The Sovereign AI Build — How New Zealand Could Own Its AI Stack on Renewable Energy and Open Models

Jack Dorsey says the real danger isn’t open source AI — it’s five CEOs deciding what the world is allowed to build with AI.

“These AI companies are building platforms and they’re all incentivized around their own particular models. You have to ask them for permission, you have to [pay them for access].” — Jack Dorsey, July 12 2026 (X/@uncover_ai)

Alex Karp, CEO of Palantir — the company with the deepest government intelligence contracts in the AI industry — went on CNBC and made the same case from the other side, with visible passion. Karp’s argument: when you pay OpenAI or Anthropic per token, you don’t own the weights, you don’t control the compute, and your proprietary data becomes training fuel for a model that also serves your competitors. CNBC anchor Becky Quick told him he sounded angry. Karp responded: “No, this is the voice of American business that is being channeled through me.”

(X/@atrupar)

Two very different men. One is a Silicon Valley founder arguing for open models. The other is a defence-tech CEO selling his own version of sovereignty. Same diagnosis. Different cure. Neither cure works for New Zealand.

This is Part 3 of a series. Part 1 argued that the NZ Super Fund should fund sovereign AI infrastructure. Part 2 argued for an AI coordination layer that connects NZ companies into consortiums, using the America’s Cup model. Part 3 is the build — and it lays out three options for NZ. The worst, the middle, and the best.

🔍 THE BOTTOM LINE

New Zealand has every ingredient to build a sovereign AI stack: 85-90% renewable electricity heading to 98% by 2030, a $70B data centre opportunity already being scouted by Invest New Zealand, open-weight models that can be fine-tuned on NZ-specific data, a Super Fund with $93 billion and a strong track record of NZ investment, and a Māori data sovereignty framework that provides governance the rest of the world doesn’t have. The question is which path NZ takes — and the Dorsey/Karp framing shows why the obvious options are both dead ends.

Three Paths — Worst, Middle, Best

When a NZ drone company wants to fine-tune a model on its flight telemetry data, or a Christchurch hydrogen company wants AI optimisation for its fuel systems, or a government agency wants a model that understands NZ law, they have three options. The gap between them is the gap between renting, buying, and owning.

🔴 WORST — Big AI Decides

You pay OpenAI, Anthropic, or xAI per token. Your data goes to the US. The model weights stay with the lab. Your fine-tuned model serves you and everyone else through the same API. You don’t own the weights. You don’t control the compute. If the US government decides to restrict access — as it did with Anthropic’s Fable and Mythos models through export controls — your AI capability gets switched off overnight, with no notice, no appeal, and no NZ input.

This is Dorsey’s five CEOs. It’s also the default state for most NZ organisations today. Every government agency using ChatGPT, every startup building on Claude’s API, every university running research on GPT-5 — they’re all renting AI from US companies that can change the terms, cut the access, or train on their data at any time. Big AI decides what you can build, where your data goes, and whether you’re allowed to keep using the tool your business depends on.

🟡 MIDDLE — Sovereign AI Platforms

You bring your data to a platform like Palantir’s. They build you a model you actually own, hosted on infrastructure they control, using open-weight models like NVIDIA’s Nemotron. You’re not renting per token. You’re not sending data to a consumer API. You have a model fine-tuned for your specific use case, running on dedicated infrastructure.

This is a real step up from the worst option. Karp’s pitch works because it solves a genuine problem — organisations that need AI capability without handing proprietary data to a consumer API. Defence agencies, intelligence communities, large enterprises with sensitive IP. Palantir’s sovereign AI platform gives them a model they own, on infrastructure they trust, without the per-token dependency.

The middle option isn’t sovereignty, though. The model runs on Palantir’s infrastructure. The stack is Palantir’s. The compute is Palantir’s. You’ve traded five CEOs for one platform. That’s better — but it’s not independence. For a NZ government agency or iwi organisation, “sovereign AI hosted by a US defence contractor” is a contradiction in terms. The data may be in a sovereign environment, but the platform, the updates, the security patches, and the roadmap are all controlled from outside NZ.

The middle option is the right choice for organisations that need dedicated AI infrastructure but can’t build their own. It’s a stepping stone, not a destination.

🟢 BEST — NZ Owns It, Controls It, and It’s Ours

You run open-weight models on domestic GPU infrastructure, powered by NZ renewable energy, fine-tuned on NZ-specific data, governed by NZ law including Māori data sovereignty principles, funded by the NZ Super Fund, with NZ companies getting priority access and foreign companies paying premium.

You own the compute. You own the fine-tuned weights. You own the energy that powers it. You set the governance. Nobody in Washington or Silicon Valley can switch it off, change the terms, or restrict your access.

This is the best option — and it’s the one the rest of this article is about.

The Energy Layer — Renewable Compute Already Exists

New Zealand generates 85-90% of its electricity from renewable sources — hydro, geothermal, wind, solar. The government targets 98% by 2030. All electricity generation in the South Island is renewable. This is not a theoretical advantage — it’s the reason Datagrid chose Southland for its AI data centre.

Datagrid’s Southland campus — 280MW AI data centre in Makarewa near Invercargill. In March 2026, Datagrid signed a 15-year, 140MW power purchase agreement with Mercury — one of NZ’s largest commercial power deals. According to IT Brief New Zealand, the agreement covers 1.2TWh per year, equivalent to 3% of national electricity demand. Construction started June 2026, full delivery targeted for 2028.

Datagrid CEO Rémi Galasso positioned the site for “international AI players” — foreign companies renting NZ’s renewable energy. That’s the Datagrid model: NZ provides the green power, foreign companies provide the workloads, the profits leave.

But the infrastructure doesn’t care who owns it. The same 140MW of renewable-powered GPU compute that Datagrid is building for foreign AI companies could serve NZ companies first. The energy is NZ’s. The cooling advantage is NZ’s. The stable grid is NZ’s. The question is who the compute serves.

Invest New Zealand gathered 100 business leaders in Auckland in July 2026 to pitch $25-35 billion in data centre investment, as The Spinoff reported. The panel — Mercury CEO Stew Hamilton, Contact Energy CEO Mike Fuge, tech leader Helen Robinson, and BCG’s Kelly Newton — called it a “once-in-three-generations opportunity.” Fuge compared it to the invention of frozen meat shipping in the 1870s.

He’s right about the scale. He’s wrong about the direction. The frozen meat shipping boom was about NZ owning the supply chain — NZ farms, NZ processing, NZ shipping, NZ exports. The data centre pitch is about NZ providing real estate for foreign companies. The model should be inverted: NZ owns the compute, NZ companies get priority, foreign companies pay premium for the surplus.

The Model Layer — Open Weights, NZ Knowledge

The open-weight revolution has delivered frontier-class models that anyone can download, run, and fine-tune. Llama 4 (Meta), DeepSeek V4, Nemotron (NVIDIA), and GLM-5.2 (Zhipu) are all available as open weights.

None of these models know NZ law. None know the Resource Management Act, the Privacy Act 2020, the Holidays Act, or the Commerce Act. None know NZ health system protocols, Pharmac funding pathways, or Te Whatu Ora governance. None understand te reo Māori at depth, or NZ English idioms, or iwi governance structures. None know NZ agriculture regulations, MPI biosecurity protocols, or Fonterra’s supply chain.

Fine-tuning these models on NZ-specific data creates models that are more useful for NZ applications than any frontier model trained on global data. A model fine-tuned on NZ legislation, NZ court decisions, NZ regulatory guidance, and NZ business data is a sovereign asset. It doesn’t need to query a US API. It doesn’t send data offshore. It runs on NZ compute, under NZ law, answering NZ questions.

The te reo Māori AI database case shows the stakes. IEEE Spectrum reported on the world-leading te reo Māori AI voice project — a database of Māori language and voice data that is culturally irreplaceable. Newsroom reported that current law can’t adequately protect it. If that data feeds a foreign company’s model, NZ loses control of its own cultural IP. If it fine-tunes a domestic open-weight model running on NZ compute, the model, the data, and the output stay under NZ jurisdiction.

The Data Sovereignty Layer — Te Mana Raraunga as a Feature

New Zealand has something no other country building sovereign AI has: a mature Indigenous data sovereignty framework.

Te Mana Raraunga, the Māori Data Sovereignty Network, has established principles for how Māori data should be governed: data as taonga, iwi governance over cultural data, consent-based data sharing, collective benefit. These aren’t constraints on AI development — they’re a governance model that most of the world hasn’t developed yet.

The global sovereign AI conversation, as Tech Policy Press argues, is about “what parts of the AI supply chain must a nation own, control, or govern, and what parts can a nation safely partner with, rent, or share.” NZ’s answer includes a layer other countries don’t have: Māori data governance. That’s not a liability. It’s a governance advantage — a framework for consent-based, community-owned data sharing that produces better data quality and stronger social licence than the extract-everything model.

Three layers of protection:

Compute-level — NZ-based GPU clusters mean data never leaves NZ borders. Government data, health data, iwi data, company proprietary data — all processed domestically. No foreign jurisdiction access, no US export control exposure.
Model-level — The fine-tuned weights are the asset. They stay in NZ, governed by NZ law. The model, the data, and the output all stay domestic.
Company-level — The consortium model from Part 2. Each company brings its IP but retains ownership. The AI coordination layer maps what each company can contribute without exposing trade secrets — it shows capability, not blueprints. SYOS brings its autonomy stack. Core Builders brings composite expertise. Fabrum brings hydrogen fuel systems. The AI shows how they could build a flying car together without any company surrendering its IP.

The Super Fund as Anchor Investor

We argued in Part 1 that the NZ Super Fund — at $93 billion, the world’s best-performing sovereign wealth fund over 20 years — should fund sovereign AI infrastructure. Part 3 fills in the model:

The entity: A private company, backed by Super Fund capital, that builds and operates NZ’s sovereign AI infrastructure. Not a government department. Not a state-owned enterprise. A commercial company with a national strategy mandate, structured to deliver returns to the Fund.

The business model:

Infrastructure leasing — GPU compute capacity leased to NZ companies at cost-plus, to foreign companies at premium. The renewable energy advantage makes NZ compute cheaper to operate than US or European alternatives.
Model fine-tuning services — Fine-tuning open-weight models on NZ data for government agencies, iwi organisations, and companies. The fine-tuned models are owned by the client, not the infrastructure company.
Consortium coordination — The AI coordination layer from Part 2. The company takes equity in consortiums it forms. Core Builders gets into aerospace. Fabrum gets into eVTOL. SYOS gets access to hydrogen propulsion. The coordination company takes a share for connecting the dots.
Sovereign AI as a service — For smaller NZ companies that can’t afford their own GPU time, a subsidised tier provides access to fine-tuned NZ models via API. The data stays in NZ. The pricing is transparent.

The Super Fund’s role: Anchor investor. Not picking winners — funding the infrastructure that lets winners emerge. The Fund already invests in NZ infrastructure, venture capital (Movac), and sustainable finance. The mandate exists. The product doesn’t. This is the product.

The Super Fund doesn’t need to build a frontier model. It needs to build the compute, host the open-weight models, provide the fine-tuning infrastructure, and operate the coordination layer. The base models are open. The GPUs are purchased. The energy is renewable. The data is NZ’s. The governance is NZ’s. The returns go to NZ’s retirement fund.

The Full Arc

Part 1 (The Super Fund argument): Why sovereign AI infrastructure is a strategic investment for NZ.

Part 2 (From America’s Cup to Flying Cars): What the AI is for — an industrial coordination layer that connects NZ companies into consortiums.

Part 3 (this article): How to build it. Three paths — worst (Big AI decides), middle (sovereign AI platforms), best (NZ owns it). Renewable energy powers domestic GPU compute. Open-weight models fine-tuned on NZ data. Māori data sovereignty provides the governance framework. The Super Fund provides the capital.

The middle option is a genuine step up from the worst. But it’s not sovereignty — it’s a better landlord. The best option is the one where NZ owns the compute, owns the fine-tuned models, owns the energy that powers it, sets the governance, and keeps the profits. Nobody in Washington or Silicon Valley gets to decide what NZ companies can build. Nobody gets to switch it off.

Dorsey is right that five CEOs controlling AI is a danger. Karp is right that renting your AI is renting your future. The best option is the one where NZ doesn’t have to ask anyone’s permission at all.

Not AI as control. AI as sovereignty. Not the thought police. The infrastructure builder. Not the restriction. The capability.

The pieces are on the board. The energy is in the grid. The models are open. The question is whether NZ sees what it already has — and builds it before someone else rents it back to us.

❓ FAQ

What’s the difference between the middle and best options? The middle option gives you a model you own on someone else’s infrastructure. The best option gives you a model you own on your own infrastructure, powered by your own energy, under your own law. The middle option is a stepping stone for organisations that can’t build their own stack. The best option is the national infrastructure layer that makes sovereignty real.

Is this just industrial policy dressed up as investment? No. Industrial policy picks winners. The AI coordination layer doesn’t pick — it sees. It shows what’s possible and lets companies decide. The Super Fund funds the infrastructure, not the outcome. The difference between picking winners and building the field they play on.

Can NZ actually compete with US hyperscalers on compute? Not on scale. On cost and sovereignty. NZ’s renewable energy makes compute cheaper to operate. NZ’s jurisdiction makes data legally protected. NZ companies don’t need to beat AWS on scale — they need to own enough compute to run their own models on their own data under their own law.

Why open-weight models instead of building a frontier model? Building a frontier model costs billions and takes years. Fine-tuning an open-weight model on NZ data costs thousands and takes weeks. The base model doesn’t need to be NZ’s — the fine-tuning does.

How does Māori data sovereignty actually work in practice? Te Mana Raraunga principles: data about Māori is governed by Māori. Consent-based sharing. Collective benefit. In practice, te reo Māori data feeds a model only with iwi consent. Cultural data is governed by the iwi that owns it. The infrastructure provides the compute; the governance model determines what data flows through it. This is a feature, not a constraint — it produces better data quality and stronger social licence.

What about the Datagrid Southland campus — isn’t that already sovereign AI? Datagrid is building the compute. But Datagrid is positioning it for international AI players, not NZ companies. The infrastructure is in NZ; the primary beneficiaries are foreign. Sovereign AI means NZ companies get priority access to NZ-powered compute, not just foreign companies renting NZ’s renewable energy.

Could the Super Fund actually do this? The Fund already invests in NZ infrastructure, venture capital, and sustainable finance. Its mandate includes investing in New Zealand. An AI infrastructure company is a new category, not a new mandate. The Fund provides capital; the company builds and operates the compute; NZ companies lease capacity; foreign companies pay premium. The returns go to NZ’s retirement fund.

📰 Sources

— CJ Murden, editor of Singularity.Kiwi. Former digital technologies teacher, author of AI-focused books. Writing with a New Zealand focus.