On-Prem vs Cloud vs Colocation: UK Cost Reset 2026

The 2026 placement reset: why the question changed

For the past few years the infrastructure conversation was dominated by cloud repatriation headlines. By mid-2026 that debate has matured into something more useful for UK buyers: a workload-by-workload cost reset. One widely cited five-year reference case puts on-premises total cost of ownership at $5.6 million against $5.75 million for colocation — figures close enough that other factors decide the outcome — while public cloud runs approximately 60% higher than equivalent capacity in the same model.

That 60% premium isn't a universal law; it's the outcome of one reference scenario. But it lines up with what UK-specific modelling shows too, which is the reason this reset matters for anyone planning 2026–2029 infrastructure budgets: the gap between cloud and owned/rented capacity is now large enough, and persistent enough across independent models, that it deserves a proper line-by-line comparison rather than a policy default.

On-prem, colocation and cloud: what they actually mean now

On-premises means you own the hardware and the space it sits in — full control, full capital exposure, and full responsibility for power, cooling and physical security. It's the model most under scrutiny in 2026, and it's worth reading a balanced view on whether you should consider if the on-prem server is dead before assuming the answer.

Colocation means you still own the hardware but rent rack space, power and cooling in someone else's UK facility — you keep control of the kit and the software stack while offloading the building. If you're unfamiliar with how this differs from managed hosting, it's worth taking a moment to understand what colocation is before comparing quotes.

Public cloud strips out ownership entirely: you consume compute, storage and networking on demand and pay for what you use, with elasticity as the main selling point. Most UK organisations in 2026 aren't choosing one model exclusively — they're deciding which workload belongs where, which is why it's worth understanding how to explore public, private, and hybrid cloud models work together in practice.

The UK cost equation: CAPEX and OPEX by component

One 2026 mid-sized-organisation model breaks the components down clearly. On the on-premises side: hardware refresh costs of $300,000 to $500,000 per cycle, IT staff costs of $200,000 to $400,000 annually, and power and cooling costs of $100,000 to $250,000 annually. The same analysis puts a well-managed colocation solution at $600,000 to $900,000 annually and a robust cloud solution at $800,000 to $1.2 million annually for a comparable mid-sized organisation.

Colocation pricing itself isn't flat — density drives cost. A 2026 generative-AI TCO paper prices high-density power configurations at roughly $1,500 per rack per month, against roughly $600 per rack per month for standard-density racks. That gap matters directly to UK buyers: as GPU-dense and AI-adjacent workloads push power draw per rack higher, the colocation bill scales with density, not just floor space — and UK power availability constraints make that density premium a live commercial factor rather than a theoretical one. That's not simply a density issue either — UK energy price exposure adds a second layer of cost volatility, since power is increasingly contracted and priced separately from rack space, meaning buyers who model only floor-space cost risk missing the larger and less predictable power line on the invoice.

None of the component figures above are UK-quoted prices — they're US-modelled ranges — but the shape transfers directly: staffing, power/cooling, and hardware refresh are the three levers that determine whether on-prem beats colocation for your organisation, and all three need modelling in GBP, with VAT and cross-connect charges added, before you compare them against a cloud quote.

Worked example: three years of UK infrastructure spend

The most UK-specific hard number available is a displayed three-year TCO scenario from a UK colocation-versus-cloud calculator: £278,600 for colocation against £435,045 for the equivalent public cloud deployment. Treat this as one modelled scenario rather than a market-wide benchmark — your own workload, contract terms and utilisation will move both figures — but it's a genuinely UK-currency data point, and it corroborates the direction of the international reference case above.

Put the two data points side by side and the pattern is consistent even though the scenarios and currencies differ: colocation tends to land closer to on-premises cost than to cloud cost, and cloud carries a real premium once you're running steady, sized capacity rather than bursty, unpredictable demand. The practical takeaway for a UK budget-holder is to run your own three-year model with your actual utilisation, egress and staffing assumptions before assuming either figure applies to you.

Beyond the price tag: workload fit, control and compliance

Cost alone doesn't decide placement — workload pattern does. One 2026 workload framework classifies steady-state, 24/7, performance-sensitive workloads as colocation candidates, and variable-demand or development/test workloads as cloud candidates. A separate 2026 AI-workload analysis finds that persistent AI workloads achieve the lowest TCO in colocation over three years, while public cloud has both the highest cost and the lowest cost predictability in the same comparison.

A 2026 cloud-cost comparison adds three specific workload-fit rules worth applying directly: steady-state high-volume compute running at 70%+ utilisation can be cheaper on owned hardware at scale; high-volume storage at petabyte scale can be cheaper on owned storage than object storage plus lifecycle fees; and workloads with meaningful internet egress are heavily affected by cloud egress charges, making on-prem or colo with direct bandwidth potentially cheaper. For UK buyers with data-residency or latency-to-UK-user requirements, colocation adds a compliance and performance argument that a pure cost comparison won't capture. This matters most in procurement terms where sector regulation dictates where data can physically sit: UK IT leaders evaluating cloud contracts for regulated workloads often find that the sovereignty and residency question is settled before the cost comparison even starts, which is why colocation's UK-based, owner-controlled footprint continues to feature in procurement shortlists even when a pure TCO model favours cloud on paper.

Hidden costs and pitfalls in the UK data centre market

The single biggest pitfall in UK evaluations is comparing headline instance prices or per-rack rates without adding VAT, cross-connects, internet transit and power commitments into the model — none of these show up on a marketing rate card, and all of them show up on the invoice. In practice, these are recurring monthly line items rather than one-off fees: cross-connects and internet transit are typically billed per circuit per month on top of the base rack rate, and power is contracted separately from space — which is exactly why the same colocation product can range from roughly $600 to $1,500 per rack per month once density and power draw are factored in, rather than being a single flat figure. Any UK procurement model that stops at the headline per-rack or per-instance price will miss these recurring commitments entirely.

Power density is the second trap: the difference between standard-density and high-density colocation pricing shows how quickly a GPU-dense refresh can push a rack from a modest monthly line item into a genuinely expensive one.

Staffing is the third, and most underestimated, pitfall. On-premises infrastructure needs a dedicated in-house team to manage hardware, patching and physical resilience — a real, ongoing cost that many capital-expenditure business cases quietly omit. Colocation reduces the facilities burden but still requires hardware and operations expertise in-house or via a managed partner; cloud shifts the skills requirement toward platform engineering and FinOps discipline rather than removing it. Whichever model you choose, model the staffing line explicitly rather than assuming it nets to zero.

Making the call: when each model wins

On-premises makes sense when you have the capital appetite, in-house expertise, and steady enough utilisation to justify owning hardware outright — and when the five-year reference case's near-parity with colocation gives you confidence the maths works at your scale.

Colocation wins for steady-state, 24/7, performance-sensitive workloads, persistent AI/GPU deployments, and UK-latency or data-residency-sensitive applications, where it consistently shows the lowest three-year TCO in the AI-workload analysis above and avoids the capital burden of building your own facility.

Cloud remains the right default for variable demand, development and test environments, and short-lived experimentation — anywhere the workload's unpredictability outweighs the cost premium that steady-state comparisons consistently show for cloud capacity.

The hybrid answer: splitting workloads by economics

The strongest 2026 evidence doesn't point to a single winner — it points to hybrid placement. One 2026 analysis of hybrid colocation models frames enterprises as cutting 30% to 40% of OpEx by categorising workloads properly rather than defaulting everything to one platform: bursty and development workloads go to cloud, while persistent, storage-heavy, high-egress or GPU-dense workloads move to colocation or owned infrastructure where local power and space economics support it.

For UK IT leaders building a 2026–2029 roadmap, that's the practical conclusion: don't re-run the on-prem-versus-cloud debate as a single decision. Segment your estate, price each segment against the component breakdown above, and let the workload economics — not platform loyalty — decide where each system lives.

Sources

Every figure in this article traces to the sources below.

• •Databank — 5-year on-prem/colo/cloud TCO reference case for 2026
• •Databank — full TCO breakdown for AI workloads, 2026 update
• •Lenovo — on-premise vs cloud generative-AI TCO, 2026 edition (per-rack colocation pricing)
• •Cagelab — UK colocation vs cloud TCO calculator, 3-year scenario
• •Imperial Datacenter — colocation vs cloud vs on-prem TCO analysis, 2026
• •AWSNegotiations — AWS vs on-premises TCO, 2026 workload-fit analysis
• •Databank — how enterprises cut 30–40% OpEx with hybrid colocation models