Multi-cloud has become a board-level phrase, which is usually a sign that its costs are under-discussed. Running workloads across two providers is a legitimate architecture. It is not, on its own, a strategy, and it is certainly not an insurance policy.
What it actually costs
Every additional provider adds a permanent operating burden:
- Skills. Two sets of networking, identity, security and observability primitives, and people who know both well enough to be on call.
- Tooling. Either duplicated native tooling or a third platform to unify it, which introduces a dependency of its own.
- Data movement. Egress charges and latency between providers, which quietly shape which architectures remain affordable.
- Incident complexity. Every investigation now begins with determining which environment, which tooling and which runbook apply.
- Commercial dilution. Commitments and discounts are negotiated per provider; splitting spend weakens your position with both.
None of these are reasons to avoid multi-cloud. They are reasons to be specific about why you are doing it.
The four reasons that hold up
Data residency and regulation. A regulator or a customer contract requires data to sit in a particular jurisdiction or under a particular provider's certifications. This is the strongest reason, and it is rarely negotiable.
Acquisition or organisational reality. The business acquired a company that runs elsewhere, and consolidation would cost more than coexistence. Multi-cloud here is an outcome, not a goal.
Best-of-breed dependency. A specific capability, such as a data platform, an AI service or a regional network, exists only with one provider and is core to the product.
Genuine resilience against provider-level failure. Rare, expensive, and usually misapplied. If the two environments cannot survive without the same identity provider, the same DNS and the same team, you have two dependencies rather than one.
If your reason is not on that list, the honest question is whether one well-operated cloud would serve the business better. In most mid-market environments, it would.
Portability is a property of your architecture, not a promise from your provider. Containers and open formats help. They do not remove the operational differences between clouds.
What the operating model has to answer
If multi-cloud is justified, the operating model decides whether it works:
- One identity plane. A single directory and access model across both providers. Everything else follows from this.
- One observability layer. Logs, metrics and traces in one place, or incidents are investigated blind.
- A default placement rule. Where a new workload goes by default, and the specific conditions under which it goes elsewhere. Without this, placement is decided by whoever is on shift.
- A named owner per environment. Shared ownership of two clouds means accountability for neither.
- Egress and inter-provider traffic budgeted as a first-class cost, not discovered at the end of the quarter.
- A consolidation review each year. Environments accumulate. Some workloads should move back.
A pragmatic sequence
Most organisations are better served by deepening one cloud before adding a second:
- Make the first environment properly governed: identity, landing zone, logging, cost controls, backup and recovery with tested restores.
- Standardise the operating model: infrastructure as code, change records, incident process, documented runbooks.
- Only then extend, and only for a reason that survives the list above.
A second cloud built on top of an unmanaged first one produces two unmanaged environments and a larger invoice. The order matters more than the destination.