IMREAL.LIFE
 ·   ·  6 posts
  •      1 members
  • 10 friends

Own your AI, don't rent it: what sovereign AI means for Australian small businesses

"You don't own your AI, you rent it." It's a line doing the rounds in Australian tech circles, and it's accurate for most businesses. Your prompts, your client data and your workflows run on someone else's servers, on someone else's terms, often in another country.

Renting is fine, until the price changes, the terms change, the model you built around gets retired, there's an outage, or a client asks where their data went. Owning at least part of your AI stack is how you take those risks off the table.

Australia is asking the same question

In August, Parliament set up a Joint Select Committee on Artificial Intelligence. Its remit covers issues including the workforce, data sovereignty and national security. Submissions closed on 14 September, and the committee has to report by 30 November.

James Newton-Thomas and his team read the 540 published submissions and tallied what people asked for. By their count, the top request, made in 121 submissions, was to build sovereign AI capability. Funding worker retraining came second with 118. That's their own analysis, not an official committee figure, but the message is clear: build it here, control it here, and help people learn to use it.

A local example, with caveats

At the end of September, Australian AI company Maincode open-sourced Matilda Jev, which it calls Australia's first "decision model". Instead of generating text, it answers a typed question (pick an option, yes or no, or a score) with a probability for each option. Maincode says it was post-trained and is served on Australian infrastructure, and the weights are on Hugging Face under the Apache 2.0 licence.

The balanced view matters here:

  • The model card lists a Qwen 27B model, from Alibaba, as its base. So this is Australian post-training on an overseas open-weight model, not a model built from scratch here. Commenters on the launch post raised exactly this.
  • The headline benchmark scores were run by Maincode itself, and they've been submitted for independent verification. Treat them as claims until that happens.
  • It isn't a small download. The weights are about 49 GiB, and Maincode tested on data-centre AMD hardware.

None of that makes it a bad release. Maincode's own position is that sovereignty is about where a system runs, who controls it and how it's governed, not about inventing every layer locally. I think that's the right way to look at it, and it applies to small businesses too.

Why owning your AI matters for a small business

  • Privacy. Client files, financials and internal notes never leave your building. If you handle personal information, the Privacy Act's rules on sending it overseas (APP 8) are one more reason to know exactly where your prompts go.
  • Data sovereignty. You decide where data is stored, how long it's kept and who can see it.
  • Cost control. A local server is a known, one-off cost plus power, rather than a per-token bill that grows with use.
  • Resilience. Your models keep working when a provider has an outage, changes its pricing or retires the model you depend on.
  • Customisation. You can tune prompts, workflows and even model weights to your own business without asking permission.

What owning looks like in practice

You don't need a data centre. For most small businesses, owning your AI looks like this:

  • A modest local server. One Linux machine with a decent GPU, or a desktop with plenty of unified memory, can run capable open-weight models for drafting, summarising, classifying and searching your own documents.
  • Open-weight models. Run them with tools like Ollama or llama.cpp, with Open WebUI as a friendly front end for staff.
  • Sensitive data stays in-house. Point a local search and retrieval setup at your own files, so the answers come from your documents and nothing leaves the network.
  • Cloud where it makes sense. Use the big hosted models for heavy, non-sensitive work. Route those calls through a gateway you control, so you can switch providers and strip personal details before anything goes out.
  • Security from day one. A self-hosted model on an open port is worse than no model at all. See our checklist on hardening self-hosted AI before you put anything on your network.

Skills are the real bottleneck

Retraining came a close second in those submissions, and I'm not surprised. Hardware is the easy part. The harder part is having someone in the business who can set up a model, keep it patched, connect it to real workflows and know when the cloud is the better tool. Those skills are learnable, and you don't need to be a developer to pick them up.

Start owning yours

If you want to run your own AI rather than rent it, join the IRL community at imreal.life. The 30-Day Sovereign Builder Challenge walks you through building a private AI stack step by step, alongside other Australian builders.

Take the 30-Day Sovereign Builder Challenge. See the full day-by-day blueprint and get started here: https://imreal.life/page/30-day-challenge

Tell me in the comments what you'd want to run in-house first.

Sources

  • Favorite
  • More
Comments (0)
Login or Join to comment.

IMREAL.LIFE

Close