The Australian Privacy Act, the Critical Infrastructure Act and APRA CPS 230 all share an assumption: that regulated entities know where their data is, who can compel access to it, and under what foreign law.
When you call an inference API hosted in a US region, none of those questions has a clean answer. The provider may be a US-incorporated company subject to the CLOUD Act. Your prompts and responses sit in that provider’s logging pipeline. Even if the contract says they don’t train on your data, the disclosure surface is the contract, not the architecture.
Sovereign inference is the architecture answer. Weights run on Australian GPUs. Prompts hit an Australian endpoint. Responses come back. Logs stay onshore. Nothing crosses a jurisdiction boundary.
What changed in 2025–2026
Two things, really. First, the open-weight model gap closed. Llama 3.x, Qwen 2.5, DeepSeek V3 and Mistral Large are all competitive with frontier closed models on most benchmarks that matter for production workloads. Second, the inference economics improved enough that hosting your own models on Australian GPUs no longer carries a 3–5× cost premium over calling a US API.
Where the trade-offs still live
For some workloads, the absolute frontier — Claude Opus, GPT-5 — still matters. Amaze ships those as non-sovereign catalogue entries so customers can opt in per workload, with the sovereignty tag visible at the model card level. The point is informed choice, not blanket refusal.
