invisible: peak provisioning. Businesses design workloads for the worst-case scenario. End-of-month payroll runs, seasonal retail spikes, overnight batch jobs – the moments define the infrastructure, even if they represent only a fraction of actual operating time. “But for 25 days of the month, it’s running at half that capacity,” says Vieira. “You’re paying for the peak, continuously, whether you need it or not.” It’s a point that Ian Engelbrecht, field CTO at Veeam Software EMEA, echoes. Teams prepare for peaks, rather than patterns.
“If infrastructure is permanently provisioned for those peaks, inefficiency becomes embedded,” he says. It sounds obvious when you think about it, but the fix is dynamic rightsizing and it’s rarely done properly. This means continuously adjusting resource allocation to match actual usage instead of locking in worst case capacity forever. Cloud platforms support this natively through tools like AWS Auto Scaling, Azure Monitor and Google Cloud’s Recommender, yet they go largely unused.
NOW YOU’RE SPEAKING MY LANGUAGE
Cloud optimisation rarely comes from one big decision. Duolingo is a great example of how it actually works in practice. Back in 2024, the language learning app, which has more than 500mn users, asked its engineering team to reduce AWS spend without breaking anything. The cloud bill was high and the margin for error was slim. What the team found when they started looking was, in hindsight, predictable: storage no one needed, instances bigger than the workloads required, and databases that had become costly because no one had reviewed them for a while. Working through it methodically, they migrated caches and databases onto more cost-efficient processor architecture, rethought their EC2 instance pool and moved individual microservices, including the Python monolith that powers much of the core product, onto better-suited infrastructure. “At first, we were a little scared,” says Jean-Paul Bonny, senior software engineer at Duolingo, “but that just speaks to the efficiency of moving to better architecture.” The savings added up across the stack, with a 25% reduction in caching costs, a 50% drop in database spend within three months and a 10% reduction in compute costs, all without changing a single line of application code.
“It was a cost savings win and a reliability win coming together,” says Bonny.
What both micro waste and dynamic rightsizing have in common is that neither show up if cloud cost optimisation is treated as a one-time exercise. You cannot simply audit the environment, cut the waste and move on. “It’s about continuously reviewing, just like you would do in your own environment,” says Vieira. “The review and optimisation process needs to be a continuous thing.” The logic is straightforward – workloads change and business requirements shift – so a service that was right-sized in January may be badly over-provisioned by June, or vice versa. Without a continuous loop of monitoring, analysis and adjustment, even a well-optimised environment drifts back towards waste. A tool, like Nutanix Cloud Manager, is designed around this idea. It runs ongoing reports that flag which workloads are sitting in the right environment and which should be moved, repriced or retired. “It’ll tell you exactly which workloads should be where, for the right reasons,” says Vieira. “Maybe it’s time to move something from on-prem to cloud because it would run more optimally there. Or maybe there are workloads in cloud that should come back in-house.”
The point here is that not every workload has value, yet every workload has a cost. With cloud costs, visibility needs to come before everything else because you can’t optimise what you can’t see. “Accountability works best when it is visible,” says Engelbrecht. “The aim is not only protection, but informed control.” Visibility without action is just a dashboard. What organisations need, says Engelbrecht, is a governance structure that makes acting on what they see the default, not the exception.
One of the most common myths about cloud is that it’s inherently cheaper than on-premises infrastructure. It can be, but it isn’t always. Flexera’s ‘2025 State of the Cloud Report’ found that organisations are still overshooting cloud budgets by an average of 17%. “The total cost of ownership is really what needs to be considered,” says Vieira. “Not just the front-end cost per hour or per day, but everything that goes into it – skills, risk, performance, availability, business priorities.” The calculation changes depending on the workload, the organisation and the time horizon. A three-to-five-year TCO model, for example, looks very different from a monthly bill comparison. “For us, it’s not about whether it’s on-prem or in-cloud,” says Vieira. “It’s about the right fit for the customer.”
AI is becoming one of the most significant cloud cost drivers and unlike the waste problems that came before it, it’s arriving before organisations have any idea what they’re spending or why. GPU instances, API calls and SaaS subscriptions are being provisioned at speed, with little governance and even less clarity on what business value they’re supposed to deliver. And now, this is all showing up as bill shock. “Unvalidated AI adoption has entered the mix,” says Engelbrecht. “AI tools are deployed before business value is proven.” It’s a familiar pattern. If anything, it’s the same dynamic that gave rise to cloud sprawl in the first place: fast adoption, slow governance, deferred accountability. “A lot of organisations just want to do AI and that’s a really scary thing,” says Vieira. “You need to know what business objective you’re trying to meet and whether AI can actually get you there.” Rushing workloads into public cloud AI services before a model is validated risks both cost overruns and data exposure. His advice is to start with an on-prem proof of concept, validate the model and then scale into cloud when the use case is proven and the data governance is in place.
Cloud value in Africa is being shaped by geopolitics, regulation and foreign exchange – and 89% of organisations are already adjusting their cloud approach in response. “In this environment, realising cloud value depends on how effectively organisations respond to shifting geopolitical, regulatory and economic pressures,” says Mark Allderman, Africa cloud and digital leader at PwC South Africa. “Cloud strategies must be flexible enough to adapt to new compliance requirements, resilient to currency and cost fluctuations and aligned to national sovereignty expectations.” The PoPI Act in South Africa, Nigeria’s NDPR and Kenya’s Data Protection Act all add compliance complexity to what is already a technically demanding environment, and with 45% of organisations saying geopolitical shifts are directly affecting their cloud strategies, the pressure to get this right is only growing. None of this makes cost optimisation easier. “In 20 years, that conversation hasn’t changed,” says Vieira. “The only way to mitigate it is to invest in tooling that helps people run these services without a steep learning curve.”
The right tools create visibility and control, but without a commitment to continuous review, even the best-optimised environment drifts back towards waste. This is why cloud cost optimisation has outgrown the idea that it’s just a technical problem with a technical fix. The real issue is lifecycle management. “In the cloud, in theory, there’s an infinite amount of resources and so people just tend to deploy and say, ‘Don’t worry, we’ll sort it out next week’,” says Vieira. “That cost shock comes about very quickly. Cloud is definitely not a deploy-and- forget concept.”
* Article first published on www.itweb.co.za
Share
