Cloud or On-premises – What Really Pays Off?

Cloud infrastructures are often regarded as flexible, scalable - and economically attractive. But how viable will this assessment really be in 2025? The rising operating costs of large cloud platforms, growing demands for planning security and the increasing use of compute-intensive workloads are causing many companies to re-evaluate their decisions.

This article examines the conditions under which on-premise can make economic sense, what role FinOps plays in this - and why the cost question can only be answered in the context of specific usage scenarios.

Asking the Right Question

"Which is cheaper: cloud or on-premises?"

This question is often asked, but it is rarely productive. Infrastructure costs cannot be evaluated in isolation: they depend on usage patterns, utilization, integration requirements, the regulatory environment, and available resources.

Under these conditions, blanket answers are of little help. At the same time, the pressure to make decisions is increasing: operating costs are rising, flexibility requirements are growing, and budgets must remain predictable.

That's why it's worth asking the question in a more nuanced way:

• Under what conditions does the cloud make economic sense?
• When is in-house operation worthwhile – technologically and financially?
• And how can both models be combined without creating redundancies or disruptions?

The following sections show why it is worth taking a differentiated view – and which criteria can help.

What Practice Shows: Cost Assessment Under New Circumstances

In the past, the use of cloud services was often associated with cost advantages – primarily due to their flexibility and rapid scalability. However, it has since become apparent that this assumption does not apply under all conditions.

Many companies are finding that while public cloud offerings provide operational advantages, they do not automatically lead to lower overall costs. According to Flexera's State of the Cloud Report, 84% of companies have difficulty reliably controlling their cloud spending. Around one-third exceed their planned budget by more than 17%. At the same time, an estimated 30% of booked resources remain unused – so-called cloud waste.

Price developments are also contributing to the reassessment:

• Microsoft increased prices for Azure and Microsoft 365 services by up to 40% in April 2025
• According to Canalys global cloud operating costs will rise by around 19% in 2025 – driven in particular by growing demand for AI services, storage, and data traffic

These developments are not a temporary phenomenon. They reflect the maturity of a market that is becoming increasingly differentiated – and whose use must now be strategically planned and selectively deployed.

What Fundamentally Distinguishes Cloud and On-premises Costs

The key difference lies in the cost structure:

• Cloud costs are usage-based and variable. Depending on utilization, they increase linearly with compute time, storage space, API calls, or data volume.
• On-premises costs are largely fixed. Investments are made in advance (CAPEX), but the marginal costs per unit of use decrease as utilization increases.

As a result:

• For temporary or difficult-to-plan workloads the cloud can make economic sense
• For continuously utilized scenarios on-premise becomes more attractive—both from a cost perspective and in terms of control and predictability

Why a TCO Analysis is Crucial

A reliable assessment requires a total cost of ownership (TCO) analysis covering a period of three to five years. This is the only way to record the total costs in a structured manner—including all hidden effects.

Only when all relevant items on both sides are fully considered can economic efficiency be reliably assessed—and translated into sound decisions.

Practical Scenario: GPU Workload On-premises vs. Cloud – a Calculation Example with a Signal Effect

A particularly illustrative example of the economic differences between cloud and on-premises is provided by a TCO analysis from Lenovo, which compares GPU workloads over a period of five years. The focus is on comparing a dedicated on-premise server with NVIDIA A100 GPUs with the use of AWS p5 instances, which provide comparable GPU resources in the public cloud.

Conclusion of the comparison

• Break-even point: After around 12 months of continuous operation, the on-premises server is more economical – from this point on, every additional GPU hour in the cloud incurs direct additional costs.
• Savings over five years: Depending on the pricing model in the cloud, there is a delta of $1.5 to $3.4 million in favor of in-house operation.

What does this mean in practice?

Not every company operates dedicated GPU clusters 24/7. The calculation example presented here is a special case, but by no means a rare one. In many scenarios, permanently high utilization is realistic – especially in data-intensive and AI-related areas of application:

• SaaS providers whose databases, caching systems, or ML functions must be permanently available
• Medical research institutions that analyze large image datasets with deep learning models
• Financial service providers that use GPU-based real-time scoring, risk analysis, or fraud detection
• Media companies that continuously transcode or evaluate video formats

In these cases, continuous utilization is not the exception, but the rule. This is precisely where the usage-based cloud model reaches its economic limits. What leads to recurring costs in the cloud becomes a predictable investment with foreseeable amortization on-premises.

Further Considerations: Technical and Strategic

In addition to operating costs, there are other factors that speak in favor of in-house operation for permanently computationally intensive workloads:

• Lower latency through local data storage and direct access paths
• More control over security and data protection, especially for sensitive training or patient data
• Better predictability, especially for multi-year research or funding projects with a fixed budget
• Avoidance of vendor lock-in, for example through proprietary cloud APIs in the area of model training or deployment

It is also clear that operating your own GPU infrastructure requires expertise, physical space, and coordinated cooling and power supply.

But if you have a long-term need, the economic benefits can be really noticeable—and make strategic sense.

When is On-premise Worthwhile?

The cost-effectiveness of on-premise infrastructures depends less on the technology stack than on the usage pattern. The more stable and predictable a workload is, the more likely it is that in-house operation will pay off – especially if there are additional requirements for integration depth or regulatory control.

A rough guideline:

• Below five hours of use per day, a cloud model can be economically advantageous.
• From around six to nine hours of daily utilization, on-premises often becomes the more cost-effective option – even with moderate investments

Other typical indicators for cost-effective on-premises operation:

• Constant or high bandwidth requirements
• High egress costs for outgoing data volumes
• Close integration with existing systems or local interfaces
• Requirements for data sovereignty, availability, or auditability

FinOps as a Bridge Between Cloud and On-premises

The economical use of cloud resources has led to the development of FinOps in recent years—a structured approach to cost control, consumption analysis, and operational management.

Many organizations are finding that the underlying mindset can also be applied to on-premise environments.

Examples:

• Measure resource consumption → Optimize utilization
• Analyze workloads → Better justify operating models
• Automating provisioning → Reducing the workload on staff and avoiding errors

In short: On-premises is not automatically efficient – it becomes so when the same principles are applied as in modern cloud environments.

This creates a consistent economic view of hybrid infrastructures – not ideological, but data-based.

Conclusion: Cost-effectiveness Arises in the Context of Use

There is no blanket answer to the question of whether cloud or on-premise is cheaper – and that is precisely the key insight.

Whether a model is economically viable depends not only on price lists or performance data, but also on:

• the actual utilization
• the planning horizon
• the integration effort
• and the general conditions within the company

Cloud services offer flexibility and short time-to-market. On-premise offers control, predictability, and in many scenarios—especially with consistently high utilization—significant cost advantages.

Economically viable decisions are made when both models are evaluated in a transparent manner and combined in a targeted way.

Outlook for Part 3

In the next article, we will take a look at the term "digital sovereignty" – beyond political buzzwords. We will analyze what the CLOUD Act, EUCS, and GAIA-X mean in practice – and when infrastructure decisions also become legally relevant.