Module Four · Whakamana
Who controls the AI you rely on?
whakamana: to give authority, to give effect, to empower
Type a question into ChatGPT or Claude from a desk in Auckland and the work of answering happens on computers in another country, owned by an American company, running a program that is never sold, on chips designed in California and manufactured in Taiwan. Every link in that chain is controlled by someone. “Sovereign AI” is the question of who, and the phrase is now used to sell everything from data centres to legal software. This module explains each layer of control, shows who holds it today, and gives you the questions to ask of anyone selling sovereignty.
What “sovereign AI” names
People use the phrase to mean at least four different things. A conversation that has gone wrong has usually mixed two of them.
Data sovereigntyYour information stays in your country, under your country’s law.
The claim: what you type in, what comes back, and what gets stored all remain in New Zealand, where a New Zealand court decides who may see them. Think of a medical practice and its patient records. Whether a vendor can actually promise this depends on layers one and two of the ladder below.
Operational sovereigntyThe system keeps working if a foreign supplier changes its terms, its prices or its mind.
The claim: continuity. If an overseas provider doubles its prices, rewrites its rules or switches the product off, everything built on it stops with it. The contract layer decides how exposed you are. Holding the model yourself is the only complete protection.
Model sovereigntyYou hold the AI model itself, and can run it, inspect it and keep it.
The claim: you possess the model the way you possess software installed on your own machine, whatever its maker does next. Only the weights layer delivers this. The leading American models are never sold this way.
Capability sovereigntyYour country can build these systems itself: the chips, the data centres, the power and the people.
The claim: the whole pipeline, from silicon to trained model, exists at home. This is usually what national strategy documents mean. It requires chip factories, warehouse-scale data centres, electricity at industrial scale and thousands of specialist engineers. No small country holds it. As at 2026, no country holds all of it alone.
An organisation that has not decided which of the four it is buying will buy the wrong one.
The eight layers
To hold full sovereignty over an AI system you would need to hold every layer here. Each one opens.
The lawWhose courts can order your data handed over.
Nearly all of the big AI providers are American companies: OpenAI, Anthropic, Google, Microsoft, Amazon. In 2018 the United States passed a law called the CLOUD Act. It requires an American provider to hand over data in its possession, custody or control when a US court orders it, wherever in the world that data is stored.
So an Auckland data centre answers only part of the question. Storing your data in New Zealand brings the storage under New Zealand law and narrows day-to-day exposure. The company holding it remains answerable to American courts.
CLOUD Act, 18 U.S.C. § 2713 (2018). Checked 23 September 2026.
live sourcePerma.cc recordcaptured 23 September 2026
The processingWhere the computers that produce the answer actually sit.
Every AI answer is produced by physical computers in a data centre somewhere, doing billions of calculations. The industry calls this work inference. It happens wherever the provider has machines with capacity to spare, and that is often a different country from yours.
The New Zealand example, dated March 2026: Amazon’s Bedrock service went live in its Auckland region with Claude available. Under a feature called cross-region inference, the request goes to whichever data centre has capacity, and a setting chosen by the buyer decides how far it may travel. On the Asia Pacific setting, a request made in Auckland is handled in Auckland, Sydney or Melbourne. On the global setting, which costs about 10 per cent less, Amazon may send it to any of its commercial regions worldwide, the United States included. Your stored data stays in Auckland on either setting.
The cheaper setting is the one that carries the request furthest. A request processed in the United States brings layer one back into play.
AWS, 26 March 2026; AWS Bedrock cross-region inference documentation. Checked 23 September 2026.
live sourcePerma.cc recordcaptured 23 September 2026
live sourcePerma.cc recordcaptured 23 September 2026
The contractThe terms on which you rent access.
Nobody buys ChatGPT or Claude. Access is rented, through a subscription or a usage agreement, the way a business rents Microsoft 365. The provider sets the price, the rules, the model version and whether the service exists at all. Each of these can change on the provider’s schedule.
The pace matters here. The leading model has been superseded every few months since 2023. A deal built around one named model is out of date within a year.
The weightsThe model itself, as a file that can be owned.
Training an AI model produces, in the end, a single enormous file of numbers. The industry calls these numbers the weights. That file is the model. Whoever holds a copy can run it on their own computers, keep it forever, examine it and alter it, subject to the licence it was released under.
Some laboratories publish the file for anyone to download. Meta’s Llama models work this way, as do the leading Chinese models. The top American laboratories, OpenAI, Anthropic and Google, keep the file on their own servers and sell access to it. As at September 2026 the most capable weights you can actually download come from Chinese laboratories, mainly DeepSeek and Alibaba’s Qwen, under licences that permit commercial use.
DeepSeek and Qwen model licences, Hugging Face. Checked 23 September 2026.
live sourcePerma.cc recordcaptured 23 September 2026
live sourcePerma.cc recordcaptured 23 September 2026
live sourcePerma.cc recordcaptured 23 September 2026
The training dataWhat the model learned from, and whether anyone consented.
A model’s abilities come from the text, images and audio it was trained on. For the big models that means a very large share of everything publicly available on the internet, collected by automated scraping, mostly without asking the people who created it.
For New Zealand that includes te reo Māori. OpenAI’s Whisper speech model was trained on 1,381 hours of te reo Māori audio gathered from the web, on the paper’s own figures. Māori data sovereignty treats data like this as a taonga. It has its own page in this module.
The computeThe chips, the buildings and the electricity.
Training and running these models takes specialised processors, in the tens of thousands, housed in warehouse-sized data centres drawing electricity at industrial scale. The supply chain for those processors is one of the most concentrated in the world.
NVIDIA, an American company, designs the chips. It owns no factories. TSMC manufactures them, and its most advanced production lines are in Taiwan. The machines those production lines depend on come from one company: ASML, in the Netherlands. Over the whole chain sits US export control law, which decides which countries the most capable chips may be sold to. Every AI system used in New Zealand rests on this chain.
NVIDIA Form 10-K (FY2026); TSMC; ASML; US Bureau of Industry and Security. Checked 23 September 2026.
live sourceArchived copy pending: sec.gov blocks automated capture
live sourcePerma.cc recordcaptured 23 September 2026
live sourcePerma.cc recordcaptured 23 September 2026
live sourcePerma.cc recordcaptured 23 September 2026
The peopleThe engineers who can actually build these systems.
The number of people who have trained a top-tier model is small. Most of them work at a handful of laboratories in the United States and China, and skills at this level are bought on a global market.
A country wanting to build its own systems needs this layer alongside the hardware. New Zealand’s pool is small.
The frontier itselfThe best model in existence, which keeps moving.
“The frontier” means the most capable models in existence at a given moment. The lead has changed hands every few months since 2023. Hold all seven layers below and this one still moves on you.
Access to the top is also narrowing. Anthropic’s generally available model is Fable 5. Its most capable model is Mythos 5.1, released on 1 September 2026. Anthropic restricts that model to vetted cyber-defence and life-science organisations through its trusted access programmes. Those organisations are currently in the United States. A New Zealand organisation cannot obtain it on any terms.
Anthropic, 1 September 2026. Checked 23 September 2026.
live sourcePerma.cc recordcaptured 23 September 2026
No country holds every layer. The United States designs the leading chips and has them fabricated in Taiwan. China publishes capable weights and faces export controls at the compute layer. New Zealand holds part of the law layer and part of the data layer.
Sovereignty over these systems is held layer by layer. A purchase can secure a position on particular layers. Anyone selling sovereignty should name which ones.
Where to next
The New Zealand position
What New Zealand holds, layer by layer, what is actually on offer here, and the questions to ask anyone selling sovereignty. Dated September 2026.
ReadMāori data sovereignty
The doctrine that Māori data is a taonga, the institutions behind it, and the one New Zealand example of sovereignty achieved at the data layer.
ReadWhakamārama explains how these tools work. Whakatere covers what is safe to put into them.