Australia needs more than AI-written emails

TL;DR

Australia needs substantial investment in AI it can build and operate locally. Everyday tool use is encouraging, but business integration remains uneven and sensitive workloads need suitable infrastructure. The confidential GPU gap is access to an enabled, verifiable Australian service, not a complete absence of capable chips.

Black-and-white newspaper cartoon: Rob, in his approved Roman helmet, glasses, black T-shirt and tattooed forearms, stands at an Australian AI helpdesk holding a request for 8 H100s. A clerk offers a long printed email beside an enormous pile labelled EMAIL DRAFTS. Caption: We have drafted your email to Japan.

Australia needs to spend much more on the ability to build and run AI here. Getting ChatGPT to tidy up our emails is useful. It is also a very low bar for a national ambition.

I came back from GTC Berlin thinking about that gap. The conversations I had were about open models, local knowledge and how countries could operate AI for themselves. There was plenty of unfinished work. But people were arguing about how to build the capability.

We should be having much more of that argument here.

Are we confusing use with adoption?

Everyday AI use is encouraging, but it tells us little about how deeply an organisation has changed its work.

OpenAI’s September data puts writing at 41.6% of Australian work-related messages on individual ChatGPT plans. That includes more than email, and excludes enterprise accounts and Codex. It is a picture of what people do in a chat window, not a national productivity score. OpenAI Signals

The business gap is more revealing. In CPA Australia’s latest survey, Australian respondents reported 88% AI use, but only 10% had deeply embedded it across their organisation. Australia had the lowest broad integration among the five markets surveyed. These were accounting and finance professionals reporting on their organisations, not a census of Australian business. CPA Australia, October 2026

That is the distinction I care about. A staff member can get help writing a reply without the business changing very much. Putting AI into a claims process or an engineering workflow means connecting it to real systems, deciding what it may read, and taking responsibility when it gets something wrong.

(The email can still begin with “I hope this finds you well”. Apparently we cannot automate taste.)

Is government any further along?

Government has made real progress, but a staff assistant is only one part of the job.

The OECD’s 2026 assessment says Australia “performs strongly across most dimensions of AI in government”. That assessment draws on 2023 and 2024 evidence, and its digital-government index measures enabling foundations rather than the digitalisation of particular services. OECD country assessment

GovAI Chat is a practical example. Its current beta offers public servants models managed on Australian soil for drafting, research and analysis. It permits public or OFFICIAL information, but excludes OFFICIAL: Sensitive and PROTECTED material. Those are sensible limits for that service. They also leave a substantial question about how we support more sensitive work. GovAI Chat

My concern is that we become comfortable counting accounts and trials while the harder operational work moves too slowly. Government should help create a market for secure local capability by buying useful systems and measuring what they actually improve.

Can we obtain the compute here?

An Australian data centre does not automatically mean an Australian team can obtain the GPUs it needs.

For the NVIDIA poster paper, I needed eight H100 GPUs. The only capacity I could find was in Japan. There was none available to me in Australia for that job.

These were ordinary GPU instances. The work did not need confidential computing. I was simply trying to get eight GPUs to do the work.

Renting overseas compute is often a perfectly reasonable decision. I’ve made the case for renting while a technology is young. The problem comes when it is the only practical option, particularly for work with location or control requirements.

Australian engineers can learn plenty on overseas infrastructure. But my read is that a domestic operating industry also needs local providers and customers doing real work together. If suitable capacity is difficult to obtain here, that industry will be slower to develop.

Where are the confidential GPUs?

What I have not been able to obtain is an Australian GPU service with confidential execution enabled and a way to verify it.

Confidential computing protects data while it is being processed, within a defined hardware and software boundary. For GPU workloads, that protection must extend beyond the CPU to the accelerator. Attestation provides cryptographic evidence about the environment before we trust it with data. A padlock beside the storage account does not answer that question. NVIDIA’s confidential-computing overview

Azure’s regional availability table lists its confidential H100 series in Central US, East US 2 and West Europe. As checked on 13 November, it lists no Australian region for that series. Google’s published confidential H100 zones are also in the US and Europe. Published support still does not guarantee that your account can allocate capacity. Azure regional availability, Google’s supported configurations

AWS documents Nitro protections for GPU workloads. I could not find a comparable EC2 offering exposing NVIDIA confidential GPU mode and customer-verifiable GPU attestation. AWS’s stated security model

The hardware is arriving. ResetData describes 1,024 H200s in Melbourne. NVIDIA names Firmus and IREN among providers supporting confidential-computing deployments. Neither fact, by itself, identifies an Australian service I can order with the required attestation. ResetData, NVIDIA

Across the provider catalogues checked for this piece, I could not verify an Australian service available to order with confidential GPU execution enabled and a documented attestation path.

Honestly, this should be much easier to buy here. We are asking organisations to use AI with sensitive information, while the infrastructure needed for some of those workloads remains difficult to obtain locally.

What should we invest in?

We should invest heavily in usable local capacity and the people who can turn it into working systems.

The honest caveat is that money is already arriving. RBA staff estimated this week that Australian data-centre investment had increased around tenfold in three years. We should welcome that, then ask who can use the resulting capacity and under what terms. RBA staff analysis, 9 November

For me, sovereign AI means practical control over where sensitive workloads run, who operates them, and who holds the keys. It also means local teams can obtain enough compute to build something. A machine being physically here answers only one of those questions.

Europe offers one useful example. EuroHPC’s Playground scheme offers eligible smaller businesses a route to compute access within two working days, with onboarding help. It has conditions and limited allocations. It still gives a team a practical way to begin. EuroHPC access scheme

I would put Australian money into shared GPU capacity that smaller teams can afford, including confidential services with published verification procedures. Fund the engineers who can operate it. Use government purchasing to support deployments with measurable outcomes. Require credible plans for electricity, cooling and ongoing hardware replacement before celebrating another announcement.

More GPUs will not repair bad data or a broken process. Confidential execution does not decide what an AI is allowed to read, or make its answers correct. Those jobs still need doing.

If we leave them undone while other countries build experience, our businesses will have to compete against organisations that may be able to deliver useful AI more cheaply and reliably. We will still be paying for access, with less say over the terms.

I would like Australia to be ambitious about more than the inbox. Give local teams the compute and the responsibility to build useful systems here. Then judge the investment by what they deliver.

Sources

  1. Signals individual data, version 3 · OpenAI · 2026-10-26
  2. Aussie businesses risk falling behind in AI race · CPA Australia · 2026-10-09
  3. Digital Government Outlook 2026: Australia · OECD · 2026-06-15
  4. GovAI Chat · Australian Government
  5. Azure Confidential GPU options · Microsoft
  6. Azure products by region · Microsoft
  7. Confidential VM supported configurations · Google Cloud · 2026-11-10
  8. AWS Confidential Computing · AWS
  9. Confidential Computing Overview · NVIDIA
  10. AI-F1 launch · ResetData
  11. Confidential computing for model builders and infrastructure providers · NVIDIA · 2026-10-21
  12. Data Centre Investment in Australia · Reserve Bank of Australia staff · 2026-11-09
  13. Playground Access to AI factories · EuroHPC Joint Undertaking

Plain-markdown version of this article