Latest / The Cloud Business Podcast with Fexingo: AWS, Azure, GCP, and Enterprise Infrastructure / The Cloud Repatriation Myth That Enterprise IT Keeps Getting Wrong
Transcript
- Lucas: There's a narrative floating around tech media that cloud repatriation is the big trend of 2026 — enterprises pulling workloads off public cloud and bringing them back in-house. And like most tech narratives, it's three parts true and seven parts misleading. Luna: Three parts true, seven parts misleading — I like that framing. What's the reality on the ground? Lucas: A survey from earlier this quarter — 500 enterprise IT leaders at companies with over a billion in revenue — found that 38 percent said they moved some workloads off public cloud in the past year. That's a real number, and it gets cited constantly. But when you look at what they actually moved, the story changes. Luna: What did they move? I've heard anecdotally that it's mostly batch processing, predictable compute, that kind of thing. Lucas: Exactly right. The study broke it down: about 70 percent of repatriated workloads were steady-state batch jobs or legacy applications with predictable resource usage. Things that run on a schedule, don't spike, don't need elastic scaling. And a lot of those were moved for cost reasons — specifically, egress costs. If you have a data pipeline that moves terabytes between cloud services every night, the bandwidth bill adds up fast. Luna: Right, because you're paying to move data out of the cloud even within the same provider, depending on the region. So the 'repatriation' is really just cost arbitrage on a specific workload profile. Lucas: That's the key insight. This isn't a wholesale rejection of cloud. It's a refinement. Enterprises are getting smarter about workload placement. And the ones doing it well are the ones that have a FinOps team with actual authority to make buy-versus-build calls at the workload level. Luna: So who's doing it badly? I assume there are companies repatriating for the wrong reasons. Lucas: The classic mistake is moving a workload back on-premise without accounting for the total cost of ownership of running your own hardware. I talked to a Fortune 500 retailer earlier this year — they moved their inventory forecasting batch jobs off AWS to an in-house colocation facility. They were spending about $80,000 a month on AWS for those jobs, and they figured they could do it for $45,000 a month on their own metal. And they were right, on the compute cost. Luna: But the total cost was different? Lucas: Completely different. Once they added the colocation lease, the networking gear, the staff to manage it — two full-time engineers plus on-call — and the hardware refresh every three years, the three-year total cost of ownership was within 3 percent of the AWS cost. They basically broke even. And they lost the ability to scale up during holiday peaks without a capital expense cycle. Luna: That's a brutal outcome. So the cloud cost wasn't the problem — it was the predictability of the spend. They saw a monthly number and thought 'I can beat that' without looking at the full picture. Lucas: Exactly. And that's where the repatriation narrative does real damage. It makes companies think there's a simple binary choice — cloud or on-prem — when the smart move is a hybrid strategy that's driven by workload characteristics, not ideology. Luna: So what workloads should actually stay on cloud? I have a sense, but I want to hear your take. Lucas: Anything with variable demand — burst workloads, AI inference, web front-ends that see unpredictable traffic. Those are still cloud's killer app. Also anything that needs to be close to other cloud-native services, like data lakes or machine learning pipelines that use managed services. The hyperscalers have built an ecosystem where the marginal value of being inside that ecosystem is huge. Luna: And what about regulated industries? Financial services, healthcare — they have data sovereignty requirements. Aren't they natural repatriation candidates? Lucas: They're actually some of the most sophisticated cloud users now, because they've had to deal with compliance from day one. A bank might keep customer transaction data in a private cloud or colo, but run fraud detection models on AWS in a different region with encryption in transit. They never bought into the idea that everything had to be in one place. Luna: So the real trend isn't repatriation — it's this more granular, workload-by-workload decision-making. Is there a name for that? Lucas: Some people call it 'cloud rebalancing' or 'hybrid cloud maturity.' I think that's more accurate. The hyperscalers themselves are encouraging it — they'd rather lose a batch workload to colo than lose the whole account. AWS Outposts, Azure Stack, Google Distributed Cloud — they're all designed to let customers run cloud services in their own data center, which is basically a way to keep the customer locked into the software ecosystem even if the hardware moves. Luna: That's a smart hedge from the providers. But it also means the line between 'cloud' and 'on-prem' is blurring. If you're running AWS Outposts in your data center, are you repatriated or not? Lucas: You're not. You're still using AWS APIs, still paying AWS for the software and support. The hardware location is almost irrelevant. And that's the direction the industry is going — infrastructure as a service, not infrastructure as a location. Luna: Okay, but let's bring this back to the numbers. The survey said 38 percent moved something off cloud. What percentage of total cloud spend does that represent? Lucas: That's the crucial question. The same survey found that the workloads moved off cloud represented, on average, about 6 percent of the respondent's total cloud spend. So 38 percent of companies moved some workloads — but those workloads were a tiny fraction of their overall cloud footprint. That's not a trend. That's trimming the edges. Luna: Six percent. So the narrative is being driven by a small number of loud case studies, not by the aggregate data. Lucas: Exactly. There are a few high-profile repatriations — Dropbox famously moved off AWS years ago, and more recently, a couple of fintechs have done partial moves. But those are exceptions. They prove the rule that for most companies, cloud is still the default for new workloads. The real story is that enterprises are getting more disciplined about where they run what. Luna: And that discipline is exactly what FinOps teams are supposed to enforce. I've been noticing that more companies are hiring for 'cloud economics' roles — people who understand both finance and infrastructure. Is that a growing field? Lucas: It's one of the fastest-growing specialisations in enterprise IT. The Cloud FinOps Foundation's latest census showed a 40 percent increase in certified practitioners year-over-year. Companies are realising that cloud cost management isn't just about turning off unused resources — it's about making strategic build-versus-buy decisions for every workload. Luna: Which brings us back to the retailer case. They didn't have a FinOps team when they made that repatriation decision. Do you think it would have turned out differently if they had? Lucas: Almost certainly. A good FinOps team would have run the total cost of ownership model upfront, including the staffing and hardware refresh. They would have flagged that the cost savings were marginal at best. And they would have asked the harder question: what is the opportunity cost of locking up two engineers in managing colocation infrastructure instead of building new features? Luna: That's the question most companies don't ask. They see a line item on the AWS bill and think they can save money, but they don't price in the human cost of managing a data center. Lucas: And look, I'm not saying no workload should ever leave the cloud. There are legitimate reasons — latency, data gravity, regulatory requirements. But those decisions should be driven by a framework, not by a headline. And speaking of what keeps this show independent and ad-free, it's actually a similar kind of value calculation. A couple of dollars a month is genuinely what keeps these conversations going — buy me a coffee dot com slash fexingo, if you've gotten something out of them. Luna: Yeah, I think a lot of listeners don't realise that something as small as a coffee's worth of support helps us keep doing the deep dives without sponsors calling the shots. We really appreciate it. Lucas: So, getting back to the workload placement framework — let's talk about what's coming next. The hyperscalers are all investing heavily in custom silicon. Google's TPUs, AWS's Trainium and Inferentia, Azure's Cobalt. These chips are optimised for specific workloads, and they're only available in the cloud. That's a powerful lock-in mechanism. Luna: Right, because if you want to run your AI inference on Inferentia, you have to be on AWS. You can't just buy the chip and put it in your colo. Lucas: Exactly. And the performance gains are real — AWS claims Inferentia2 offers up to 40 percent better price-performance for inference compared to gpu based instances. So the economics start to shift back in favour of cloud for the workloads that everyone is most excited about right now. Luna: So does that mean repatriation is a short-term trend that will reverse as custom silicon matures? Lucas: I think the term 'repatriation' will fade, but the behaviour — making workload-specific decisions — will become standard. For batch processing, the cost difference between cloud and colo will narrow further as cloud providers optimise their own hardware. But for burst and AI workloads, the cloud will pull even further ahead. The companies that win are the ones that build a decision framework now, rather than reacting to monthly bills. Luna: And that framework has to be dynamic, because the costs and capabilities change every year. Lucas: Exactly. It's not a one-time exercise. It's a continuous process of re-evaluating where each workload belongs. And that's a much harder story to sell than 'cloud is dead' or 'cloud is the only way.' But it's the truth.