Someone else’s computer

First posted on LinkedIn, November 2025.

Does anyone else remember when we were all told that moving stuff to the cloud meant the responsibility for hardware failure would lie with the cloud vendor?

Here’s roughly where we’ve landed instead. Azure App Service is a single-region service: if the region goes down, your application goes down with it. Microsoft’s own documentation is admirably clear that reliability is “a shared responsibility”, that regional failover is a decision the customer orchestrates, and that you should design, test and run your own failover rather than assume the platform will move you somewhere else. Which is to say: if the datacentre your app lives in falls over, that’s your app to look after, and nobody is quietly relocating it on your behalf.

To be fair to Microsoft (and I do mean this), nothing there is dishonest, and it’s better documented now than it has ever been. It’s also not what most of us were sold a decade ago. The pitch was that racking, cooling, power and disk failure became somebody else’s problem. All true! But somewhere in the excitement, a lot of organisations heard something rather different: that availability had become somebody else’s problem. It hadn’t. The cloud took the hardware off your hands and left the architecture firmly in them.

And architecture costs money. Multi-region isn’t a checkbox, it’s a second set of everything plus the data replication, the failover orchestration and the testing to prove it works. When the business case for the migration was written, I’d wager that second set of everything wasn’t in the spreadsheet.

So are we at the end of the cloud rush?

Presumably the next big thing will be big metal boxes we keep in our own buildings with all our stuff stored on them. Revolutionary.

Only half joking, because the numbers have genuinely turned. In Barclays’ CIO survey at the end of 2024, 86% of CIOs said they planned to move at least some workloads from public cloud back to private or on-premises infrastructure, the highest figure that survey had ever recorded. IDC had around 80% of organisations expecting to repatriate some compute or storage inside a year.

But before anyone declares the cloud over: the same research puts full exits at roughly 8%. Almost nobody is closing their cloud account. What’s actually happening is that the steady, predictable, always-on workloads (the ones that never needed elastic scale in the first place) are being looked at again, and often they don’t survive the look. Cloud remains a brilliant answer for bursty, variable, growing things. It’s an expensive way to run a server that does the same amount of work every day for six years.

What the argument is really about

“Cloud” and “on-premise” were never the interesting question. The interesting question is whether anybody has decided, explicitly and in writing, what happens when the thing you depend on isn’t there. Because that decision gets made either way: deliberately, by you, in a design; or accidentally, by default, at 3am on the day it matters.

So the ones I’d ask of your own estate:

• If the region hosting your most important application became unavailable this afternoon, what actually happens, and who moves it?
• Does anyone outside the IT team know that answer, and have they agreed it?
• Which of your workloads are genuinely elastic, and which are steady-state ones paying elastic prices?
• What did the migration business case assume about resilience, and did we ever build the bit it assumed?

How many organisations will be bringing things back on premise over the next five years? More than the sales decks suggest, fewer than the headlines claim, and the ones who do it well will be the ones who worked out what they were paying for in the first place.

There is no cloud. It’s just someone else’s computer, and their support contract.

Update, July 2026

A wee bit has happened since. Microsoft’s guidance has, if anything, got firmer: the current advice is explicitly that customers should own and rehearse their own failover rather than expect the platform to initiate it. Meanwhile a string of high-profile outages across the big providers has made the point more vividly than any architecture review ever could, and AI workloads have become a genuine repatriation driver in their own right: steady, hungry, GPU-heavy and punishingly expensive to rent by the hour.

The direction of travel isn’t back to the server room, though. Gartner reckons around 90% of organisations will be running hybrid infrastructure by 2027, which is the honest end state: workloads placed where they earn their keep, rather than wherever the strategy slide said everything was going. “It depends” was the right answer all along. It just doesn’t fit on a slide.