Australian-made AI is independent AI

In October last year I wrote that we’d moved beyond sovereign AI. The intent behind the word was right (control, security, capability), but the framing had become too loaded and too defensive. We started talking about Australian-made AI instead, meaning AI made here, trained here and aligned with our needs and values.

The more we build Matilda, the more we find ourselves working with organisations that value independence, and a growing share of them are outside Australia. We’re having the same conversation in New Zealand and Singapore that we have in Melbourne and Canberra. Those organisations don’t need their AI to be Australian-made. They need what Australian-made AI gives an Australian organisation, which is AI that nobody else can switch off or change. We call that independent AI, and it’s what Australian-made AI has been about all along.

What makes AI independent

Independent AI isn’t AI that acts on its own, and it doesn’t mean building every chip and every model at home. It means AI an organisation, or a country, can keep running, change and rely on without anyone else’s permission. In July I told SmartCompany we’d ditched “sovereign” because “you can’t test it”. You can test independence, with five questions about any AI your organisation depends on.

1. Who holds the model’s weights?
2. Where does it run, and whose laws reach it there?
3. Who decides when it changes?
4. Who can switch it off, or cut off your access?
5. Whose values and requirements was it trained to?

If every answer is you, your AI is independent. If any answer is a company or a government somewhere else, it isn’t, however close the data centre is.

This year showed who decides

On 12 June the US Government, citing national security, ordered Anthropic to suspend all access to its newest models, Fable 5 and Mythos 5, “by any foreign national, whether inside or outside the United States”. Anthropic had no reliable way to check nationality in real time, so it switched both models off for every user. Fable 5 came back 19 days later. Mythos 5 came back first for a set of US organisations, and in September its successor was still available only to a set of US organisations. Two weeks after the order, OpenAI released its newest models to “a small group of trusted partners” first, at the US Government’s request.

Twelve days after the order, APRA’s Therese McCarthy Hockey warned Australian financial institutions that “a critical business process, control or cyber-defence capability that depends on a single offshore frontier AI model may be disrupted not only by an outage or cyber incident but by a regulatory decision made overseas”. Thailand’s SCBX learned a quieter version of the same lesson in 2024. When its provider rolled out a new version of the model it relied on, “the model’s performance in Thai unexpectedly degraded, and the team had no way to intervene”.

Australian-made AI is Australia’s answer

In July the Prime Minister told the University of Sydney that “if we are always dependent on someone else, somewhere else, we will always be vulnerable”, and that Australia “can be much more than a data warehouse for AI products made overseas”. The National AI Plan says locally developed models “may provide benefits such as reflecting unique cultural context or language, supporting innovation and mitigating risks around data security”. In October Parliament’s AI committee told the Prime Minister that Australia should not be “merely a consumer of AI technologies developed elsewhere”.

Matilda is built for exactly that. It runs on our own GPUs in Melbourne, Maincode is Australian-owned, and when the US order landed in June it didn’t touch us, because we don’t rely on models hosted in the US in the first place.

New Zealand and Singapore want the same thing

New Zealand’s government guidance on cloud services warns that the risks of foreign law “are still present in onshore data centres where IT products and services are owned by offshore vendors”, because “it’s not possible to fully contract out of the legal framework of another country”. Its AI strategy says the country’s advantage lies in becoming “sophisticated adopters who can quickly identify, adapt, and deploy AI solutions to local challenges”, not in competing with Google or OpenAI. Adapting the best open-weights models and running them under New Zealand control is how that strategy and that warning fit together.

Singapore’s national models already work that way. SEA-LION’s current models are built on open models from Google, Alibaba, Swiss AI and NVIDIA, adapted to give “better data representation and alignment compared to Western or Chinese models”. In 2024 Singapore’s government told Parliament that where sensitive data is involved, “open-source models may be finetuned for use but must be deployed on Government servers and computers”. In June its digital minister, Josephine Teo, said that “every country must be able to use AI on its own terms, in service of its own people”, and that owning every layer of the stack is “neither realistic nor helpful for most countries”.

The rest of the world wants it too

In February Emmanuel Macron told the AI summit in New Delhi that “India chose granular and smart, and Europe chose sovereignty and scale. But both chose independence, and both were right.” Canada’s AI strategy says its AI infrastructure must be “operated under Canadian control and Canadian law”, not dependent on platforms that can be “restricted or withdrawn at the discretion of a foreign government”. In Deloitte’s survey of 3,235 business leaders in 24 countries, 77% said they now factor an AI solution’s country of origin into their vendor selection.

Some countries insist on training their own base models. Korea requires the models in its national AI project to be trained from scratch, to address “technological, cultural, economic, and security dependence stemming from heavy reliance on global AI models”. Switzerland’s Apertus was trained from scratch too. Many more do what Singapore does and adapt the best open-weights models to their own languages, laws and needs. Japan’s Sakana AI says the focus of Japan’s model development “should be in post-training”, and Thailand’s Typhoon team built a Thai model this way with “only an academic budget”.

Independent of us, too

An agency in Wellington or a bank in Singapore shouldn’t swap a provider in San Francisco for one in Melbourne and call it independence. Independent AI has to be independent of us as well, and that’s why we build on open weights.

We take the best open-weights models and put them through the Matilda training system, which aligns them to your ideological, cultural and enterprise requirements and trains them on your own material. You run the result in your own instance, in our Melbourne facility or in your own data centre. You hold the weights under the base model’s licence, and if you leave, they come with you.

Some of the best open-weights models come from Chinese labs and some from American ones, and both governments now want a say over who gets the best models. Washington showed that in June, and on 16 October Beijing proposed export controls that cover the weights of its largest and most capable models, while leaving out models already published as open source. Weights you already hold stay on your hardware whatever either government decides next. With open-weights models coming from Google, Mistral, NVIDIA, OpenAI, Alibaba, Moonshot AI and Z.ai, you aren’t tied to one lab or one country.

Australian-made AI is independent AI made in Australia. New Zealand-made and Singapore-made AI are the same idea made in New Zealand and Singapore, and they can all start from the same open-weights models and still end up as separate models, each answering only to its owner. If you want AI that answers to you, wherever you are, talk to us.