Telekom as an inference provider — and why that solves more than a technical problem

I recently got access to T-Systems’ AI Foundation Services — and with it, among other things, to GLM 5.2. Short version: I am glad to see Telekom position itself as a provider of AI inference. This is not a footnote. It solves a problem that comes up in nearly every conversation I have.

The question I actually get asked

The portfolio companies I work with are uneasy. Not about whether AI is worth it — that question is settled. About who gets to see their business processes. US providers, Chinese providers: in both cases something is left over.

And yes, the legal ground exists. With the right contracts in place, the data may not be used for training. And still that remainder sits there — and anyone who dismisses it as irrational has not been listening. A contract is a promise about something you cannot see.

It is sharpest with ERP vendors. Their value is not the code. Their value is the process: the particular way a workflow is modelled, ground down over years with customers. The worry is not “someone reads my data”. It is: if my process ends up in the training data, anyone can suddenly rebuild it.

I think that conclusion is wrong — and it does not matter

I disagree with it. Just because someone can rebuild a product does not mean they can place it successfully in the market. Trust and the customer relationship are the real moat, not the reproducibility of a workflow. Anyone who believes their business rests on nobody understanding how they do it usually has a weaker business than they think. A separate entry on thissoon

But that conviction is of no use to the person who has to sign the approval today. A worry does not disappear because you refute it. It disappears when you take away what it rests on.

Which is exactly what AI Foundation Services do

T-Systems runs an inference platform with more than 15 models behind an OpenAI-compatible API — switching costs you a line of code. The catalogue lists open models such as Llama 3.3, Mistral, Qwen3, GPT-OSS and, yes, GLM 5.2 (currently in preview) alongside proprietary ones.

What matters is what the docs say under Enterprise Trust. For the models T-Systems operates itself, in the T-Cloud in Germany, the wording is explicit: prompts and responses are “never saved or used to train or fine-tune any model”, they are “not stored” and “not viewable” — not by Telekom, not by third parties. Request and response only, no retention, no training. (Enterprise Trust)

That is precisely the commitment that takes away what the worry above rests on. Not “we are not allowed to”, but “it is not stored”.

Where does the prompt end up?

Straight to a US or CN provider No training, by contract. Still a promise about something you cannot see.

Proprietary models via the platformGPT, Claude, Gemini Forwarded to Azure or Google Cloud. Their terms apply, processing may happen worldwide.

Open models in the T-CloudLlama 3.3, Mistral, Qwen3, GPT-OSS, GLM 5.2 not stored · not viewable · never used to train. Request and response, nothing else. Operated in Germany.

The sovereignty argument rests on the bottom row — not on the provider, but on the model category.

The catch you need to know about

That commitment covers the category T-Systems operates itself — the open models in Germany, GLM 5.2 included. The proprietary models (GPT, Claude, Gemini) are forwarded to Microsoft Azure or Google Cloud, where the respective provider’s terms apply and, per the documentation, processing may happen worldwide.

So if you want the sovereignty benefit, stay on the European-operated side of the catalogue. That is not a criticism — it is the operating manual. And it is the first time I can tell an ERP vendor: take a strong open model, it runs in Frankfurt, your process is stored nowhere.

I had the chance to test it over the weekend. Speed, output quality, throughput: I am impressed. This is not a compromise you accept out of regional loyalty — it is a serious option.

And it is a better answer than what I could offer the sceptics before. The alternative was local hosting: expensive, narrow in model choice, heavy to operate, and rarely current enough for what the work demands. There is now a third path between “send everything to a US hyperscaler” and “put GPUs in our own basement”.

I think this will win over a sceptic or two.

Sources

AI Foundation Services — documentation. Model catalogue, OpenAI-compatible API (https://llm-server.llmhub.t-systems.net/v2), notes on EU data sovereignty.
Enterprise Trust — the data-handling model, verbatim. Source of the quotes: no storage, no training use for the EU-operated models; forwarding of the proprietary models to Azure and Google Cloud.
T-Systems launches AI Foundation Services for enterprises. Deutsche Telekom, press release.
AI Foundation Services. T-Systems, product page.