At a glance
The Department of Finance has added Google's Gemma, NVIDIA's Nemotron and Writer's Palmyra to GovAI, the AI service built for the whole Australian Public Service.
Finance says every model used by the GovAI chat service must be hosted onshore, with data processed only within Australia.
Amazon offers NVIDIA's Nemotron 3 Super from both Sydney and Melbourne, and publishes no setting that would send a request for it overseas.
The two Writer Palmyra models Amazon carries are offered from four US locations and no Australian one.
GovAI's pages name its cloud suppliers but not the data centres, and its environment for sensitive material is still being security-assessed through 2026.
Finance adds Gemma, Nemotron and Palmyra to GovAI
Three new families of AI model are now reachable through GovAI, the Department of Finance's artificial intelligence service for the Australian Public Service. iTnews reported on 30 July that the service has taken on Google's Gemma, NVIDIA's Nemotron and Writer's Palmyra, joining the OpenAI, Anthropic and Mistral models already there.
Finance told iTnews that all models used by the GovAI chat service must be hosted onshore, with data processing also only taking place within Australia. A model can only meet that condition if its maker already runs it from an Australian data centre. That gives Amazon and Microsoft, GovAI's two cloud suppliers, considerable influence over what the Australian government can offer its own staff, because a model neither has switched on locally is a model Finance cannot put in the chat tool.
GovAI launched in 2025 and describes itself as a service running on "secure, Australian-based infrastructure". The wider service is in open trial for public servants with a GovTEAMS licence, the government's cross-agency collaboration platform. GovAI Chat, the version intended for every APS employee, is listed on GovAI's own site as coming in mid-2026, and iTnews describes it as in beta trial.
The model brokerage admits offshore models after a risk assessment
The model brokerage is the second of GovAI's two routes in, and it is what agencies use when they build AI into their own systems. There Finance allows exceptions: an agency can ask for extra models, including ones hosted overseas, once a risk assessment is done and controls are in place. The chat tool, which is what most public servants will ever see, is the route bound by the onshore condition.
Finance's wording to iTnews was that onshore Australian-hosted models remain the default, and overseas hosting is treated as an exception subject to governance, classification limits and security requirements. A model with no Australian home is still available to agencies through the brokerage, once a risk assessment clears it.
The GovAI model access page names two suppliers, Microsoft and Amazon, and says the cloud platforms underneath are assessed under IRAP, the government's security assessment scheme, as suitable for data classified PROTECTED. That is a level above Official: Sensitive, and the directory of Australian data centre sites tracks which facilities hold government certification. GovAI's own environments are assessed separately, and its hosting page has the Official: Sensitive pilot still going through IRAP during 2026. The model page is candid about the trade-off: because Australian-hosted models come through specific agreements, there may be a delay before the newest versions turn up.
Nemotron runs in Sydney and Melbourne with zero cross-border routing
Amazon publishes, model by model, which of its data centre locations can serve a request and whether the request may be passed elsewhere when a location gets busy. The strictest setting keeps the request where it was sent. Looser ones let it be answered anywhere within a named group of Amazon locations, which may span several countries, or anywhere in the world.
Nemotron 3 Super is on the strictest setting in all eleven locations Amazon offers it, and Sydney and Melbourne are two of them. Amazon publishes no option at all for passing a Nemotron 3 Super request to another location, so a request sent to Sydney is answered in Sydney.
Palmyra and Gemma 4 have no Australian listing on Amazon
The two Palmyra models Amazon carries, X4 and X5, are listed in four American locations and nowhere else, and Amazon's only published route to them runs through a United States setting. Palmyra does not appear in Microsoft's published catalogue of partner models either, which covers both of GovAI's suppliers.
Google's Gemma divides by version. The Gemma 3 model Amazon carries in Sydney and Melbourne is the pretrained base checkpoint rather than the instruction-tuned version normally put behind a chat tool. Gemma 4, the newer generation, is not on Amazon in Australia at all, running from three American locations and Frankfurt by Amazon's June announcement.
Model on Amazon | Australian availability | Cross-border routing |
|---|---|---|
NVIDIA Nemotron 3 Super 120B | Sydney and Melbourne | None offered |
Google Gemma 3 27B PT | Sydney and Melbourne | None offered |
Google Gemma 4 | None | Three US locations and Frankfurt |
Writer Palmyra X4, X5 | None | Four US locations only |
Source: Amazon Bedrock model cards, 2 August 2026.
Finance has not said which of its two suppliers is hosting the three new families, and GovAI's model list has not been updated to show them. If Amazon is the route, then on the condition Finance described, Palmyra and Gemma 4 sit behind the risk-assessment door rather than in the chat tool.
Amazon and Microsoft route Asia Pacific requests across five countries
On Anthropic's Claude Sonnet 4, Amazon's Asia Pacific setting can answer a request sent from Sydney out of Tokyo, Seoul, Osaka, Mumbai, Hyderabad or Singapore, as well as Sydney and Melbourne. Amazon sets that destination list model by model, and states that prompts and results might move outside the location they were sent to even though stored data stays put.
Microsoft's equivalent works the same way. Its Asia Pacific zone is defined in Microsoft's documentation as Australia, Japan, Korea, Singapore and India, a list Microsoft says it may add to without prior notice, and only its single-location settings keep processing in one country.
Amazon has since added a narrower Australian setting, used by Anthropic's Claude Sonnet 4.5, which from a Sydney or Melbourne starting point sends requests only to Sydney and Melbourne.
Onshore models need accelerators racked in Australian facilities
Gemma and Nemotron are open-weight models, meaning the files behind them are published and can be downloaded and run on your own hardware, subject to each maker's licence. On Amazon the government does not run its own copy: it pays by usage and Amazon owns the machines. For a model Amazon will not pass to another location, serving Sydney and Melbourne means Amazon has put the hardware to run it in both cities.
Running your own copy moves that cost onto the customer. Microsoft's version, still in preview, reserves dedicated AI chips for a single customer, who picks a chip type such as an NVIDIA H100 and holds a separate allowance for it in each location. Google's route to running Gemma yourself also means provisioning your own accelerators rather than drawing on shared capacity.
Either path, buying by usage or standing up your own deployment, only satisfies an onshore condition if the chips are racked in an Australian building. That is the demand hyperscalers in Australia and the GPU-focused operators beneath them are competing for, and it explains why a policy sentence about data processing ends up as a question about AI data centre capacity. It arrives alongside the Whole-of-Government Cloud Computing Policy, under which cloud has been the default consideration for new Australian government digital investment since 1 July 2026. That policy binds non-corporate Commonwealth entities, is optional for corporate ones, and excludes the national intelligence community.
Agencies are already buying the largest and newest AI systems outside GovAI altogether. The Fair Work Commission signed Anthropic's first direct Australian government contract in March, and the Australian Signals Directorate followed in April with an agreement awarded without competitive tender. Anthropic is inside GovAI as well, with Claude 4.5 Sonnet listed on the brokerage through Amazon.
What to watch
The GovAI model list. The published page names Microsoft and Amazon as the two suppliers and does not yet show Gemma, Nemotron or Palmyra among the available models.
The Gemma version. The Gemma 3 checkpoint Amazon serves from Sydney and Melbourne is a base model rather than a chat-tuned one, and Gemma 4 is not on Amazon in Australia, so which version Finance has taken and on whose platform decides whether it can reach the chat tool.
The IRAP assessment. GovAI's hosting page has the Official: Sensitive pilot environment still going through security assessment during 2026.
The Australian-only setting. Amazon now offers Sydney-and-Melbourne-only routing on Anthropic's Claude Sonnet 4.5, and how far that spreads across other suppliers' models is one thing to watch for what the chat tool can offer next.