Self-Hosted AI: From Fallback to Standard

The End of the Compromise: Why Our Own Model Is Now the Standard Companies that want to use AI should not only ask which model is the most capable. Equally important are where company data is processed, which dependencies are created, and how costs develop as usage grows. A convincing answer has to take all of these questions into account.

At AITEZA, we have been running our own language models for almost two years. We started with DeepSeek R1. Our main motivation back then was digital sovereignty. Today there is a second one: the quality of self-hosted models is now good enough to take on a considerably larger share of our day-to-day work.

This changes the role of self-hosted AI. What was once an alternative for particularly sensitive content is increasingly becoming the standard model for daily work.

For us, digital sovereignty means that companies can decide for themselves where and under what conditions their data is processed. Internal knowledge, contract documents, and confidential project files should not have to be sent to an external model provider simply because employees want to work more efficiently.

Sovereignty also means being able to switch model providers. Anyone who ties their entire AI usage to a single provider becomes dependent on that provider’s pricing, product decisions, and technical terms. That is why the freedom to choose between different models is a core part of AITEZA. It is also a reason why many customers deliberately choose us over Copilot and comparable solutions.

For a long time, though, that decision came with a noticeable compromise. Self-hosted language models offered more control over data processing and a more predictable cost base. But on difficult tasks they did not always match the quality of the leading commercial models. As soon as large documents had to be analyzed, complex connections explored, or multiple steps reliably chained together, additional model capability was often needed.

Our answer to that is a hybrid approach. AITEZA combines self-hosted models with the ability to integrate powerful external models via API. These so-called frontier models complement our own infrastructure exactly where their extra capability is actually needed. Companies no longer have to commit permanently to one of the two worlds.

That choice should not depend on task difficulty alone. The confidentiality of the content has to be considered too. A complex question is no license to send sensitive data to an external service. Which models may be used for which information should therefore be defined by clear operational rules.

With our self-hosted Qwen3.8 model, the balance between internal and external processing now shifts significantly. Based on our experience so far, it reaches a quality on many typical company tasks that we previously only got from leading commercial models.

What matters to us is less about individual benchmark scores than about working with real company documents. Can information be reliably combined from documents? Are connections captured correctly in a research task? Do usable text drafts come out that do not need to be fully rewritten? Can the model handle a task that requires several steps building on one another?

In these areas, the model convinces us more and more in daily use. That does not mean it replaces every frontier model in every discipline. On particularly difficult questions, external top models still deliver better results. But a different question is more important for deciding on your own AI infrastructure: How much extra capability is actually needed for the work that really comes up?

If a self-hosted model already completes a task at the required quality, using a more expensive external model does not automatically deliver a corresponding benefit. A large share of daily work can then stay within your own controlled environment.

By “self-hosted” we do not necessarily mean a server in the customer’s office. Our model runs in a German data center. What matters is that processing happens in an environment we control and that the inputs do not have to be passed on to an external model provider for that processing. The location alone, however, does not replace a data protection concept. Access rights, retention rules, and responsible handling of the results remain just as necessary.

Economically, this approach becomes more interesting too. Some of our customers spend several thousand euros a month on using commercial models. With token-based billing, costs rise with the volume of inputs and outputs processed. Long documents, frequent follow-up questions, and multi-step workflows can increase the volume further. The more intensively AI is used, the more important reliable cost planning becomes.

Our Qwen3.8 system costs less than 1,000 euros a month and can support the AI day-to-day of a company with around 200 employees. No variable token costs are incurred for using this self-hosted model.

That figure is not, however, a blanket guarantee for every usage scenario. The actual capacity needed depends, among other things, on how many people work at the same time, how large their documents are, and what response times are expected. The number of employees alone says little about the required computing power.

Self-hosted AI is not infinitely scalable without additional costs either. If utilization grows beyond the available capacity, the infrastructure has to be expanded. The economic advantage therefore does not lie in supposedly free computing power, but in a calculable base for regularly recurring work. On top of that come the usage-dependent costs of external models as needed.

For the technically interested, our setup can be described briefly. We run Qwen3.8 27B with FP8 quantization on an RTX 6000 Ada with 48 GB VRAM. We use vLLM as the serving engine. The quantization reduces the memory footprint of the model weights, while vLLM handles serving the model for use. Whether such a setup fits a given company should be judged based on the actual tasks and expected utilization.

For us, this development is an important step. We no longer run our own language model mainly because we want to keep data under our own control and cap costs. We use it because the quality convinces us in production, everyday use.

The hybrid approach remains in place. The self-hosted model takes on a growing share of daily work. External frontier models complement it precisely where their use is professionally sensible and permissible for the data involved. This lets companies use AI more broadly without automatically tying every additional application to an external model provider and its usage-based costs.

Find out more

Discover how AITEZA can help you in your day-to-day work. Or simply book a slot for a product demo. We look forward to hearing from you.