The Vatican Runs Its Own Private AI Model, and It Speaks Latin

A clip from the podcast This Week in Tech features a Catholic priest describing something most people would not expect from the Church: a fully private, locally-run large language model. No cloud, no third-party API, no data ever leaving the building. The conversation runs 87 seconds and stays grounded in specifics, which is what makes it land.

The priest explains that the Church runs its own models on dedicated hardware. He name-checks an Acer Veriton workstation built around Nvidia's Blackwell chip: roughly a petaflop of performance, 4 terabytes of storage, 128 gigabytes of memory. The point of the local box is control. Because nothing touches the internet, the model can be used for what he calls "internal forum" cases, sensitive information that needs to be summarized, translated, or transformed without ever being exposed.

The training data is the interesting part. The Church digitized several hundred years of texts specific to Catholic history and doctrine, then trained on that corpus. The result learned the language, the culture, and yes, Latin. The priest says teaching it Latin was one of the very first things they did, partly so it could help translate older sources. It is a completely closed model that draws on no outside information, and he makes a point of noting it will openly say "I cannot answer that" rather than fabricate.

The visual framing is a standard tech-podcast setup: split-screen between the host in a loud floral shirt and the priest in a clerical collar against a bookshelf backdrop, with TWiT branding and podcast platform badges along the bottom. On-screen captions track the conversation phrase by phrase.

Key Takeaways

• Private, on-premise LLMs are a real pattern for organizations with sensitive data, not just a theoretical option. The Vatican's "internal forum" use case (confidential material that legally and ethically cannot leave the institution) is exactly the scenario where a local model beats any cloud API.
• Domain-specific training on a proprietary corpus pays off. Several hundred years of digitized Catholic texts produced a model fluent in the institution's own language and history, including Latin, which no general-purpose model handles well.
• The "I cannot answer that" behavior is a deliberate design choice. For an institution where accuracy on doctrine matters, a model that refuses rather than hallucinates is a feature, not a limitation.
• Consumer-grade hardware now clears the bar. A single Blackwell-based workstation with a petaflop of compute is enough to run a serious private model, which lowers the barrier for any organization weighing build-vs-buy on AI.