The Right Operating Model for Enterprise AI: Public Cloud, Private Cloud, or Self-Hosted LLM?

Jaber Kakar, Head of Information Security
26. October 2026
A ransomware attack on the Japanese cloud provider IDC Frontier in October 2026 paralyzed the systems of approximately 495 companies and local governments. The incident illustrates just how far-reaching the consequences of an attack on a central IT service provider can be.
The threat raises a crucial question: Where and how should companies and government agencies operate their AI systems to protect sensitive data, mitigate security risks, and at the same time remain economically viable and independent?
In diesem Beitrag:
In a nutshell:
- Operating Models: Companies can use large language models (LLMs) as external AI services or operate them themselves—in the cloud or on their own infrastructure.
- Data Sovereignty: It is not just the location where data is stored that determines how well sensitive data is protected. Access rights, data flows, and control over processing are also crucial.
- Security: Every operating model comes with its own risks and responsibilities. Even self-hosting does not guarantee IT security.
- Costs and Performance: External AI services often allow for a quick start. With high, consistent utilization, operating your own system can be cost-effective.
- Digital sovereignty: Dependencies on providers, opportunities for changing models, and exit strategies should be taken into account from the very beginning.
- Government Agencies, Defense, and Critical Infrastructure: Stringent requirements for security, data control, and availability necessitate a careful selection process and, where appropriate, tiered or hybrid solutions.
The attack on IDC Frontier underscores that companies and government agencies must consider not only the technical capabilities but also security risks and dependencies on external service providers when selecting their IT infrastructure. When deploying AI systems, companies and government agencies must therefore carefully weigh how much control they need over data, infrastructure, and security measures.
The appropriate operating model depends as much on costs and performance requirements as it does on the desired level of digital independence.
What are the different operating models for enterprise AI?
Companies can use large language models (LLMs) as external AI services or operate them themselves. The main differences lie in who is responsible for technical operations, what infrastructure is used, and how much control the organization retains over its AI systems.
Terms such as API, public cloud, private cloud, and self-hosting are often used interchangeably, even though they describe different aspects: An API is a technical interface for accessing an AI model. “Public cloud” and “private cloud” refer to deployment environments, while “self-hosting” describes the independent operation of a model.
To facilitate a clear comparison, we distinguish between four typical operating models below. This classification is intentionally simplified. In practice, hybrid models and other technical architectures are also possible.
Checklist: Is our AI agent sufficiently secure?

An Overview of the Four Operating Modes
External AI Services (Third-Party Operation): Shared or Dedicated Model Environments
With external AI services, a provider handles the technical operation of the language model. Companies access it via an API, for example, and do not need their own model infrastructure.
1. Shared AI Services: Multiple customers use an AI service operated by the provider. This option typically allows for a quick start and flexible, often usage-based billing. How data is processed and protected depends on the provider, the technical architecture, and the contractual agreements.
2. Dedicated AI Services: The provider offers a model environment intended exclusively for the company. This can provide additional options for configuration, resource management, and isolation. However, a dedicated model instance does not necessarily mean that the underlying infrastructure is available exclusively to a single customer.
Self-hosted LLM: Running the Model in the Cloud or On-Premises
With self-hosting, the organization itself assumes responsibility for operating the language model. However, it does not necessarily have to use its own hardware.
3. Self-hosted LLM in the cloud: The company operates and manages its language model on rented cloud resources. For example, it controls the model software and its configuration, but remains dependent on the cloud provider for the underlying infrastructure.
4. Self-hosted LLM On-Premises: The language model runs on locally operated infrastructure, such as the organization’s own data center. This provides extensive control over data processing and network access. At the same time, the organization is responsible for model operations and must also ensure hardware, maintenance, security, and availability.
Public Cloud, Private Cloud, and Sovereign Cloud: What’s the Difference?
The terms “public cloud” and “private cloud” describe the deployment environment, not directly the responsibility for operating the model.
In a public cloud, a provider makes cloud resources available to various customers.
A private cloud, on the other hand, is intended exclusively for a single organization. It can be operated either in the organization’s own data center or by an external service provider. The hardware does not have to be owned by the company itself. A private cloud is also not automatically protected against cyberattacks—the key factors are technical security measures, access controls, and responsibilities.
The three possible scenarios
| Variant | Where is the infrastructure located? | Who owns the hardware? |
|---|---|---|
| Private Cloud in Your Own Data Center | Internal |
Usually for the company |
| Externally Hosted Private Cloud | With the service provider | Often with the service provider |
| Private Cloud Managed by a Service Provider | In-house or outsourced | Depending on the contract |
There is no single, universally accepted definition of a sovereign cloud. Generally speaking, a sovereign cloud is expected to meet additional requirements regarding data control, legal jurisdiction, and digital independence. The specific guarantees that actually exist depend on the particular technical and contractual arrangements.
The choice of cloud environment alone therefore does not determine either security or data sovereignty. The overall architecture, the protective measures, and the distribution of responsibilities are the decisive factors.
Comparison Table: What are the differences between the various operating models?
| Criterion | Shared AI Service | Dedicated AI service | Self-hosted cloud | Self-hosted on-premises |
|---|---|---|---|---|
| Model Operation and Maintenance (Updates, Patches, Model Changes) | Vendor | Vendor | Company | Company |
| Infrastructure | Providers | Provider | Cloud providers | Locally operated |
| Model Selection | Mostly the provider’s current top-of-the-line models | Mostly models from the provider, depending on the offering | Typically open-source and open-weight models | Typically open-source and open-weight models |
| Control Over Model Operations | Low | Contract-dependent | High | High |
| Setup effort | Usually low | Low to moderate | Medium to high | High |
| Cost Structure | Often usage-based | Often capacity- or contract-based costs | Cloud resources and in-house operations | Hardware and in-house operations |
| Scalability | Usually flexible | Depends on reserved resources | Depends on cloud capacity | Depends on local capacity |
| Data Sovereignty | Depends on the contract and architecture | Depends on the contract and architecture | High level of control over the model and data is possible; infrastructure and legal jurisdiction remain with the cloud provider | Maximum control possible |
The information provided describes typical characteristics and does not constitute a universal evaluation of individual offerings. Responsibilities may vary, particularly with regard to managed services.
In our article on open-source AI, we explain the possibilities offered by open AI models and how they differ from proprietary solutions.
Data Sovereignty and Data Protection: What Data Can Be Shared, and Where?
Whether an AI system is operated in the public cloud, in a dedicated environment, or on a company’s own infrastructure affects how companies can control and protect their data.
When choosing an AI operating model, several questions are crucial:
- Where is the data stored?
- Who can access it?
- Where is the data processed?
- What information is shared with third-party service providers?
In addition to prompts and uploaded documents, this also applies to company data, model outputs, and logs.
Data Control and Data Protection
When using external AI services, companies should verify what happens to prompts, uploaded documents, retrieved company data, model outputs, and log data: Are the inputs stored, used for model training, or shared with subcontractors? Self-hosted LLMs allow for greater control over data flows and access rights. However, when operating them in the cloud, dependencies on the infrastructure provider remain.
When personal data is processed, the requirements of the GDPR must be met. These include, among other things, an appropriate legal basis, provisions regarding data processing on behalf of others, and, where applicable, international data transfers. Neither a private cloud nor self-hosting automatically guarantees GDPR compliance.
Special Requirements for Government Agencies, the Defense Sector, and KRITIS
Federal agencies, the German Armed Forces, and KRITIS operators are subject to additional security requirements depending on their area of operation. For classified information such as VS-NfD, the relevant security guidelines must be followed. Operating on one’s own hardware alone is not sufficient for this purpose.
Regardless of the operating model, clear rules are needed to determine which AI applications may be used and which data may be processed. A structured AI governance system helps define responsibilities and manage risks. Learn more in our article on ISO 42001 and AI governance.
The NIS 2 regulation also imposes cybersecurity risk management requirements on affected companies and organizations. Read more about this in our article on NIS 2.
LLM Security: How secure are the various operating models?
Data sovereignty alone does not guarantee IT security. Even self-hosted large language models (LLMs) require comprehensive security measures. The key factors are which vulnerabilities arise and who is responsible for securing them.
What Cyberattacks on Japanese Companies Reveal About Security Vulnerabilities
The attack on IDC Frontier mentioned at the beginning illustrates just how heavily companies depend on the security and availability of their IT service providers. Security experts also warn that AI could help attackers overcome language barriers and detect vulnerabilities more quickly.
Japan used to have a certain degree of natural protection due to its language. Emails or phone calls from people who were not fluent in Japanese stood out and aroused suspicion. But with the advent of large language models (LLMs), this protection has disappeared. However, there is no proof that AI was actually used in these specific attacks.
The key finding: Missing security updates, weak access controls, and inadequately protected systems pose a risk regardless of the operating model.
What security risks are associated with APIs, the cloud, and self-hosting?
Every AI operating model comes with its own vulnerabilities. These include APIs, model servers, plugins, and interfaces to corporate data. Typical risks include:
- Vulnerabilities and missing updates: Outdated software can allow attackers to gain access to AI systems and connected applications.
- Inadequate access controls: Permissions that are too broad increase the risk of data leakage and misuse.
- Insecure software supply chains: Models, libraries, and external components may contain vulnerabilities or tampered content.
- Prompt injection: Manipulated inputs can cause AI systems to bypass intended instructions or perform unwanted actions. This is particularly relevant for AI agents with access to external tools. Learn more in the article “Using AI Agents Safely.”
With external AI services, the provider handles key aspects of technical security. However, companies remain responsible for secure integration, access permissions, and the handling of their data. With self-hosting, additional responsibilities include model operation, security updates, monitoring, and responding to security incidents. In cloud environments, companies and infrastructure providers share certain responsibilities.
That is why “security by design” should begin as early as the selection of the operating model. This includes a secure architecture, clearly defined responsibilities, access restrictions, and continuous monitoring. The new ISO/IEC 27090:2026 standard supplements ISO 27001 with guidance on AI-specific threats such as data poisoning, model theft, and prompt injection, and provides technical guidance on these issues.
In addition to technical safeguards, trained employees and clearly defined security processes are crucial. The ML Gruppe supports companies and government agencies with information security training, awareness training, and ISMS consulting to help them identify and effectively mitigate security risks.
Realistically Comparing AI Costs and Performance
Security and data sovereignty are key criteria when choosing an AI operating model. In practice, however, companies and government agencies must also keep cost-effectiveness in mind. This involves not only the prices for using the model, but also the total cost of ownership (TCO) as well as requirements for performance, availability, and scalability.
How much does an AI operating model actually cost?
With external AI services, billing is often based on tokens consumed—that is, the units of text processed. Dedicated services, on the other hand, may incur fixed capacity or provisioning costs.
Self-hosting incurs costs for computing power—particularly for high-performance graphics processing units (GPUs)—as well as for setup, maintenance, security updates, and personnel. Rented cloud infrastructure also incurs costs. With on-premises operations, there are additional investments in proprietary hardware and its ongoing operation.
At what point does self-hosting become worthwhile?
When usage volumes are low or fluctuate, external AI services are often more cost-effective: Companies typically pay on an as-needed basis and do not have to maintain their own model infrastructure.
A tipping point can occur when a language model is consistently operating at high capacity. In this case, fixed operating costs are spread across many requests, while usage-based API costs continue to rise. Whether self-hosting actually becomes more cost-effective depends, among other things, on model size, GPU utilization, electricity and personnel costs, and the prices of the external service.
Performance: It’s Not Just About Computing Power
In addition to cost, companies should also compare model quality, response times (latency), scalability, and availability. More powerful models often require more computing resources, while smaller, specialized models may be sufficient for certain tasks.
External services can often flexibly handle peak loads. With self-hosting, sufficient computing capacity must be available or provisioned to handle these peaks. Therefore, the key factor is not the most powerful model, but one that reliably meets the functional, technical, and economic requirements.
Digital Sovereignty: How Independent Will the Company Remain?
In addition to security and cost-effectiveness, digital sovereignty plays an important role in choosing an AI operating model. This refers to the ability to have self-determined control over AI systems and data and to switch providers or models as needed.
Vendor Lock-in: When Switching Providers Becomes Difficult
Strong vendor dependence (vendor lock-in) arises, for example, when AI applications are closely tied to proprietary interfaces, specific model functions, or a particular cloud infrastructure. This can make switching providers later on technically complex and expensive.
Even self-hosting does not automatically protect against dependencies: Companies may, for example, be reliant on specific hardware, software platforms, or external service providers.
Consider exit strategies from the very beginning
To ensure that companies remain capable of acting effectively in the long term, they should consider the following points when selecting an operating model:
- Portability and model switching: Can applications be migrated to a different language model or infrastructure with reasonable effort?
- Data export: Can data, configurations, and relevant logs be exported in full and in usable formats?
- Open standards: Do standardized interfaces and interoperable technologies facilitate a switch to a different provider?
- Exit strategy and contingency plan: How can AI operations be continued in the event of a provider change or provider failure?
This independence is particularly relevant for federal agencies, the defense sector, and KRITIS operators. In these areas, long-term availability, security of supply, and control over critical IT dependencies must be taken into account alongside economic considerations.
Digital sovereignty therefore does not mean doing without external providers, but rather consciously limiting dependencies and creating genuine alternatives for action.
Decision-Making Guide: Which AI Operating Model Is Right for Which Application?
The appropriate AI operating model depends on the data’s protection needs, security requirements, available resources, and operational conditions. The following overview shows typical use cases and possible solutions.
Which operating model is suitable for which protection needs?
| Use Case | Key Requirements | Possible Operating Models | Key Prerequisites |
|---|---|---|---|
| Bundeswehr and Defense | Security, Data Sovereignty, Isolated Environments | Self-hosted on-premises, self-hosted in a suitable cloud environment | Compliance with security regulations, required clearances, controlled access |
| Federal Agencies | Data Protection, Data Sovereignty, Compliance, Procurement Requirements | Dedicated AI service, self-hosted cloud, or on-premises; for suitable data, also a shared AI service | Review of data processing, security requirements, legal bases, and vendor dependencies |
| KRITIS Operators | Availability, Resilience, Vendor Management, Emergency Preparedness | Dedicated AI service, self-hosted cloud, or on-premises; shared AI service if applicable | Fail-safe design, emergency response plans, security monitoring, controllable service providers |
| Companies with lower security needs | Cost-effectiveness, scalability, rapid integration | Shared or dedicated AI service, possibly self-hosted cloud | Data protection, access controls, transparent costs, contractual provisions |
The table is intended as a guide, not as a set of blanket recommendations. Particularly in the case of classified information, admissibility depends on the specific classification, the applicable security requirements, and the suitability or required clearance of the operating environment. Hybrid solutions may also be appropriate.
Checklist: Is Our AI Operating Model Sustainable?
Before making a decision, companies and government agencies should consider the following six questions:
- Protection Requirements: What data is processed, and how sensitive is it?
- Requirements: What legal, regulatory, and organizational requirements apply?
- Data Control: Where is data processed, and who has access to it?
- Responsibilities: Who is responsible for model operation, IT security, maintenance, and security updates?
- Cost-Effectiveness and Performance: Are the total costs, model quality, scalability, and availability appropriate?
- Exit Strategy: Can the provider, model, or infrastructure be changed if necessary?
Answering these questions helps you systematically evaluate the various operating models and make an informed decision based on your specific needs.
Conclusion: When it comes to AI operating models, there’s more to it than just the technology
Whether it’s an external AI service, a self-hosted LLM in the cloud, or operation in your own data center: Each model offers advantages, but also entails risks and responsibilities. The operating model must align with the data’s protection needs, security requirements, and available resources.
However, choosing an operating model is not a one-time decision. New AI capabilities, changing threat landscapes, and regulatory requirements may necessitate a reassessment. Digital sovereignty therefore also means being able to continuously monitor and adapt AI systems and, if necessary, further develop them independently of individual providers.
“The key question for the future will no longer be where an AI model runs, but how much control we retain over its actions. The more autonomous AI systems become, the more important it will be to have an architecture that ensures security, transparency, and human decision-making autonomy in the long term.” — Jaber Kakar, Head of Information Security
Using AI Safely and Responsibly
Whether in the cloud or self-hosted, information security starts with clear responsibilities and the right expertise. The ML Gruppe supports companies and government agencies with practical training and consulting on information security.



“The key question for the future will no longer be where an AI model runs, but how much control we retain over its actions. The more autonomous AI systems become, the more important it will be to have an architecture that ensures security, transparency, and human decision-making autonomy in the long term.” — Jaber Kakar, Head of Information Security