The fantasy of leaving the cloud

Posted on: 7 July 2026

Every so often the notion resurfaces that the subscription model has run its course, and that we are drifting back towards software bought once and installed on the machine in front of us, the way it was twenty years ago. This time the thought carries more force than usual, seductive because it promises to reclaim something, yet the difficulty is that the numbers refute it outright, and it is precisely the strength of the wish that tells us something exact about where control is actually going.

Consider the figures before indulging the story. Worldwide spending on public cloud, on Gartner's estimates, passed seven hundred billion dollars in 2025, up by more than a fifth on the year before. The subscription economy as a whole continues to swell towards eight hundred and fifty billion, and something close to fifteen thousand new subscription applications appear every month. When eighty-six per cent of chief information officers tell the Barclays survey they intend to bring some workload back in house, only around eight per cent, on IDC's reckoning, are planning a full exit from the cloud. The mass return to the software of old does not exist in the data. It exists in the wish, which is a different matter entirely.

The right question, then, is not whether the rental model is dying, because it is not, but why so many would like it dead just now, and the answer is economic before it is nostalgic.

Software sold as a service looked cheap for fifteen years thanks to two subsidies now expiring together. The first was capital at zero cost, which allowed the firms in the sector to sell below cost in order to take share and bind the customer to them. The second, more recent, is artificial intelligence offered at a flat rate when every single request is dear to whoever supplies it. That flat rate was bait. In June GitHub moved to metered billing for its coding assistant, admitting with unusual candour that the product now devours too much compute to hold at a fixed price. Uber, which had adopted that tool at scale, watched its annual AI budget evaporate in a matter of months, the real bill arrives now, and when it arrives the arithmetic of renting in perpetuity changes sign for everything steady and predictable.

This is the moment a choice that looked eccentric stops being so. When 37signals, the firm behind Basecamp, left Amazon to buy its own servers and reported seven million dollars saved over five years, plenty of people raised an eyebrow. Today the same sum is being done by a good many others, not out of ideology but because renting forever a thing you use constantly is the most expensive way to own it. We are not looking at a war between two models, but at the same cost reappearing from two directions: the supplier pushed the flat rate to tie you in and raises it the moment you are tied, while the party that owns its own hardware pays that cost once rather than forever, and rising.

Then there is the matter of ownership, which in today's software has become close to a paradox. A programme bought once was a thing of yours, whereas a subscription is a permission someone else can revoke, and now and then they do revoke it. When its new flagship model appeared, one large laboratory pulled the previous version out of the application overnight, leaving those who worked with it to find it gone by morning. The great outages of the past year taught the same lesson by the rough route, from the Amazon region that periodically drags half the web down with it to the Cloudflare failure in November that rendered thousands of sites unreachable: the cloud is someone else's computer, and when it breaks it grants you no say. This, more than price, is the exposed nerve, the point at which the tenant discovers he never owned anything.

Up to here, though, it would be only the familiar pendulum, in which computing has swung back and forth between the local machine and the remote centre since computers have existed, from the mainframe to the personal computer, from the personal computer to the web, from the web to the cloud. To anyone who was in the room when the render farms of digital cinema moved off in-house servers onto rented ones and then came home again, following nothing but the price of compute, the oscillation has lost any power to surprise. I watched the same thing happen in the tools for making music, migrated from physical hardware to the plug-in to the monthly fee. The pendulum always returns, but it never returns to quite the same point.

There is a precedent worth more than any forecast. A century ago factories produced their own electricity with generators on the premises, then the centralised grid rendered them uneconomic and everyone plugged into the socket, only to discover later that for certain critical uses, the hospitals, the data centres, anyone who cannot afford the current to fail, the generator on site never disappeared but became the silent backbone. Not a return to the past, then, but a stratification: the grid for the bulk, the private plant where an interruption costs too much. Computing is tracing the same shape.

And here is the novelty, the thing that makes this turn different from all the earlier ones. For the first time the most demanding and most coveted workload of the moment, artificial intelligence, has reasons of its own to sit on the machine in front of you rather than in the remote centre, and not out of nostalgia but for structural reasons. There is data sovereignty, which for anyone handling sensitive information or working inside European rules is not a detail but the constraint that decides where a thing may run. And there is ownership in the most literal sense, because a downloaded model is a file on your disk, yours on today's laptop and tomorrow's, while responsiveness, the fact that it answers without the journey to a distant server, comes almost as a bonus. In June NVIDIA unveiled a class of chip built to run serious models on a laptop, while Microsoft slipped into Windows models that work on the device without ever calling the cloud. The centre of gravity of computing, for the first time in years, has a technical reason to come home.

Put together, these pieces describe not a return but a bifurcation, in which the aggregate goes on centralising, cloud spending still climbing and subscriptions multiplying, while in the meantime a thin, high-value layer departs on its own account. These are the regulated workloads and the ones that cannot bear delay, along with everything that artificial intelligence makes it worth keeping close to whoever uses it. Finance in practice, healthcare, anyone working with data that cannot cross a border or a jurisdiction. So in the last analysis it is not the eight per cent that flees which makes the story, but which eight per cent departs and why, because that layer is exactly the part where control weighs more than convenience, the part that fifteen years of painless renting had persuaded everyone to delegate.

The person who cancels the subscriptions and pulls a model onto the laptop believes he is going backwards, to the years of software in a box. He is doing something different and more interesting: not returning to a place but shifting the centre of gravity while the ground beneath moves. The file on the disk you can hold in your hand against the permission that evaporates with an update to the terms, the choice is not between old and new, it is between two opposite ways of standing inside a system that neither of them controls entirely.