What is Sovereign AI?

I've recently become really interested in the concept of Sovereign AI after hearing Sarah Kaur from Portable discuss it on the In The Blink of AI podcast. So I decided to dig deeper.

What exactly is Sovereign AI?

Basically we're talking about AI models and infrastructure that is developed by a nation (this can include public and private institutions), deployed and controlled within that nations borders, and controlled by its own institutions.

In practice, this means Australian government departments, critical services, local businesses, and the general public can access powerful AI models without relying on foreign-owned platforms or data centres.

Let’s break that down

An AI model might be a Large Language Model (LLM) like ChatGPT (OpenAI) or Claude (Anthropic), both commercial and US-based, or open-source options like Meta’s Llama. Whether global or sovereign, all models rely on three critical ingredients:

1. Data: training a LLM requires massive amounts of data. A sovereign model needs access to high quality, locally curated datasets, collected responsibly with proper consent in order to train a model with local context.
2. Compute: to train the model on this data we need access to local super computers jammed packed with thousands of GPUs (computer chips perfect for AI training) that work through the data, build patterns and connections to create these AI models. This means building data centres to house these systems which require physical space, energy resources and capital.
3. Talent: finally we need to keep AI talent in the country and grow that talent pool by providing access to these models and infrastructure. Creating a culture of innovation and research so we avoid the brain-drain of AI experts heading off to Silicon Valley.

We also need to ensure these sovereign models can be deployed on local infrastructure and made accessible to public and private organisations. This allows for proper oversight and governance of how the models operate, ensuring transparency and accountability. It also means equitable access without relying on foreign entities or governments, and greater cultural sensitivity. Because its important to remember every AI model reflects the biases, priorities, and perspectives of those who create it.

In upcoming posts, I’ll explore the benefits, challenges, and global examples of Sovereign AI — including what’s happening here at home.

Want to dive deeper? Have a look at Maincode’s whitepaper "Deploying Sovereign AI" and their locally developed model, Matilda.

So would you use an Australian-trained AI model over ChatGPT or Gemini? Why or Why not?

This article was originally published on LinkedIn.