Southland’s AI factory: what it could mean for AI, data sovereignty and Māori

Southland is becoming part of the global AI infrastructure story in a very physical way. Datagrid New Zealand has received resource consent for a hyperscale data centre at Makarewa, north of Invercargill, promoted by the developer as New Zealand’s first “AI Factory”. Southland District Council describes six data halls across three modules and up to 240 MW of IT capacity for AI training, data processing and storage. Datagrid describes a 280 MW hyperscale campus, together with a high-capacity international fibre connection through the Tasman Ring Network.

Those are extraordinary numbers by New Zealand standards. The project is important. It could create substantial new computing capacity in Aotearoa and give New Zealand a larger role in global artificial intelligence, high-performance computing and cloud infrastructure.

But an AI data centre in Southland is easy to misunderstand. The most important distinction is this: a data centre used to train an AI model is a compute location. It does not tell us where the training data came from, who has authority over those data, or whether the resulting model represents a New Zealand or Māori worldview.

A model could be trained in Makarewa using data gathered from servers and websites around the world. The GPUs would be in Southland. The training corpus could still be overwhelmingly international.

What has actually been approved

Southland District Council says Datagrid NZ Partnership Limited has received the necessary resource consents to establish and operate the hyperscale facility at 342 and 370 Flora Road East, Makarewa.

The consented development includes six data halls across three modules covering about 9.5 hectares, up to 240 MW of IT capacity, a Grid Exit Point substation, a cable landing station, cooling and water-treatment infrastructure, stormwater systems, internal roads and support buildings, and 84 emergency backup generators. It also includes telecommunications infrastructure associated with the Tasman Ring Network and a subsea connection to Australia via Ōreti Beach.

Datagrid describes the wider campus as a 49-hectare project with 280 MW of hyperscale capacity. Its March 2026 announcement said the project had received full resource-consent approval and was purpose-built for AI training, inference and high-performance computing.

The difference between the Council’s 240 MW IT-capacity figure and Datagrid’s 280 MW campus figure is worth retaining rather than trying to force into one number. They appear to describe different parts or definitions of capacity. For public explanation, both should be attributed to their source.

See Southland District Council’s Datagrid resource consent page and Datagrid’s resource consent announcement.

The power requirement

AI training requires enormous amounts of electricity because modern training systems can use thousands of high-end GPUs continuously for weeks or months. Inference for widely used models can also become a huge ongoing workload.

Datagrid and Mercury announced a 140 MW long-term power purchase option arrangement in March 2026. Datagrid described the agreement as supporting AI, high-performance computing and sovereign-cloud workloads at the Southland campus. The developer says the 140 MW arrangement corresponds to around 1.2 TWh a year, a very large industrial electricity load.

Datagrid markets the Southland location around New Zealand’s renewable generation, southern climate and potential natural-cooling efficiencies. Those advantages are real considerations in data-centre economics, although the eventual environmental performance will depend on the final technical design, utilisation, cooling system and electricity supply over time.

See Datagrid’s Mercury agreement announcement and Mercury’s operating information.

AI training is computation, not a warehouse of New Zealand data

The phrase “AI training centre” can create an image of a building being filled with New Zealand information and teaching a model about Aotearoa.

That is not what the term means.

Training occurs when computing hardware processes a training corpus and adjusts model weights. The data may be stored locally, streamed from another data centre, copied temporarily into the facility or assembled from many sources. The physical location of the GPUs does not establish the origin of the data.

Imagine a future customer leases a large GPU cluster in Southland to train a global language model. The training corpus might contain Common Crawl, books, source code, images and licensed datasets from many countries. The computation could occur entirely in New Zealand while only a tiny fraction of the training material relates to New Zealand.

The resulting model would not become a New Zealand model simply because its electricity came through a Southland substation.

That distinction matters when words such as “sovereign” begin to appear around data centres.

Data residency is not data sovereignty

Keeping information physically in New Zealand can be valuable. It may simplify some jurisdictional questions, support resilience and reduce dependence on overseas infrastructure. For organisations with strict data-residency requirements it can be a major advantage.

But data residency answers only one question: where is the information physically stored or processed?

Data sovereignty asks a much larger set of questions. Who has authority over the information? Who determines its purpose? Who can access it? Can it be copied elsewhere? Who controls administrators and encryption keys? Which company owns the infrastructure? Which contract applies? What legal jurisdictions can reach the data? Can the information be used to train another model? Who controls derived outputs?

For Māori data, those questions extend further into collective authority, tikanga, whakapapa, benefit, representation and appropriate use.

A server rack located in Southland does not answer those questions automatically.

Infrastructure sovereignty is still important

None of this reduces the importance of local computing infrastructure. Quite the opposite.

One of New Zealand’s existing weaknesses in AI is dependence on computing infrastructure located elsewhere. Large AI workloads require expensive GPUs, high-capacity networking, reliable electricity and specialised cooling. Very few New Zealand organisations can build serious AI compute independently.

A large domestic AI facility could change what is technically possible. New Zealand organisations might be able to run high-performance workloads closer to home. Government, research institutions and businesses could potentially obtain local compute. Organisations requiring New Zealand-based inference might gain additional choices.

The significant question for Māori is whether Māori organisations can participate in those choices rather than merely having a very large global AI facility built in Aotearoa.

What a Māori sovereign AI service could look like

There is no public evidence at present that the Datagrid project includes a dedicated Māori AI service or Māori data-governance architecture. That should be stated clearly.

But the infrastructure raises an interesting future possibility.

A Māori-controlled or Māori-governed computing environment could lease or operate dedicated hardware inside New Zealand. Māori organisations could retain their own document stores and encryption. Locally approved open-weight models could run on dedicated GPUs. Access controls could be set by participating organisations. Sensitive source collections could remain separate. Logging and retention could be designed around the kaupapa rather than inherited from a global consumer chatbot.

This would not require creating a Māori equivalent of OpenAI. It could involve operating existing open-weight models on infrastructure whose deployment and information flows are much more tightly controlled.

That is a realistic form of infrastructure sovereignty.

Storage and compute are coming together

Te Pā Tūwatawata provides a useful parallel. It demonstrates a Māori-owned approach to distributed data storage, with emphasis on control over location, governance and infrastructure.

The logical next step in the AI era is computation. Storage determines where information rests. Compute determines where the information is processed.

If Māori organisations increasingly control storage but must send their most important information into offshore AI services whenever they want advanced processing, part of the sovereignty problem remains unresolved.

Conversely, Māori-controlled storage combined with Māori-governed compute and locally operated open-weight models would create a much stronger technical position.

That is not automatically a role for Datagrid. It is a strategic opportunity that facilities of this scale make easier to imagine.

AI inference may matter more than training

There is also a tendency to focus on headline-grabbing model training because it consumes spectacular amounts of compute.

For most Māori organisations, inference is more immediately relevant.

They are unlikely to train a frontier foundation model from scratch. The cost would be enormous and the benefit questionable. But they may want to run existing models privately over Treaty documents, environmental reports, organisational knowledge, GIS metadata or approved archives.

Inference infrastructure could therefore be more strategically useful than participation in global pre-training.

A shared GPU service in New Zealand could allow organisations to run large open-weight models that would be impractical on individual desktop machines. The key governance question would be whether those workloads are genuinely isolated and whether the source information remains under the organisation’s authority.

GIS workloads are part of the opportunity

AI infrastructure is not only about chatbots.

High-performance computing and GPUs can support remote-sensing analysis, LiDAR processing, computer vision, environmental modelling, change detection and large geospatial datasets. AI-assisted mapping can involve detecting features in imagery, classifying land cover, processing historic map scans or analysing large collections of field photographs.

For iwi and environmental teams, local compute could support taiao monitoring, erosion analysis, habitat mapping, flood modelling, vegetation classification and other workloads where the spatial data themselves may be sensitive or strategically important.

Again, the location of the computer is only one part of the architecture. The source imagery, derived layers, model outputs and access arrangements still need appropriate governance.

The environmental debate is not finished

A facility of this scale also has ordinary physical-world consequences. Energy demand, noise, cooling, generators, water, construction effects and landscape impacts remain legitimate issues regardless of the attractiveness of the phrase “AI factory”.

On 4 September 2026, Southland District Council announced that an independent commissioner had declined a request to initiate a review of the project’s consent conditions. The request raised concerns including low-frequency noise, cooling-system noise, acoustic modelling, substation and generator noise and whether additional conditions were needed. The commissioner found that the statutory threshold to initiate the review had not been met. Council says normal compliance monitoring and enforcement responsibilities remain.

This should not be portrayed either as proof that every environmental concern is resolved or as evidence that the consent is defective. It is a regulatory decision about whether the legal threshold for reopening conditions was met.

See the Council’s 4 September 2026 decision.

The more useful sovereignty question

The Southland project invites two very different stories.

One story says New Zealand will host global AI infrastructure powered by relatively renewable electricity and connected to international networks. That could bring investment, jobs, connectivity and new computing capability.

The other asks who gets to use that capability and on what terms.

For Māori data sovereignty, the second question is more interesting.

Can Māori organisations obtain domestic AI compute without surrendering control of their information? Can infrastructure contracts recognise collective data authority? Can dedicated environments be independently audited? Can locally hosted open-weight models operate over Māori-controlled storage? Can data be prevented from entering unrelated model-training pipelines? Can the people governing the information decide where the workload runs?

Those questions move Māori data sovereignty beyond storage and into the AI infrastructure layer.

What Southland does and does not prove

The Datagrid project demonstrates that large-scale AI compute in New Zealand is becoming technically and commercially plausible.

It does not prove that global models trained there will contain more New Zealand data. It does not create Māori authority over training datasets. It does not make an AI model culturally local. It does not by itself provide Māori data sovereignty.

But it may create infrastructure on which much more locally controlled AI can eventually be built.

That distinction is worth holding onto. The opportunity is not that the computers happen to sit in Southland. The opportunity is that Aotearoa may soon have enough computing capacity to make new architecture choices possible.

The final article in this series brings the pieces together: What this means for Māori GIS: from protecting data to controlling the AI architecture.