Cloud and infrastructure

Fewer than one in a hundred cloud customers change provider in a year. The finding is the Competition and Markets Authority's, published on 31 July 2025 in the final decision of its cloud services market investigation, and it deserved more attention in British boardrooms than it got.

The rest of that document is equally direct. UK customers spent £10.5 billion on cloud services in 2024, with spending growing by nearly 30% a year across the preceding four years. Microsoft and AWS each hold a share of infrastructure supply the CMA put in the 30 to 40% band, with Google in the 5 to 10% band. The inquiry group concluded that competition in these markets is not working well, and recommended that the CMA board consider using its digital markets powers to investigate whether Microsoft and AWS should be designated as having strategic market status.

Exit is the design problem

A market in which almost nobody moves is not necessarily a market in which everybody is happy. The CMA identified the specific frictions: egress fees charged on data leaving a provider, technical differences that make services hard to compare or integrate, and Microsoft's licensing practices, which it found were adversely affecting how well AWS and Google could compete for customers running that software.

For a head of infrastructure this reframes a familiar decision. Every architectural choice is also an exit cost decision, and exit costs are paid years later by somebody who was not consulted. Managed services that shorten delivery today are the same services that price a move tomorrow, and that is a legitimate trade to make deliberately. It is only a mistake when nobody writes down which side of it they chose.

The CMA was more relaxed about two mechanisms that customers often assume are the trap. It found committed spend agreements to be widespread but not, in their current form, harmful to competition, on the basis that rival providers can profitably compete against them. It reached a similar view on cloud credits, while noting that the position could change if their value or scope grew. Both findings are worth reading in the original if your renewal conversation is about a discount tied to a volume commitment.

Meanwhile the architecture most British organisations actually run stopped being a compromise and became the plan. Very little of consequence is single-cloud, almost nothing serious is entirely on-premises, and the interesting engineering question is no longer which model wins but which workloads have a genuine reason to sit where they do. That argument is taken on directly at hybrid stopped being the thing you settled for.

The counter-move that gets discussed most, moving workloads back off public cloud, is also the one with the weakest public evidence in Britain. There is no reliable national dataset of repatriation. Where this section reports a figure it will name the organisation and the filing it came from, and where the number does not exist it will say so rather than borrowing a vendor's estimate. The count is kept at who actually moved back, and what it saved them.

The second constraint is electricity

Compute capacity in Britain is now limited by the grid at least as much as by capital. Two changes in the last two years made that explicit. On 12 September 2024 the government designated data centres as Critical National Infrastructure, the first such designation in almost a decade, the previous additions being the space and defence sectors in 2015. Then on 15 April 2025 Ofgem approved the TM04+ connections reform package, which embeds "ready" and "needed" criteria into how requests for electricity connections are managed.

Read together, those say something uncomfortable about capacity planning. A data centre is now important enough to be defended by the state and constrained enough that its build date depends on a queue position it does not control. Anyone whose 2029 capacity plan assumes British colocation at a given price should know which queue their supplier is in, and the honest answer is that most procurement teams have never asked. The electricity side of that story is followed at the constraint on British compute.

The compute line is rarely the expensive one

Infrastructure budgets are argued over in the units cloud providers bill in, which are instances, storage tiers and network egress. The costs that actually move a large organisation's total are usually one layer up: the database licence that costs more on a rival's platform than on its owner's, the support contract attached to an operating system, and the middleware nobody has been able to remove for eleven years. A migration business case built only from compute prices is a business case that will be wrong by the second invoice, and the licensing half of the argument sits in SaaS and enterprise software.

The corollary is a warning about numbers. A cloud saving quoted as a percentage, with no workload, no region and no term attached, can be produced to order in either direction, and both vendors and their challengers produce them. Where a figure appears in this section it will carry the workload it describes and the document it came from, or it will not appear.