For most districts, using a commercial AI tool is the easy answer. But, what if we ran the AI ourselves? Hospitals, banks, law firms, and government agencies have already gone down this road, not because they love managing servers, but because they handle sensitive information and want control over where it goes and who touches it.
Every time a school district sends student or staff generated data to an outside AI vendor it is trusting the vendor’s contracts, security, and data practices. Self-hosting offers another path: the school, coop or state keeps control of the data, the model, and the infrastructure itself.
This is easier to do today than it would have been even two years ago. Tools like Ollama and vLLM let organizations run capable open-weight and even open source AI models on their own hardware. The auditing and access controls that a school system would need are also available. None of this requires building AI from scratch anymore.
No school in Minnesota has tried this, yet, as far as I know, but Merced County's education office in California has already piloted self-hosted AI tools across departments serving 20 school districts, using Ollama and Open WebUI. It's a small-scale proof of concept, not proof that this works at classroom scale or saves money long-term. But it shows an education agency can actually do this. Minnesota might just be good enough and smart enough to do this, and doggone it, people like us.
Owning the infrastructure also changes how teachers and students feel about the tool. Much of today's push to be "more human" in the classroom is really a response to commercial AI and pretty much all of the other apps used in schools being controlled by a distant company, not anyone in the building. That distance breeds distrust. When a school controls its own AI server, the relationship shifts: teachers can see what the system is doing, students can ask how it actually works, and the tool feels like something the school uses on its own terms. Students and teachers might even try to do some retraining of the model. The unease about AI in schools may have less to do with AI itself than with who holds the keys.
Self-hosting doesn't have to mean every district buys its own servers. Germany offers a different model: a shared, government-coordinated AI system called BildungsLLM, built on Llama 3.3 and run through a national cloud provider, all within German data centers under GDPR rules. Individual schools don't run anything themselves; they just use the shared service. Another plus is that BildungsLLM (and its portal frontend AIS.chat) is integrated into a school's Moodle system, so the data flow remains strictly contained within a closed, secure network sovereign ecosystem
That's an option worth considering. I don’t expect many Minnesota districts will actually use Moodle again, but a large district might run its own AI system. A cluster of small districts might rely on a state or regional consortium instead. And there's a third option that is also possible: a small, air-gapped AI server, disconnected from the internet, sitting in a single school or classroom. It wouldn't replace a district's main system, but it could be a low-stakes way to teach how AI actually works, what a model is, how it responds to prompts, why it sometimes gets things wrong, all without any data ever leaving the building. For schools wary of handing data to any outside company, an air-gapped box is a way to start teaching AI literacy with almost no risk. It’s a great tool for experimenting and teaching students about AI.
Self-hosting requires hardware, electricity, security, and staff time. A small district with light AI use may still find a commercial subscription cheaper. But for a large district or service cooperative with thousands of users, the calculation is different. The money goes toward infrastructure the school system actually owns, instead of writing increasingly larger checks to a vendor indefinitely.
K-12 also gets to go second. Healthcare, finance, and government already worked out the hard parts of running AI responsibly, including plenty of costly mistakes. Schools can borrow what worked without repeating what didn't. This doesn’t mean self-hosting is right for everyone. Small and rural districts often lack the IT staff to run a self-hosted AI safely, and commercial vendors offer support contracts, uptime guarantees, and someone else absorbing the operational risk. There are trade-offs.
The real question isn't whether every school should run its own AI servers. It's whether co-ops, and large districts will build enough of their own AI infrastructure and know-how to keep control over the increasing technology, data, and costs their schools will depend on more each year.
