Architecture  /  AI Infrastructure
AI Infrastructure · Deep dive 01

When the model can't leave the boundary

Australia banned DeepSeek on government devices in 2025, open weights included in some directions. The procurement and data-residency case for running a capable model inside PROTECTED.

Draft outline · Sovereignty / policy lens
The anchor

In 2025 Australia (alongside South Korea, Taiwan, Italy and others) banned DeepSeek on government devices, and some directions extended to downloadable open-weight models, citing foreign-access and data-residency concerns. For a PROTECTED workload the question is not which SaaS model to trust, it is how to run a capable model where the data never leaves the boundary.

Sources we build on
Journalism
TechCrunch: calls to ban PRC-produced AI models

Independent reporting on the policy momentum and the arguments driving government bans.

Primary
Australian Government DeepSeek direction / cyber.gov.au

Primary source for the Australian government position on DeepSeek and foreign-access risk.

Article outline
  1. Why the bans happened. Foreign access, data residency, and why open weights got swept in.
  2. The PROTECTED constraint. The data cannot touch a public API, full stop.
  3. Running a model inside the boundary. Azure options, GPU sizing, classification-aware architecture.
  4. Capability versus control. The honest trade-off against frontier SaaS models.
  5. Supply chain of the model itself. Provenance of weights and dependencies.
How it aligns to what we do

A sovereignty and policy piece, not a security-scare piece. It leads with the procurement and residency argument a government reader is already having internally, and positions us as vendor-agnostic and sovereign-first, which is the product's whole stance.

Points to hit
Control it ratifies
ISM / E8 Supports ISM controls on data sovereignty, classification handling and offshore access; the architecture keeps PROTECTED data inside the accredited boundary.
← All deep dives AI Infrastructure →