Simulate the Nation: What would it take to build an AI Country Model (country-scale AI World Model) for New Zealand and Australia

Almost every country wants a digital twin. Almost none of them are asking for a world model. That distinction sounds pedantic until you look at what each one can actually do. A digital twin tells you what your country looks like right now, in extraordinary and growing detail. A world model tells you what your country would look like if something changed - a flood, a policy, a port closure, a decade of population drift - before any of it happens. New Zealand and Australia have both spent the last several years building the first thing. Neither has seriously proposed building the second. This article is about what it would actually take.

This one returns to a question the world models article raised but never pushed to its logical extreme: if a world model can learn to simulate a room, a warehouse, or a city block, what stops it from simulating a country? If you haven't read the rest of the series - From Code to Concrete, Data Residency, Data Sovereignty, and the Rise of Inference Sovereignty, Data Has Currency, Where Reason Ends and Act Begins, and the governance piece that followed it - the ideas in this piece lean on all of them: sovereign data, an internal data market, a generative architecture that actually understands cause and effect rather than pattern-matching to it, and an honest reckoning with what happens when that understanding starts shaping real decisions.

1. Digital Twin Is the Sense Layer. World Model Is the Reason Layer.

Virtual Singapore, launched in 2014, is still the reference point everyone reaches for: a 10-centimetre-precision 3D replica of an entire city-state, built from 25 terabytes of laser-scanned geospatial data.^[1] The UK's National Digital Twin Programme, running since 2018, and Australia's own Digital Atlas of Australia - a federated platform now connecting hundreds of curated national datasets on geography, population, economy, and environment across every level of government^[2] - are built on the same premise. So is the conversation New Zealand has been having since 2022, when Te Waihanga, the Infrastructure Commission, recommended accelerating investigation into city, regional, and national digital twins as a tool for spatial planning.^[3]

All of these are extraordinary achievements, and all of them are doing the same fundamentally descriptive job: representing what already exists, as accurately and currently as possible. That is Sense, in the language this series has used throughout - governed, high-fidelity data about the world as it is. None of them, on their own, can tell you what happens next. A digital twin can show you where every pipe in Auckland runs. It cannot tell you which fifty of those pipes will fail first under a given rainfall scenario, and why. That second question - the causal, generative, counterfactual one - is what a world model answers, and it is a categorically different piece of technology from a twin, not a more advanced version of one.

Here's how different this actually is from the generative AI most boardrooms are already comfortable with - and the gap is exactly where the opportunity sits. An LLM has read an enormous amount of text about the world and become extraordinarily good at predicting what word plausibly comes next - it knows what people have said about floods, but it has never seen water actually behave. A world model is trained on the world's own dynamics rather than on descriptions of them, learning what physically happens next rather than what is statistically likely to be said next.^[14] That is the entire difference, and it is also exactly the difference between a technology that can write a very convincing paragraph about a flood and one that can actually show a planner where the water goes.

2. Four Different Bets on What a World Model Should Be

What makes this the right year to raise the idea is that four really different approaches to world models have all moved from research to serious capital and real products in the last twelve months, and each one maps onto a different piece of what a country-scale World Model would need to be.

Odyssey - the world model company I covered in an earlier piece in this series, co-founded by New Zealander Jeff Hawke - builds causal, multimodal systems that learn to simulate the world frame by frame, generating minutes-long interactive video that responds coherently to intervention, in a way no LLM can, because it is not generating a description of an event, it is generating the event.^[11]

AMI Labs - Yann LeCun's venture, launched in March 2026 with Europe's largest-ever seed round at $1.03 billion - is built on LeCun's own public argument that language models are close to a dead end for anything that has to operate reliably in the physical world, because fluent text prediction and genuine understanding of cause and effect are not the same skill.^[12] Its world models learn directly from sensor data rather than from text or images, targeting exactly the industrial, robotic, and process-control settings where an LLM's confident, plausible-sounding wrongness is a genuine safety problem rather than an inconvenience.

World Labs - Fei-Fei Li's spatial intelligence company, which has raised $1.23 billion and shipped its first product, Marble, to commercial availability in February 2026 - starts from a similar critique to AMI Labs, aimed at a different gap. Li's own framing is that language models are "wordsmiths in the dark": eloquent, but with no grounding in physical space at all.^[13] Her taxonomy splits "world model" into three functions - a renderer that generates a consistent environment, a simulator that gives it physics and consequence, and a planner that turns perception into action - which is, functionally, Sense - Reason - Act by another name.

Prometheus, Jeff Bezos's venture, is the odd one out in the best way - and the one whose underlying architecture is, deliberately, the least publicly detailed of the four. What is known is the ambition rather than the mechanism: it is aimed not at simulating the world for its own sake but at compressing the loop between designing something and knowing its real-world consequence, turning a decade-long engineering cycle into something answerable in a fraction of the time.^[10]

None of these four companies is building a country model. Between them, though, they are building every component one would need - a generative renderer, an abstract sensor simulator, a physics-grounded planner, and a compressed design-consequence loop - which is precisely why this is a live engineering question now rather than a decade away.

3. The Proof of Concept Already Exists, at National Scale

This isn't a hypothetical - it's already working, in production, for one part of the world: the climate. NVIDIA has built a system called Earth-2 that can generate detailed, fast-running climate simulations, trained on fifty years of weather data, running far faster than the traditional models governments have relied on for decades.^[4] This isn't a lab demo. Taiwan's national weather agency is already using it, tuned specifically on Taiwan's own data, to predict exactly where typhoons will make landfall - feeding directly into real evacuation decisions for an island that has been hit by more than 136 typhoons since 2000.^[5]

Sit with what that actually is: a national government, in production, using a generative world model - not a dashboard, not a digital twin, a model that generates plausible futures on demand - to make real decisions about where to evacuate people. That is precisely the pattern this article is proposing be extended beyond climate. Climate-in-a-bottle is the proof of concept. The open question for New Zealand and Australia is not whether this class of technology works at national scale - Taiwan has already answered that for one domain - but whether either country is willing to extend the same generative approach to economy, infrastructure, ecology, and population, rather than leaving it parked at weather.

4. Two Distinct Opportunities

New Zealand and Australia are starting from really different bases, and each of those starting points is worth reading as an opportunity in its own right, rather than as two competitors on the same track.

New Zealand's opportunity is coherence. It is small enough, and governed simply enough - one national government, one set of national datasets, no state-level federation to negotiate - that a single, unified, whole-of-country world model is very much a realistic engineering target rather than an abstraction. LINZ (Toitū Te Whenua) already holds the national geospatial and cadastral backbone such a model would need as its foundation, and New Zealand's own digital twin conversation has recently taken what one practitioner called a deliberately "back to basics" approach, prioritising national data standards ahead of any specific technology bet - exactly the discipline a first-mover would need.^[6] The open question is less about capability and more about ambition: whether New Zealand chooses to fund a model of this scope itself, in partnership, or through sovereign-cloud arrangements that keep the sovereignty questions from the second article in this series firmly in view.

Australia's opportunity is scale and existing infrastructure. The Digital Atlas of Australia is already, by a wide margin, the most mature descriptive national data platform in the region - an Integrated Geospatial Infrastructure connecting geography, population, economy, and environment data across federal, state, and agency boundaries, with a live partnership between Geoscience Australia, the Australian Bureau of Statistics, and the Social Services portfolio specifically aimed at place-based policy.^[2] Australia also has the compute, capital, and critical-minerals-driven appetite for large-scale geoscientific modelling already in motion - Geoscience Australia's Resourcing Australia's Prosperity programme is producing continental-scale prospectivity models across 36 critical minerals right now.^[7] The opportunity here is less about proving the concept from scratch and more about extending infrastructure that already works: federation is a genuine coordination challenge, but it also means eight jurisdictions' worth of data, compute, and institutional appetite are already partially assembled, waiting to be pointed at a generative rather than purely descriptive goal. A voice from Australia's own digital twin community has been candid about the one thing that has to come first regardless: most Australian digital twin initiatives stall in what she called "pilot purgatory" precisely because governance work started after the technology work rather than alongside it.^[6]

Read together, the two opportunities are complementary rather than competing: New Zealand is positioned to prove that a single, coherent country model is buildable at all; Australia is positioned to prove what a federated system of models, built on already-mature existing infrastructure, can do once ontology and governance are solved. Either result would be useful to the other country.

5. What It Would Actually Take

Data. A country model needs to fuse categories of data that currently live in entirely separate institutional worlds. Geospatial and cadastral (LINZ, Geoscience Australia). Climate and atmospheric (NIWA, the Bureau of Meteorology). Geohazard (GNS Science, Geoscience Australia's own geohazard programmes). Population and economic (Stats NZ, the ABS). Infrastructure condition and capacity (Te Waihanga, state infrastructure bodies). And, critically, private-sector data that sits entirely outside government - utility networks, telecommunications load, insurance claims history, agricultural yield data - which is exactly the category of asset the internal data brokerage model from earlier in this series was built to govern, extended from an organisation to a nation.

Ontology. This is the least glamorous requirement and the one every existing digital twin initiative underestimates. Datasets from geology, economics, ecology, and infrastructure do not share a common semantic structure - a "risk" in an insurance dataset and a "risk" in a geohazard dataset are not the same object, and a model cannot reason causally across domains it cannot first describe in a shared language. Academic assessments of Singapore's own city-twin governance have flagged exactly this maturity gap: sophisticated visualisation sitting on top of fragmented, poorly reconciled underlying ontology.^[8] In terms of what actually determines whether a country model works, building that shared ontology is harder than building the model itself, and it is the step most national programmes skip in their rush to ship a visible map.

Governance. Everything the second and third articles in this series argued about data sovereignty and internal data brokerage applies here at national scale, with higher stakes. Where does inference actually run - onshore, sovereign cloud, or offshore compute leased from a hyperscaler? Who owns the output of a model trained on pooled agency and private-sector data - the contributing agency, the model operator, or the public the data was collected from? For New Zealand specifically, this cannot be treated as a generic data governance question: land, water, and geospatial data carry Treaty obligations that already shape how LINZ operates, including active integration work with the Māori Land Court's own land data systems.^[9] A country model that does not build Māori data sovereignty into its governance from the first design decision, rather than retrofitting it later, will not be a legitimate national asset - it will be a government IT project that happens to be very large.

6. The Provocation: What a Country Model Could Actually Do

This is where I want to really speculate, because the value of a generative country model is not a better map - it is the ability to ask "what if" and get a real answer. What happens to the national grid if a major earthquake takes out a single transmission corridor through the central North Island? What does a decade of accelerated climate migration actually do to regional housing demand, five years before it shows up in a census? What is the realistic economic and logistical consequence of a single Cook Strait ferry route being unavailable for six months? These are not hypothetical curiosities - they are exactly the kind of question Treasury, emergency management agencies, and infrastructure planners currently answer with static scenario planning and expert judgement, because no tool exists that can simulate the actual causal chain and show its work.

This is also the same ambition Jeff Bezos described for Prometheus in the previous instalment of this series - an AI system that compresses a design-and-consequence loop that currently takes a decade into something answerable in an afternoon.^[10] Prometheus is applying that idea to jet engines. There is no technical reason the same class of model, applied to national infrastructure and policy, could not do the same thing for the decisions that actually determine how well a country absorbs the next crisis - climate, seismic, or economic - rather than merely documenting it after the fact.

7. The Honest Risks

A model detailed enough to simulate a country accurately is also, by construction, a model detailed enough to simulate every person and institution in it - which is precisely the kind of capability this series has argued, repeatedly, needs governance designed in before deployment rather than bolted on afterward. A generative national model concentrates an extraordinary amount of inferential power over policy in whoever controls it, with limited external ability to audit why it produced a given output. It is also, unavoidably, a piece of critical national infrastructure in its own right - a single, high-value target whose compromise would be far more consequential than a leaked dataset. None of this is a reason not to build one. It is a reason the governance work has to run in parallel with the technical work from day one, not trail eighteen months behind it the way it typically has for every digital twin initiative that came before.

Conclusion: The Conversation Neither Country Is Having Yet

New Zealand and Australia have both done the unglamorous work of building descriptive national data infrastructure - LINZ's geospatial backbone, the Digital Atlas of Australia's federated data ecosystem, a decade of digital twin investigation on both sides of the Tasman. Neither has yet asked the next question out loud: now that we can see the country this clearly, are we willing to build something that can imagine its future as rigorously as it represents its present? Taiwan has already shown, for one domain, that the technology is ready before most governments think it is. The gap left in New Zealand and Australia is not technical. It is the willingness to have the ontology, governance, and sovereignty conversation before the demo, rather than after it - exactly the discipline this entire series has been arguing for, one pillar at a time.