When a business looks for an alternative to ChatGPT, it is rarely looking for a better chatbot. It is looking to stop sending its data to a third party, to get out of a per-user subscription that grows with the team, or to stop depending on US law. None of these three problems is solved by switching cloud providers: they are solved by changing where the AI lives. This guide walks through the real options, from least to most sovereign, with their honest limitations.
As long as the assistant runs on someone else's servers, three questions remain open, whatever the name on the invoice:
For a small structure, these questions weigh more heavily than for a large group, as we detailed in why an SME has more to lose with cloud AI. They serve here as a reading grid: each option below answers them differently.
This is the most common answer, and it deserves a fair treatment. OpenAI states that it does not train its models on business data by default (ChatGPT Business, Enterprise, API), whereas the consumer versions can feed model improvement. If training was your only concern, a business tier addresses it contractually.
Everything else stays put. The company operating the service remains subject to US law, and therefore to the CLOUD Act, even when the servers are in Europe. We took this mechanism apart in why "hosted in Europe" is not enough. Your data still lives at a third party, and the per-user subscription keeps running. The contract improves, the situation stays the same.
A European operator without an American parent company (Mistral AI, for example) falls outside the scope of the CLOUD Act. That is genuine legal progress, and for some businesses it is enough.
But the model does not change: your data still lives at a third party, you still depend on its terms and on its survival, and the subscription remains. You have moved the trusted third party, you have not removed it. GDPR compliance for your own usage also remains entirely your responsibility, as we explain in cloud assistants and the GDPR.
At the other end of the spectrum, you can run an open model (such as Llama or Mistral) directly on your own machine, with a tool like Ollama. Nothing leaves your premises, no subscription, no foreign law. On the three opening questions, it is a perfect score.
The limit lies elsewhere: you get an engine, not a workstation. No team management or access rights, no knowledge base indexing your business documents, no organized history, and maintenance falls on you. Community interfaces fill part of these gaps, at the cost of an assembly you have to build and maintain yourself. For a curious developer, it is a playground. For a team that needs to work, it is a construction site.
The fourth option combines the previous two: a complete application, but installed on your server. The team interface, the accounts and access rights, the knowledge base answering from your documents, the memory that persists from one conversation to the next: everything lives on your infrastructure. And the language model remains your choice: an open model running locally through Ollama so that nothing leaves, or a cloud model with your own keys when a use case justifies it.
This is the principle of a sovereign AI, which we detailed for SMEs, and it is what Gungnir does: an assistant that installs on your premises, model-agnostic, with no per-user licence, whose code is published and auditable. On the three opening questions: your data lives at your place, your law applies, and if the vendor disappeared tomorrow, your instance would keep running.
The spending changes in nature. A cloud AI is paid for in per-user subscriptions, every month, indefinitely, at a price the provider can revise. An AI installed on your premises is paid for differently: a server (a VPS is enough to start), model usage (your own API keys billed on consumption, or hardware if you go fully local), and deployment support if you have no technical team. There is no hidden cost that grows with the size of the team, and no one can revise the price of what belongs to you.
Open-source models (such as Llama or Mistral) are free to license and run on your own hardware. Free to license does not mean free to operate: you need a server, setup time and maintenance. The right calculation is total cost of ownership, set against per-user subscriptions that add up every month.
It solves part of it: OpenAI commits to not training its models on business data by default. What remains is jurisdiction (an American company stays subject to the CLOUD Act, wherever its servers are), the location of your data at a third party, and the dependence on a subscription whose terms you do not set.
Yes, with open models running locally: nothing leaves your infrastructure. Let's be honest about the trade-off: the most powerful models on the market remain cloud services. Many business uses (writing, summarizing, questions over your documents) run very well locally, and a model-agnostic platform lets you mix both according to how sensitive each use is.
No. Installing a platform like Gungnir is designed to be simple (a server, one command), and deployment support is part of the offering for structures without a technical team. The goal is not to turn you into a system administrator.
