AI video for companies whose material can't touch the cloud

Every AI video tool worth using has the same first step: upload your material to someone else's servers. For most companies, that's fine. For some, it's the end of the conversation. This post is for the second group.

The problem nobody puts in the product demo

AI video generation has become genuinely good. Script generation, image prompting, motion, avatars, voice: the full production chain now runs on models instead of camera crews. Marketing teams everywhere are using it.

Unless they can't.

If you work in banking, insurance, healthcare, defence, or public administration, you know the pattern. Someone on the team finds a tool. The output looks great. Then compliance asks one question: where does the input go?

And the input is the problem, not the output. A product video script contains unreleased features. Internal training footage contains employees' faces and voices. A customer case study contains customer data. All of it leaves the building the moment you press upload, and the tool's terms of service decide what happens next.

So the honest answer for most regulated companies has been: no AI video. Not because the technology isn't ready, but because the deployment model isn't.

What changed

The open source video stack caught up. Not to the level of the best proprietary models, but past the threshold where it produces work you can actually publish.

We recently put this to the test on ourselves. Together with our partner chim.ai, we produced Xinity's own avatar video on Chimaira, their platform for agentic, character based video storytelling, and we deliberately built as much of the pipeline as possible on infrastructure we own:

Script and image prompts: Qwen3.6, via the Xinity API. An open weights model, running on our own hardware. Every draft, every internal detail in the prompt, every iteration stays on the machine in our office.

Motion: Wan 2.2 i2v. Also open source. It is not the strongest video model on the market, and we won't pretend otherwise. It is the best one you can run yourself, and for most corporate video use cases that is enough.

Character animation and voice: still cloud, for now. We could hide this. We'd rather not. The avatar animation currently runs on a cloud service, and the voice comes from ElevenLabs. chim.ai has its own frame by frame animation solution that can be deployed on premise when a project requires it, and usable open source voice models already exist. A fully on-prem pipeline is a setup decision today, not a research problem.

That last paragraph is the point of this post. Sovereign by architecture means you can inspect the architecture, including the parts that aren't sovereign yet.

What "on prem" actually means here

When we say the data stays in the building, we mean something specific:

The models run on hardware you own. In our case, and for most of our customers, that's an ASUS Ascent GX10 sitting in the office. No API calls to a hyperscaler, no inference on rented GPUs in someone else's data center.

The weights are open. Qwen3.6 and Wan 2.2 are downloadable. There is no vendor who can change the terms, raise the price per token, or deprecate the model your workflow depends on.

The interface is standard. Our inference layer is OpenAI compatible, so the tooling your team already knows works unchanged. Switching from a cloud API to your own endpoint is a URL change, not a migration project.

Nothing phones home. No telemetry, no training on your inputs, no "we may retain data to improve our services." The clause doesn't exist because the connection doesn't exist.

Who this is for

You probably don't need this if your video material is already public: product shots, published brand assets, stock-adjacent content. Use the cloud tools, they're excellent.

You probably do need this if any of the following sounds familiar:

• Your legal team reviews every SaaS tool for data processing agreements, and AI tools keep failing the review

• Your video material includes employees, customers, patients, or unreleased products

• You operate under BaFin, FINMA, or FMA supervision, or you're preparing for DORA and NIS2 audits where every external data flow needs documentation

• You've already decided against cloud LLMs for text, and video is the same decision one modality later

What working with us and chim.ai looks like

The division of labour is simple. chim.ai is an AI company covering consulting, training, and development, and Chimaira is their video platform: agentic, character based video storytelling, from avatar design through the animation pipeline to the finished video. Xinity provides the inference infrastructure it runs on: the hardware, the model serving, the OpenAI compatible endpoint your prompts go to.

The result is a video pipeline where the sensitive parts, the scripts and the source imagery, never leave your infrastructure, and where the remaining cloud components are documented, optional, and replaceable.

If someone in your company regularly says "we're not allowed to upload that," that sentence is the reason to talk to us.

Xinity builds open source, OpenAI compatible, on premise LLM inference for regulated enterprises in the DACH region. The avatar video described in this post was produced with chim.ai on their Chimaira platform and is published on our LinkedIn.

Drafted with AI, reviewed by the Xinity team.