Why “hosted in Australia” isn't quite the answer.

Ask a software vendor where your data lives and you'll usually get a confident answer. Sydney. Melbourne. The Australian region. That answer is worth having, and it's also answering a narrower question than the one most people think they asked.

Two questions that sound like one

Residency is geography. Which country are the bytes stored in, which country are they processed in, and can the vendor show you that in writing. It's checkable, it's contractual, and plenty of providers now offer it as a normal product feature.

Sovereignty is law. Which legal system can compel the organisation holding your data to hand it over. That reach follows the company, not the hardware. A US-headquartered provider running a data centre in Sydney is still a US company, and remains subject to US legal process wherever the disks are.

Neither fact is a scandal, and neither one means cloud AI is unsafe. The problem is narrower: vendor marketing tends to answer the geography question when a buyer has asked the legal one, and buyers rarely notice, because the two get described with the same words. If your professional obligation, your insurer or a client contract says something about foreign access, a map of data centres does not address it.

What the CLOUD Act Agreement did, and what it left alone

The Australia-US CLOUD Act Agreement took effect on 31 January 2026. It formalises how the two governments make requests to each other for electronic evidence, replacing a slower and clunkier process. That's a real improvement in a genuinely useful area, and it's a government-to-government arrangement.

Here's what it doesn't do. It doesn't put a private Australian business beyond the reach of US legal process where the provider is a US-headquartered company, and the physical location of the storage isn't the deciding factor in that question.

A documented example makes this concrete without any need for speculation. In May 2025, in the New York Times litigation in the United States, Magistrate Judge Ona T. Wang ordered OpenAI to preserve and segregate output log data that would otherwise have been deleted, on a going-forward basis. That order sat above user deletion requests. Its scope covered ChatGPT Free, Plus, Pro and Team, plus API use without a zero data retention agreement. ChatGPT Enterprise was excluded, and API customers who had a zero data retention agreement were not affected. The order was lifted around 26 September 2025.

The lesson isn't that a vendor broke its promises. Two more useful things come out of it. A vendor's data commitments can be overridden by a court in a country your business has no relationship with. And the contract tier you happened to buy decided whether that applied to you, which is not something most small businesses have ever checked.

For a great many Australian businesses this is a tolerable risk, honestly assessed. It stops being tolerable when a professional obligation, a client's contract clause or a regulator's expectation says the data has to sit outside that reach.

The practical middle path: residency that really is Australian

If your requirement is genuinely about storage location, the middle path is good and it's available now. AWS Bedrock in the Sydney region (ap-southeast-2), using the Australian Geo inference profiles, gives real Australian data residency for Claude. Sonnet 4.5 and Haiku 4.5 gained Australian Geo profiles in April 2026, and Opus 4.6 is the first Opus-class model that can be pinned to Australia at all.

Worth knowing: Anthropic's direct API has no Australia region. If you want a frontier-class Claude model with Australian residency today, Bedrock in Sydney is the route, and that's the whole list. Residency also carries a small premium. Anthropic's data-residency pricing through Bedrock and Google Cloud runs about 10% above standard for Sonnet 4.5, Haiku 4.5, Opus 4.5 and future models in the regions where each is offered, so budget for it rather than being surprised by it.

Now the scrupulous part. This arrangement solves a residency requirement properly. AWS is a US company, so it does not answer a jurisdiction requirement, and anyone telling you otherwise is selling. For most businesses that distinction never becomes load-bearing and the Sydney region is the right call: strong models, Australian storage, nothing to rack, nothing to maintain.

Australian-owned, on Australian soil

One step further along is Australian-owned infrastructure. Sharon AI is an example: an Australian-owned operator running GPU capacity in the Equinix SYD3 and SYD5 facilities and in NEXTDC's M3 data centre in Melbourne. Australian company, Australian buildings, Australian law.

Two honest notes. There's no public pricing, so it's a quote-only conversation, and nobody can tell you what it costs from published information. And a third party still sits in the data path, which means the contract, the access controls, the deletion terms and the audit trail all still matter. What this tier removes is the foreign-parent question. It doesn't remove the need to read what you're signing.

On-premise, where the question stops being about anyone else's legal system

Run the model on hardware you own, in a building you control, with no outbound path for the data, and the jurisdiction question dissolves. No stronger contract is involved; there's simply no other party left to compel. That's the only architecture where the answer stops depending on someone else's legal system.

The regulator has said as much. The OAIC's guidance on commercially available AI products notes that deploying an AI system locally is “likely to be more privacy-preserving as it limits the risks of third party access to the data”, in contrast with cloud arrangements where servers may sit outside Australia and personal information could be disclosed internationally. That's the privacy regulator describing local deployment as the privacy-preferable architecture, which is a stronger statement than any vendor could make about itself.

It isn't free, and it isn't right for everybody. On-premise means capital instead of a subscription, real maintenance hours, and a capability envelope that's narrower than the frontier. Our honest write-up of what local AI actually costs includes the seat counts where it stops making financial sense, and what open models are genuinely good enough for covers where the capability gap does and doesn't bite. If you want the whole picture in one place, the local and private AI page lays out how JDCS approaches it.

One caveat, stated plainly. This is general information, not legal advice. Which obligation applies to your business, and whether it's phrased as residency or jurisdiction, is a question for a lawyer who has read your contracts. What a technical adviser can do is tell you what your systems actually do with the data, which is usually the part nobody has written down. Related reading: our guides on whether your business data is safe with AI and the December 2026 Privacy Act changes.