Tacit and AI Sovereignty: Owning the Translation Layer
The easiest way to make AI sovereignty sound impossible in New Zealand is to define it as owning the entire AI stack.
We do not own the semiconductor supply chain. We do not operate a hyperscale cloud platform. We do not own a frontier foundation model.
This is the wrong definition.
Small countries like ours have exercised sovereignty under conditions of interdependence. The question in AI sovereignty is whether New Zealand can preserve meaningful decision-making power over the AI systems that shape our economy, public services, culture, and institutional capability.
For New Zealand, the right question is whether we can create leverage within a stack that will inevitably be shared with others.
Historically, that leverage comes from three places:
- Becoming useful to the global ecosystem.
- Preserving the ability to switch between external providers.
- Building internal capability to reduce permanent dependency.
This is the gap Tacit is designed to address.
The Translation Layer: Where Sovereignty Lives
Tacit is building a global platform for producing specialised language models (SLMs).
Tacit's role is to capture the unwritten, contextual human judgment - the tacit knowledge of our organisations - and build the translation layer between foundation models and trusted real-world deployment. Our platform captures human judgment, constructs training and evaluation environments, and produces SLMs that can be deployed and operated within each organisation.
This translation layer is where New Zealand can participate, preserve optionality, and build internal capability across three distinct horizons:
1. Owning the Workflow and Trust Layer
The AI ecosystem has many layers. Chips and frontier models sit near the base of the stack, but they are not where most organisations experience value. Someone still has to turn raw model capability into reliable services, trusted workflows, governed deployments, and measurable outcomes.
Frontier models are trained on explicit knowledge (the indexed internet: books, code, Wikipedia). What they lack is tacit knowledge: the deeply contextual, unwritten human judgment, cultural nuances (like Tikanga or local governance structures), and operational "know-how" embedded in New Zealand organisations.
New Zealand can be a proving ground for capturing this tacit knowledge. We are close-knit enough for fast feedback, digitally capable enough for serious deployment, and constrained enough to force practical thinking about sovereignty.
Xero did not need to own Windows, the browser, or the cloud infrastructure to build a globally relevant accounting platform. It owned the workflow and trust layer for a specific domain. The same logic applies to AI.
2. Achieving Provider Independence
Switching between providers requires the ability to measure what constitutes good AI. Without task-specific, statistically defensible evaluation capabilities, organisations are forced to accept vendors' claims.
Provider independence requires evaluation independence. Big Tech cannot be allowed to grade its own homework; the translation layer must remain strictly vendor-agnostic.
If New Zealand invests in independent evaluation capabilities, our organisations can compare models, negotiate from a position of strength, identify unacceptable risks, and switch between providers when cost, performance, trust and geopolitics change.
3. Securing the Data and Knowledge Boundary
New Zealand can and should encourage its organisations and industries to develop the capabilities to take mature workflows from frontier models into locally developed, hosted, and operated SLMs.
The point is not to reject frontier models, but to avoid renting them permanently for work that has become both stable and economically viable to own, while ensuring our private institutional data never leaves our sovereign boundaries.
Crucially, this creates both a security boundary and a talent boundary. The risk is not only that sensitive institutional data crosses the border through external APIs. It is that New Zealand's expertise stops compounding locally.
New Zealand has long worried about the visible brain drain: skilled people leaving the country. Frontier-model dependence creates a quieter version of the same problem, an unseen brain drain, where the judgment of our people is converted into capability elsewhere.
Locally developed, hosted and operated SLMs reverse that flow. They keep data, weights, execution and accumulated expertise within New Zealand's control.
The Global Blueprint
Veldhoven is a small town in the Netherlands, historically known for cigars and textiles. In 2021, its population was about 45,500, roughly a third of Dunedin's. It is also home to ASML, the only company in the world that produces EUV lithography machines. ASML does not design GPUs, manufacture chips or run cloud platforms. Yet the most advanced chipmakers depend on its machines. That is why a small town became a choke point in the global AI race: it owns a critical enabling layer of the semiconductor supply chain.
The parallel for Tacit is simple: a small place can become globally important when it owns a layer that the rest of the ecosystem needs. For New Zealand, that layer is the translation layer that encodes frontier-model capability into specialised, measurable and locally governed AI systems.
That is New Zealand's chance to contribute to the global ecosystem without owning the whole stack.
Build it here. Prove it under pressure. Take it global.