The Quiet Return of On-Premise Computing
After a decade of moving everything to the cloud, some companies are bringing workloads home. The reasons are cost, AI hardware and control, not nostalgia.
For much of the 2010s the direction of corporate IT seemed settled: move workloads to public cloud providers and close the data centre. That trend has not reversed, and cloud spending continues to grow. But a noticeable group of companies is doing something that once looked old-fashioned. They are moving selected workloads back to servers they own or lease in colocation facilities, a practice often called repatriation.
The first reason is cost predictability. Cloud pricing is excellent for variable demand and experimentation. For steady, predictable workloads running around the clock, owning hardware can be cheaper over several years, especially once data-transfer charges are included. Several software companies have published accounts of large savings after moving stable services off public cloud, though critics note that these companies had unusually strong in-house engineering teams.
The second reason is AI. Training and running large models requires graphics processors that have at times been scarce and expensive to rent. Organisations with constant inference demand are calculating whether buying their own accelerators, or reserving dedicated capacity from specialist providers, beats paying on-demand rates. Power and cooling become the constraint: many older corporate data centres cannot supply the density modern AI racks need, which pushes firms toward purpose-built colocation sites.
The third is control. Regulators in several regions have tightened rules on where data may be stored and who may access it. Some companies in finance, healthcare and government-adjacent sectors prefer infrastructure where they can demonstrate physical control. Cloud providers have responded with sovereign-cloud offerings and on-premise versions of their own services, blurring the line.
Repatriation has costs that spreadsheets underestimate. Running hardware requires staff for procurement, maintenance, security patching and capacity planning. Companies lose some of the managed services that made cloud attractive. And hardware bought today may be outdated sooner than expected if AI chips keep improving quickly.
The realistic outcome is hybrid. Most organisations will keep bursty and experimental work in public cloud, place steady workloads where they are cheapest, and use common tools so that software can move between environments.
Why it matters: infrastructure decisions shape costs for years. The lesson is not that cloud was a mistake, but that each workload deserves its own calculation. Limitation: public case studies are self-selected success stories; failed repatriations are rarely published.
Every claim above is sourced to a document, a named person, or a record we hold. Where we could not verify a claim, we say so. Read our standards and corrections policy →