EU-hosted model APIs for Office and PowerPoint plugins
Compare EU-hosted AI model APIs for Office and PowerPoint add-ins. Evaluate providers on DPA availability, sub-processor terms and per-document pricing.
In this article7

What shipping inside someone else's tenant adds
The sub-processor question your customer will ask
An Office or PowerPoint add-in does not run in the vendor's own environment. It runs inside the customer's Microsoft 365 tenant, processing documents the customer owns, so the residency question arrives from the customer, in writing, before a procurement decision: where does the document content go, which company processes it, under which agreement, and who sits in the processing chain. The vendor is the one who has to answer, because the vendor is the processor of the customer's personal data and the customer's controller cannot accept an answer it cannot audit.
- A written statement of where prompts and document content are processed, from the model provider's own legal or trust pages, with a date.
- A data processing agreement (DPA) the customer's controller can accept, covering the vendor as processor and the model API as a sub-processor.
- A sub-processor list and a change-notification procedure, so the customer can object to a new downstream provider.
- A billing unit the vendor can price per call, because a per-seat platform fee cannot be passed through to per-use add-in customers.
The obligations are not informal. EDPB Opinion 22/2024, adopted in October 2024, sets out what controllers must verify before engaging a processor and what a processor must do before engaging a sub-processor, including ensuring each link in the chain carries equivalent guarantees[1]. An add-in therefore inherits a chain of at least three parties: the customer, the vendor, and the model API provider. If the provider publishes no sub-processor position, the vendor has nothing to hand its customer except a request for one.
Residency claims also need to be read against jurisdiction, not just server location. The US CLOUD Act lets US law enforcement compel a US-based provider to disclose data it controls regardless of where the servers sit, which is why a hosting-region label alone is not a complete answer to the customer's question[2]. The written answer an ISV ships should name the provider, the processing location it publishes, and the agreement terms, each with the date they were checked.
DPA availability per provider
The commercial question follows the compliance question. An ISV pricing a per-call feature inside an Office add-in can budget against Lyceum's published per-token rates, Qwen3.5 9B at $0.15 per 1M input and $0.20 per 1M output tokens, with cached input at $0.04 (read 2026-09-22)[3][4], instead of a per-seat licence it cannot pass on. The table below sets that billing question beside the DPA and sub-processor questions for eight European providers.
| Provider | DPA availability | Sub-processor position | Billing unit and published rate (retrieved 2026-09-22) |
|---|---|---|---|
| Lyceum | – | – | Per 1M tokens, USD excl. VAT: Qwen3.5 9B at $0.15 input / $0.20 output, cached input $0.04 |
| IONOS | – | – | Per 1M tokens, USD: Qwen3.5-9B at $0.11 input / $0.17 output on the AI Model Hub |
| Nebius | – | – | Per 1M tokens, USD: GLM-5.3-Flash at $0.15 input / $0.50 output in Token Factory |
| Open Telekom Cloud | – | – | Per 1M tokens, EUR: €0.20 input / €0.65 output for T-Cloud-hosted models, from a €1,000 monthly minimum commitment; cached input billed at 15% of the input price |
| OVHcloud | – | – | – |
| Scaleway | – | – | Per 1M tokens, EUR excl. tax: qwen3.8-27b at €0.60 input / €3.30 output, cached €0.12, Paris region |
| STACKIT | – | – | Per 1M tokens, EUR: standard model at €0.15 input / €0.25 output, EU01 Germany South region |
| TensorX | – | – | – |
A cell reading "–" means the provider's page lists no value as of the retrieval date. Several of the gaps are material for a vendor preparing a written answer: OVHcloud's public price list states hourly rates for GPU instances and carries no per-token model API rate[9][10], and TensorX publishes no per-token rate on a public pricing page. A sub-processor position for model inference is not stated on these pages, not even in general privacy policies,[11] so that document has to be requested in writing from each provider before it can appear in an add-in's customer-facing terms.
This gap is the practical difference between shipping and stalling. A separate comparison of European serverless inference providers scores region disclosure, pricing unit and DPA terms provider by provider, which is the evidence set a customer's data protection review will ask for. The minimum a vendor should hold in writing before launch is a dated price list, a DPA the provider will sign, and a residency statement for the specific model endpoint.
Cost per document, not per token
Per-token rates are wholesale units; an add-in feature is priced per use. Take a stated workload: a 10,000-token document, ten user queries per document, and 500 output tokens per query, with the document context resent in every request. That is 10,000 fresh input tokens on the first call, 90,000 cached input tokens across the nine follow-up calls where a cached rate is published (100,000 fresh input tokens where it is not), and 5,000 output tokens per document. Worked against published rates, the cost per document differs by a factor of roughly five across the providers compared, and caching changes the result again.
| Provider and model | Input rate | Cached input rate | Output rate | Cost per document (worked example) |
|---|---|---|---|---|
| Lyceum, Qwen3.5 9B (USD, read 2026-09-22) | $0.15/1M | $0.04/1M | $0.20/1M | ≈$0.006 |
| Nebius, gpt-oss-120b, eu-north1 (USD) | $0.15/1M | – | $0.60/1M | ≈$0.018 |
| Open Telekom Cloud, T-Cloud-hosted model (EUR) | €0.20/1M | 15% of input price | €0.65/1M | ≈€0.008 |
| Scaleway, qwen3.8-27b, Paris (EUR, before tax) | €0.60/1M | €0.12/1M | €3.30/1M | ≈€0.033 |
| STACKIT, standard model, EU01 Germany South (EUR) | €0.15/1M | – | €0.25/1M | ≈€0.016 |
The example assumes the cheapest usable caching behaviour: one fresh-context call and nine calls billed at the cached rate where one is published, all fresh input where it is not. The rows are not like-for-like, because the models, currencies and regions differ, and the Open Telekom Cloud figure is example pricing for T-Cloud-hosted models on a plan whose minimum monthly commitment is €1,000[7]. It still separates the two design decisions: Scaleway charges €0.60 per million input tokens for qwen3.8-27b against STACKIT's €0.15 per million standard input tokens, a fourfold spread on the input side, and providers that publish no cached rate leave the resend cost at full price.
Prompt caching is therefore a shipping requirement, not an optimisation. A plugin resends the same document context across every query a user makes in a session, so the input side dominates the bill; Lyceum prices cached input at $0.04 against $0.15 fresh for Qwen3.5 9B[3], and Scaleway at €0.12 against €0.60. How these rates move over time, across a fixed basket of open-weights models, is tracked weekly in the European Inference Cost Index, and the mechanics of rate cards and per-million-token billing are set out in a separate guide. The choice an ISV ships against is the combination: a DPA its customer will accept, a sub-processor position it can document, and a billing unit that turns into a per-document cost it can defend.
Choosing an API you can ship
Frequently asked questions
Which AI API can I ship in an Office add-in?
Several European providers offer hosted model APIs. An ISV must choose an API provider that offers a formal Data Processing Agreement (DPA) and clearly documents its own sub-processor chain to satisfy enterprise compliance requirements.
Do I need a DPA to resell an LLM API?
Yes. When your add-in processes customer data via a third-party AI model, the API provider acts as your sub-processor under the GDPR. You must have a DPA with them to pass the necessary data processing terms on to your customers.
What does an AI add-in cost per document?
Costs depend on the model and token usage. For example, Lyceum lists Qwen3.5 9B at $0.15 per 1M input tokens and $0.04 cached, about $0.006 per document in the worked example, and STACKIT's standard model costs €0.15 per 1M input tokens. Because add-ins often resend the same context, choosing an API with prompt caching reduces the effective cost per document.
Where does an add-in send its data?
An add-in sends user data to the infrastructure hosting the model API. For EU compliance, ISVs often select European endpoints. However, developers must verify the physical datacentre locations with their chosen provider, as some may use external cloud providers for their infrastructure.
Sources
Rankings and scores on Optinest cannot be bought — no commercial relationship changes how a provider is assessed. How we score
Related reading
All Providers →European serverless inference providers: region, price and DPA
Compare European serverless inference providers based on published disclosures for model processing regions, pricing units, and transparent DPA terms.

Deploying a Hugging Face model as an API endpoint
Evaluate the routes to deploy a Hugging Face model as an API endpoint. Compare managed endpoints, serverless platforms, and VMs across European providers.

Choosing a model backend for AI agents in the EU
Compare European inference providers for AI agents on tool calling, rate limits and pricing. See which EU endpoints support parallel and forced tool calls.