On-premise AI for schools: data sovereignty, predictable cost, and the Mac mini case study

The standard pitch for AI in education is a cloud subscription. Monthly per-student fees. API costs that scale with usage. Privacy policies that require lawyers to interpret. Terms of service that change without notice.

IndiLearn's architecture is a deliberate rejection of that model. Every school that deploys IndiLearn in its full on-premise configuration gets a Mac mini in their server room, runs all AI inference on their own network, pays nothing per query, and holds a structural guarantee that no student data ever leaves the building.

This is not primarily a cost argument, although the cost case is strong. It is a procurement argument. Cloud-based AI education tools cannot credibly tell a school principal or DET procurement officer that student data is safe, because the structural reality of cloud inference makes that claim untenable. On-premise inference makes it true by design.

The cloud economics problem for schools

Cloud AI APIs are billed per token — per unit of text processed. For an education platform, this creates a cost structure that scales with exactly what you want to scale: student use. Every lesson, every student submission, every feedback generation is a cost event. The more the platform is used, the higher the bill.

For IndiLearn's feedback platform, a realistic estimate at Sonnet-level pricing with prompt caching is AU$6–12 per student per year in API costs — for 100 feedback calls across a ten-week unit, plus theme synthesis. At that rate, a 600-student school spends AU$3,600–7,200 annually in API fees alone, every year, forever, scaling as students use it more.

That is before platform fees, support costs, or the compliance overhead of managing a cloud data processor relationship under the Australian Privacy Principles.

Why the Mac mini is the right hardware for school AI

The conventional assumption for AI inference hardware is a GPU server. High-VRAM NVIDIA cards, dedicated cooling, server rack, specialist installation. This is the right answer for some use cases. It is the wrong answer for a primary school's server room.

Apple's M-series Mac mini is the right answer for school AI for three reasons: cost, power, and form factor.

• Cost: A Mac mini M4 Pro with 48GB unified memory costs AU$2,000–3,000. An equivalent-performing NVIDIA GPU configuration costs AU$5,000–15,000 in hardware alone, before installation and infrastructure.
• Power draw: The Mac mini draws approximately 30W under AI load. A dual-GPU server draws 600W+. Over a full school year of use, the electricity cost difference alone is material — and the Mac mini is fanless under most workloads, meaning no noise and no additional cooling requirement.
• Form factor: The Mac mini is 12.7cm square and 5cm tall. It sits on a shelf. No rack, no rack unit, no specialist installation. A school with an existing IT cabinet can deploy it in an afternoon.

Unified memory architecture: the technical advantage

The reason the Mac mini can run large language models that previously required expensive GPU hardware is Apple Silicon's unified memory architecture. In traditional systems, the CPU and GPU have separate memory pools. For LLM inference, the model weights must be loaded into GPU VRAM — which is limited, expensive, and the primary constraint on which models can run on a given machine.

Apple Silicon eliminates this distinction. The CPU, GPU, and Neural Engine all share the same memory pool. A Mac mini M4 Pro with 48GB of unified memory can use all 48GB for model weights — equivalent to a dedicated GPU with 48GB VRAM, which does not exist in the consumer market at anything close to the Mac mini's price point.

Cost comparison: cloud vs on-premise at scale

School size	Cloud API cost / year	Mac mini cost (once)	Break-even
200 students	AU$1,200–2,400	AU$2,000–3,000	12–24 months
400 students	AU$2,400–4,800	AU$2,000–3,000	6–12 months
600 students	AU$3,600–7,200	AU$2,000–3,000	3–5 months
1,000 students	AU$6,000–12,000	AU$2,000–3,000	2–4 months

The hardware cost does not increase with student count. The cloud cost scales linearly. For any school with more than approximately 200 students using the platform meaningfully, the on-premise model delivers cost savings within the first year.

Data sovereignty: the procurement argument cloud AI cannot make

Cost is the obvious argument. The more important one is data sovereignty.

When a school runs AI inference on a cloud API, the following things are structurally true regardless of what the vendor's privacy policy says: student data leaves the school network, is processed on servers the school does not control, is subject to the laws of the jurisdiction where those servers are located, and is handled under terms of service that can change at any time.

On-premise inference makes none of this true. The data is processed on a machine the school owns, on a network the school controls, subject to Australian law, under terms that cannot change because there is no third-party cloud provider in the loop.

How IndiLearn deploys on the school's Mac mini

IndiLearn's on-site deployment uses Ollama as the local inference server. Ollama exposes an OpenAI-compatible API at the local network address, meaning IndiLearn's application code communicates with the local model in exactly the same way it would communicate with a cloud API. The privacy guarantee is architectural, not application-level — no code change is required to move from cloud to on-premise.

The CoachingProvider abstraction in IndiLearn's codebase ensures this. The application never sees whether inference is running locally or in the cloud — it calls the provider interface and receives a response. The school's Mac mini configuration swaps one implementation for another without the teacher, student, or application being aware of the change.

Related articles