Australian law almost never requires data onshore, and a Sydney region of a US platform gives you residency while jurisdiction stays with the provider's home courts, which I work through in What Australian compliance requires before you bring in AI. This article is for the organisations left over, the ones whose requirement is sovereignty itself. Disclaimer: Veriet runs on Claude, so weigh my caution about leaving the frontier platforms against that, and none of this is legal or financial advice.
Sovereignty ends at open weights
Every frontier model is operated by a US company, Anthropic, OpenAI, Google and Microsoft included, and under the Clarifying Lawful Overseas Use of Data Act (CLOUD Act) a US provider can be compelled to produce data in its control wherever the server sits, so no region setting, contract clause or local partnership changes whose courts have reach. The one sovereign arrangement that does exist here, the government's A$2 billion top-secret cloud with AWS (opens in a new tab), serves defence and intelligence alone. Anthropic's July 2026 memorandum of understanding with the Australian government and its talk of local capacity improve residency, which already satisfies almost every Australian rule, so what stays open is the jurisdictional gap.
Open weights close that gap mechanically, because a model whose weights sit on your hardware sends nothing back to its maker, so there is no provider for a foreign court to compel, and the cost of that shows up in capability, operating effort and hardware.
What the open models offer in August 2026
Most executives assume the capability gap is wider than it is, because my reading of the public benchmark trackers is that the strongest open-weight models now sit months behind the frontier, somewhere between 6 and 12 depending on the task, and the gap narrowed through 2026. Even OpenAI ships open weights now, and its gpt-oss release (opens in a new tab) of August 2025 is Apache licensed with the larger model running on a single 80GB GPU.
The leaders are Chinese, with Moonshot's Kimi, DeepSeek, Alibaba's Qwen and Zhipu's GLM at the top of the open trackers while Meta has stepped back from the open frontier. An Australian sovereignty program running on Chinese-trained weights sounds like a contradiction, but the jurisdictional logic holds, because the weights run on your hardware and nothing flows back to the trainer. What the origin changes is the assurance work, since the question moves from who can reach the data to how the model behaves, and you answer that one with evaluation instead of contracts.
A fully Australian-trained alternative stays out of reach for now. The most visible local effort, Melbourne's Maincode and its Matilda model (opens in a new tab), runs inference onshore and dropped the sovereign label as the scale reality set in, and its reported infrastructure spend in the tens of millions sits orders of magnitude below what frontier training runs cost, so an Australian frontier model is years away at best, which puts it outside any procurement decision you make now.
The managed layer barely exists here
Adopting open weights cheaply needs a managed inference service with a public model catalogue, per-token pricing and a service agreement, the shape AWS Bedrock gives the closed models, and you cannot buy that in Australia today. ResetData (opens in a new tab) advertises sovereign inference APIs on local GPU clusters, and everything on its site routes to a sales conversation, with no published catalogue, pricing or service levels I could find. Sharon AI and Firmus are building GPU capacity at scale, which is the raw ingredient you would still have to turn into a service yourself, while the global managed-inference providers list Australia as configurable on request at best.
A sovereign deployment today therefore means running it yourself, with a platform team, security hardening, an evaluation harness, and a model refresh every few months because the open leaderboard changes that often. For an organisation without a strong engineering bench, that operating cost dwarfs the hardware and it comes back every year.
Hardware prices are at the top of a shortage
Buying has got worse this year, because AI demand has pulled memory manufacturing toward high-bandwidth parts, and TrendForce-derived reporting puts DRAM contract prices up roughly 50% in the first quarter of 2026 alone (opens in a new tab), with enterprise solid-state drive contract prices up about 80% in the same quarter and the shortage forecast to run into 2027 and 2028. At the system level, Tom's Hardware reports (opens in a new tab) Nvidia's next-generation racks at up to US$8.8 million each. Rental tells a softer story, because competition among the new GPU clouds has held rates for last-generation chips roughly flat, which is why renting in-country GPUs beats buying racks for almost anyone starting now.
Buying hardware in a shortage means paying the top of the cycle for equipment that depreciates on a model-release calendar, so I would rent until the memory market normalises, and the forecasts say that is 2027 at the earliest.
Who should build anyway
Some organisations hold a constraint that ends the debate: work adjacent to defence and intelligence classifications, operators inside the My Health Record system's localisation rule, or contracts and mandates that forbid offshore processing outright. For them my advice is to build narrow, because you are paying a premium for jurisdiction and the premium scales with how much you put behind it. Rent Australian GPUs rather than buying, run one strong open model rather than a menu, point it at the few workflows that carry the constraint, and let everything else in the organisation use the frontier platforms under the contractual controls I set out in the compliance piece.
What would change my mind
I would flip this position on 3 developments, and I expect at least the first within a year. A managed inference service in Australia with public pricing and service levels, from ResetData, a hyperscaler or someone new, would remove the operating burden that does most of the damage. Memory prices normalising would fix the capital side. And the open models reaching parity with the frontier, which their trajectory through 2026 makes plausible, would remove the capability tax entirely. This verdict is dated 25 August 2026 and it is the kind that ages quickly.
Until one of those arrives, the order of operations stands: settle the compliance question with contracts and residency, which the compliance piece shows the law supports, and hold sovereignty for the narrow cases that can name the constraint forcing it.