“Half our compliance and every timesheet now runs through some AI tool, and I couldn't tell you which country it sits in or who could pull the plug. If the price triples or the login stops working one morning, that's my business, not theirs.”
— WA drilling contractor, $28M revenue, 110 workers
Should Australia Build Its Own LLM? What Sovereign AI Actually Means for a WA Mining Supplier
The question Australia keeps arguing about
Few questions in Australian technology policy generate as much heat and as little resolution as this one. On one side: building a large domestic language model is an expensive distraction from the real work of putting AI to use. On the other: a nation without its own AI capability carries a strategic dependency it cannot fully see or control. Both positions contain genuine insight, and neither is the whole answer.
The clearest version of the sceptic's case came from Telstra chief executive Vicki Brady, who said plainly that Australia does not need to build large language models — it is a global game, and the real opportunity is applying AI to problems that are uniquely ours: bushfires, floods, protecting the Great Barrier Reef. That is commercially coherent. The trouble is it answers a narrower question than the one that actually matters.
The capital reality: the frontier is off the table
Start with what everyone agrees on. Competing at the frontier — training a model to rival the best from OpenAI or Anthropic — needs capital at a scale Australia has never contemplated for AI. Europe's best-funded builder, France's Mistral, has raised only a small fraction of what the leading US labs have pulled in, and those US figures run into the tens and hundreds of billions. No plausible Australian commitment closes that gap. On the pure economics, the sceptics win: a taxpayer-funded moonshot to out-build Silicon Valley is not a serious plan.
But “we can't win the frontier” is not the same as “do nothing”. The case for sovereign capability was never really about beating OpenAI. It is about control, transparency, and what happens when a foreign technology company's interests diverge from Australia's. Frame it that way and the question changes from “build a model?” to “which layers of the stack must we control to stay resilient?”
Three choices, not two
The public debate is stuck on a binary — build one or buy from abroad. There is a third option that is the one most businesses actually face: control the application and data layer while staying free to swap the model underneath.
Sources: Australian Government National AI Plan (industry.gov.au, Dec 2025); Senate Select Committee on Adopting AI (aph.gov.au); Expectations of data centres and AI infrastructure developers (Mar 2026); public company disclosures. Figures are latest-available and approximate.
Where the sovereignty argument actually bites
Australia has almost no visibility into how offshore models were built, on what data, or with what assumptions baked in. That is uncomfortable, but the sharper exposure is legal and commercial:
- The US CLOUD Act lets American authorities compel data held by US-linked providers no matter where the servers physically sit — so “the data is in a Sydney data centre” is not the protection it sounds like if the provider is US-owned.
- Access can be withdrawn or repriced for reasons that have nothing to do with you — export controls, a policy change, a commercial pivot, or a model being retired.
- Models trained mostly on US data encode US legal, cultural and safety assumptions you cannot renegotiate — and that you inherit whether they fit Australian conditions or not.
This is why some Australian builders have stopped using the word “sovereign” loosely. An AI is only as sovereign as its weakest layer — and an Australian-branded company that is foreign-owned is not sovereign in any way that matters when a foreign court comes asking.
The risk hiding in plain sight: concentration, not just cost
The debate usually assumes Silicon Valley is a stable, benign utility that will always be there. That assumption is doing a lot of unexamined work. Frontier capability is concentrated in a handful of US firms whose economics — enormous capital burn, circular vendor financing, unproven unit costs — carry real fragility. Depending on suppliers that might consolidate, pivot or stumble is a strategic exposure, not just a purchasing decision.
Read that way, the argument flips. The smart play is not to match US spending — it is to make sure Australia (and Australian businesses) can stand up a credible alternative quickly if access is ever cut. Sovereignty as insurance and optionality, not sovereignty as a trophy.
What Australia is actually doing in 2026
The good news is the policy has already moved past the binary. Australia's National AI Plan (December 2025) pulled together more than A$460M of AI funding, and the Senate's Adopting AI committee recommended against a government-built national LLM and for incentivising private sovereign capability instead. In March 2026 the government tied data-centre approvals to a social-licence test: operators are expected to give Australian startups and researchers access to compute and to build local engineering capability.
Meanwhile the private sector is already building. Maincode has trained an Australian model (Matilda) and self-funded a Melbourne “AI factory” on its own hardware, and Sovereign Australia AI is building foundational models (Australis, plus the open-source Ginan) on onshore GPUs. Neither is trying to beat OpenAI at the frontier — both are building the layers Australia can realistically own.
Achievable policy responses
If the goal is resilience rather than a trophy, the achievable moves are clear — and most of them cost far less than a frontier model.
Do this
- Fund the layers, not a flagship: back sovereign compute, low-carbon energy, clean data and skills, and let the private sector build the models.
- Legislate a real definition of “sovereign” — ownership, board, tax residency, and legal reach (CLOUD-Act-immune) — and make it a procurement gate, not a slogan.
- Use government and GBE procurement as the anchor buyer: a guaranteed demand floor is worth more to Australian builders than another grant round.
- Back “sovereign-ready” open-weight capability — the ability to fine-tune and host the best open models onshore — as insurance if offshore access is ever withdrawn.
- Mobilise the two real advantages: distinctive Australian data (health, geospatial, environmental) and cheap low-carbon power for compute.
Avoid this
- A taxpayer-funded moonshot to out-spend OpenAI at the frontier — the capital gap is an order of magnitude too wide, and the Senate committee said as much.
- Treating “sovereign” as a marketing label while the entity is foreign-owned and still reachable under the US CLOUD Act.
- Assuming Silicon Valley is a stable, neutral utility — capability is concentrated in a handful of firms whose interests need not match Australia’s.
- Ignoring the talent constraint: subsidise a model, lose the engineers to an offshore acquirer, and the capability leaves with them.
The single highest-value, lowest-cost move is a legal definition of “sovereign” used as a procurement gate. Once government only buys AI that is genuinely Australian-controlled and CLOUD-Act-immune, the market builds the rest — because there is finally a demand floor worth building for.
What this means for a WA mining supplier
You are never going to build a language model, and you do not need to. But you make a sovereignty decision every time you adopt a tool. For a $10M–$50M drilling, labour-hire, fabrication or mechanical-services business selling into BHP, Rio, FMG, Roy Hill and FQM, the question is not “should Australia build an LLM?” It is: where do your timesheets, incident reports, tickets and compliance packs actually live, and whose law can reach them?
That is the layer you control. The resilient choice is an Australian-built application layer that keeps your operational and compliance data onshore, under Australian law, and is not welded to a single foreign model that could be repriced or switched off overnight. It is the same distinction we drew in From Hierarchy to Intelligence — whether AI is a bolt-on chatbot or the system your desk actually runs on.
Sovereignty, made concrete
Your data stays onshore
Timesheets, FIFO records, incident notifications and compliance packs held in an Australian instance under Australian law — not a black box in a foreign jurisdiction
No single point of foreign control
The application layer is yours; the model behind it can be swapped, so a repricing or a switch-off offshore does not stop your back office
The productivity is still yours
$555K–$1.1M+/year of recovered admin labour across the 17-agent suite — sovereignty and ROI are not a trade-off
You cannot build a national LLM. You can decide your data never leaves the country.
Frequently Asked Questions
Should Australia build its own large language model?
The honest answer is: not a taxpayer-funded frontier model, but also not nothing. Competing with OpenAI or Anthropic at the frontier needs capital an order of magnitude beyond anything Australia has committed — Europe’s best-funded builder, Mistral, has raised a fraction of what the US leaders have. Australia’s December 2025 National AI Plan and the Senate’s Adopting AI committee both landed in the same place: government should fund the foundations (compute, energy, data, skills) and incentivise private builders, rather than run its own LLM. The sharper question is which layers of the AI stack Australia must control to stay resilient — and that is a question every business, not just Canberra, answers for itself.
What does “sovereign AI” actually mean?
Sovereignty runs in layers: the data centre, the compute hardware, the model, the applications, and the governance around them. An AI is only as sovereign as its weakest layer. A common trap is an Australian-branded company that is foreign-owned and therefore still reachable under the US CLOUD Act — which lets US authorities compel data held by US-linked providers wherever it sits. Genuine sovereignty means Australian ownership, an Australian board, profits and tax staying here, and data held under Australian law. For most businesses the layer they actually choose is the application and data layer — where their records live and whose law protects them.
Is Australia already building sovereign AI capability?
Yes, mostly through the private sector. The National AI Plan (December 2025) brought together more than A$460M of AI funding and, in March 2026, the government published Expectations of data centres and AI infrastructure developers that tie a project’s social licence to giving Australian startups and researchers access to compute. On the model side, private builders are moving: Maincode has trained an Australian model (Matilda) and self-funded a Melbourne “AI factory”, and Sovereign Australia AI is building foundational models (Australis and the open-source Ginan) on onshore hardware. The government deliberately did not launch its own national LLM.
What does the sovereign-AI debate mean for a WA mining supplier?
You are never going to build a language model, and you do not need to. But you do make a sovereignty decision every time you pick a tool: where do your timesheets, incident reports, tickets and compliance packs actually live, and whose law can reach them? The resilient choice is an Australian-built application layer that keeps your operational and compliance data onshore, under Australian law, and is not welded to a single foreign model that could be repriced or switched off. That is your slice of AI sovereignty — realised in one business, without a billion-dollar model.