The decision that costs the most to get wrong
The cloud-vs-on-premise debate stopped being religious around 2024. In 2026 the question is operational: for this workload, in this regulatory context, at this growth profile — which model actually wins on TCO over three years? That's a finance question disguised as an architecture question, and most companies still answer it badly.
This guide is the version we wish more clients had read before signing a five-year cloud commit or before pouring a million euros into a server room. It covers when cloud wins, when on-premise wins, when hybrid is the only honest answer, the repatriation trend, a usable TCO worksheet and a five-step decision process.
Outline
- When cloud genuinely wins in 2026
- When on-premise wins in 2026
- Why hybrid is the realistic default
- The repatriation trend and what triggers it
- TCO worksheet — the line items that actually matter
- A five-step decision process
1. When cloud genuinely wins in 2026
- Variable or spiky load. Marketing campaigns, end-of-quarter reporting, seasonal e-commerce. Paying for peak as opex beats provisioning for peak as capex.
- Global reach. Edge presence in 30+ regions in a week is something on-premise simply cannot match for any reasonable budget.
- Fast time-to-market. New products, MVPs, customer pilots. Cloud kills the "rack and stack" window entirely.
- Managed services as a moat. Hosted Postgres, queues, identity. If you'd otherwise hire two engineers to run it, the cloud price is cheap.
- Disaster recovery without a second site. Cross-region failover replaces a cold DR datacentre for most mid-market workloads.
2. When on-premise wins in 2026
- Steady, high-utilisation workloads. A database running at 70% CPU 24/7 is up to 3× cheaper on owned hardware over three years.
- Data sovereignty and sector regulation. Healthcare, defence, certain financial subsectors and public sector still pull workloads on-prem to satisfy NIS2, DORA or national clauses.
- GPU and AI training at scale. Cloud GPU pricing in 2026 still punishes long, predictable training runs. Owned (or co-located) GPU servers pay back in 12–18 months for steady users.
- Predictable monthly bills. Some boards simply cannot operate with cloud bill variance. On-premise gives you a fixed depreciation line.
- Latency-bound workloads. Factory floor, retail POS, trading desks — physical proximity wins.
3. Why hybrid is the realistic default
Most mid-market companies don't choose cloud or on-premise. They drift into a mix and try to govern it later. The healthier version is intentional hybrid:
- Stable core (ERP, file services, primary database) on-premise or in a private colo.
- Variable edge (web, APIs, customer-facing apps) in public cloud.
- Burst capacity for peaks via cloud, leaving the core at modest utilisation.
- DR mirror in the opposite model from the primary.
Pick one identity provider, one network plane, one observability stack across both. Otherwise hybrid becomes "two infrastructures, double the headcount".
4. The repatriation trend and what triggers it
By 2026, cloud repatriation has stopped being a Basecamp curiosity and is a quiet line item in many CIO reports. The triggers are consistent:
- Egress fees eating 15%+ of monthly cloud spend.
- GPU costs for long inference or training runs.
- Compliance reinterpretation under NIS2 or sectoral rules.
- Mature steady-state workloads that have been in cloud for 3+ years at >60% utilisation.
- M&A forcing consolidation onto an existing on-premise estate.
Repatriation is rarely a full retreat. It's usually pulling 20–40% of spend back — the part that never benefited from cloud's elasticity.
5. TCO worksheet — the line items that actually matter
Don't compare list prices. Build a three-year TCO with these lines for both options:
| Line item | Cloud | On-premise |
|---|---|---|
| Compute | Reserved + on-demand mix | Hardware + 3-year depreciation |
| Storage | Per-GB + per-class | Drives + redundancy + spares |
| Egress / WAN | Per-GB outbound | Carrier links + MPLS / SD-WAN |
| Licensing (DB, OS, hypervisor) | Often bundled | Often separate, sometimes cheaper |
| Operations headcount | Lower for managed services | Higher; include on-call burden |
| Power, cooling, space | Included | Real and growing in 2026 |
| Backup and DR | Per-GB + retrieval | Tapes / second site / cloud target |
| Risk premium | Vendor lock-in | Hardware obsolescence |
Add a 15% contingency to both. Anyone who hands you a TCO without contingency is selling, not advising.
6. A five-step decision process
- Inventory by workload. Don't decide for "the company" — decide for each workload class.
- Profile utilisation. Average and peak CPU/RAM/IOPS over 90 days. Steady high utilisation favours on-premise.
- Map regulatory constraints. Some workloads have a single legal answer.
- Build a 3-year TCO per option. Use the table above. Include exit costs.
- Pick a target mix, then govern it. Document which class of workload goes where, who approves exceptions, and revisit every 12 months.
Key takeaways
- Cloud wins for variable load, global reach, fast time-to-market, managed services and lightweight DR.
- On-premise wins for steady high-utilisation workloads, sovereignty, sustained GPU work and predictable billing.
- Intentional hybrid is the realistic 2026 default — but only if you keep one identity, network and observability plane across both.
- Repatriation is no longer fringe. Egress, GPU and compliance are the usual triggers.
- Decide per workload, with a three-year TCO and exit costs included.
If you'd like an independent TCO review before committing to a multi-year cloud contract or a new server room, get in touch — you'll talk to an engineer, not a sales rep.