Why the real variable in your infrastructure decision was never the platform.
For most of the last decade, "cloud vs. on-prem" had a pretty settled answer. Cloud was where you were heading, and on-prem was the thing you were migrating away from. That consensus is cracking, and the 2026 data explains why.
Start with cost. AI workloads are reshaping the economics of data centers in ways that are already showing up on utility bills, not just cloud invoices. PJM Interconnection, the grid operator covering a large stretch of the Mid-Atlantic and Midwest, projects $6.3 billion in added grid costs over the next three years, driven largely by data center demand. Virginia, home to the densest concentration of data centers in the world, has seen residential electricity rates climb roughly 13% as that demand competes for capacity. Goldman Sachs expects power prices tied to data center growth to keep rising, forecasting increases through 2027, against a backdrop of roughly $7 trillion in projected global data center buildout by 2030. The Electric Power Research Institute has pointed to the same underlying cause: AI training and inference are power-hungry in a way that older, general-purpose cloud workloads simply weren't, and the grid hasn't caught up.
None of that means cloud is suddenly the wrong choice. It means the cost curve that made "just move it to the cloud" an easy default is bending, and it's worth understanding why before assuming it still holds.
Then there's the repatriation question, and this is where I think a lot of the public conversation overcorrects. Barclays surveyed CIOs and found 86% had moved at least one workload back from public cloud to on-premises infrastructure. That number gets quoted constantly, and taken alone, it sounds like a stampede back to the data center. It isn't. IDC's research puts full repatriation, meaning organizations that moved an entire workload back and kept it there, at only 8-9%. The rest is far more selective: specific workloads coming home while the broader cloud footprint keeps growing. Gartner's numbers back that up. Public cloud spending grew 21.5% in 2025 to $723.4 billion, and Gartner expects 90% of enterprises to be running a hybrid model by 2027. Cloud isn't shrinking. It's specializing, and on-prem is taking back the workloads where it actually has an edge.
Where does on-prem have that edge? Two places, mainly. The first is raw compute economics for sustained, predictable workloads. Nvidia's own data suggests roughly 30% savings on GPU costs for organizations running AI workloads on owned infrastructure versus renting equivalent capacity in the cloud, once you're running at high, consistent utilization. Cloud's pricing model is built for elasticity, and you pay a premium for that flexibility. If your workload doesn't need the elasticity, you're paying for a feature you're not using. The second is regulatory. Frameworks like DORA and NIS2 in the EU, and the tightening interpretation of GDPR's data residency requirements, are pushing certain industries toward infrastructure they can directly control and audit. That's not a preference. For a growing list of companies, it's a compliance requirement.
That regulatory pressure points at the deeper issue, which isn't really about servers at all. It's about where your data, and increasingly the intelligence your systems derive from that data, actually lives. Nutanix and others have framed this as a sovereignty question rather than a trust question, and I think that's the right frame. It's not that public cloud providers are untrustworthy. It's that once your data and your models leave your direct control, you've made an architectural decision with downstream consequences, whether or not you meant to. That's worth doing deliberately.
So what actually decides this, if it's not "cloud is safer" or "on-prem is safer"? In my experience, it comes down to talent. ISC2 puts the current cybersecurity skills gap at roughly 90% of organizations reporting they don't have the staff they need. The World Economic Forum's 2025 research found only 14% of organizations believe they have the talent on hand to execute their technology strategy. That gap is the variable nobody puts on the slide, and it's the one that actually determines whether either choice works.
Here's the part I'd push back on hardest: moving to the cloud doesn't make an infrastructure problem disappear. It relocates it, usually with an egress fee attached. If your team doesn't have the expertise to configure, secure, and maintain infrastructure well, that gap follows you to whatever platform you land on. Cloud misconfigurations are still the leading cause of cloud data breaches. On-prem without a competent team is just as exposed, arguably more so, because you own the perimeter and the patching schedule and the physical security, with nowhere to point when something goes wrong.
That's why I don't think this decision should be made platform-first. It should be made talent-first, workload by workload. Ask what expertise the decision actually requires: someone who understands your compliance obligations, someone who can evaluate whether your physical security is genuinely adequate for on-prem, someone who can tell the difference between a workload that benefits from cloud elasticity and one that's just sitting there accumulating a monthly bill. Most mid-sized organizations don't have that person on staff, and hiring a full-time CIO to make these calls rarely pencils out.
That's the gap fractional leadership is built for: someone who's made this call before, who can sit down with your environment and tell you honestly which workloads belong where, instead of defaulting to whatever's fashionable this year. If that's the conversation you need to have, I'd welcome it.
