Google takes Gemini into its sovereign cloud: what NZ firms get

By DataForge team18 November 20265 min read

Google is adding Gemini to its partner-run European sovereign cloud, and NZ buyers can use the new key and audit controls as a checklist for any AI supplier.

What Google announced in Brussels

Google announced on 18 November that Gemini models are coming to Google Cloud Dedicated. Dedicated is Google's sovereign cloud offering run by local partners. Customers on Dedicated have had only the open models in Google's model garden until now. Google's post gives no date and no list of countries for Gemini on Dedicated. Techzine reported the models as available already. We're going with Google's own wording, which describes them as on the way.

Computer Weekly's report from Google's Sovereign AI Summit in Brussels sets out how the products fit together. Google Cloud Data Boundary adds residency, access and encryption controls to the public cloud. Google Cloud Dedicated is isolated infrastructure run by an independent local partner. Google Distributed Cloud runs in the customer's own data centre and can be connected or fully air-gapped. Gemini already runs on Distributed Cloud. In Germany, Dedicated is in preview with Thales from three Berlin data centres, and full operations are due in the second quarter of next year.

The model news got the headlines, but the control features will matter longer. Confidential external key management (cEKM) is now in preview. With cEKM, the customer or a local partner holds the root encryption keys outside Google Cloud, and key operations run inside a hardware-attested confidential computing environment. A Telefónica partnership adds Spain to France, Germany and Italy as countries where organisations can hold keys locally. Google also launched Data Boundary frameworks, which map regulatory requirements to enforced controls. A Swiss framework is available now, and one for the EU's Digital Operational Resilience Act is planned. Audit Manager gives auditors continuous, machine-readable evidence of those controls. Computer Weekly also reports that Dedicated customers can block software updates in exceptional circumstances and keep running for up to 12 months if their connection to Google is cut.

Where the control sits

Google's European push answers a question NZ buyers are starting to ask too. Data stored in a local region still leaves the bigger risks open: who can reach the data, who holds the keys, and whether the service keeps running if the provider is cut off. Marianne Janik, Google Cloud's vice-president for EMEA North, told Computer Weekly that a stack is only as sovereign as the layer the customer can't control or replace. As a vendor argument the line is convenient, but it's also a fair test for buyers.

For an NZ business, the useful part of the announcement is the shape of the controls. Holding the root keys outside the provider means a legal demand or an insider at the provider can't decrypt data without the customer's cooperation. Machine-readable audit evidence turns a sovereignty claim into something an auditor can check. A defined period of disconnected operation gives the board a number to plan against. NZ buyers can ask any AI supplier for those three things, whether the supplier is a hyperscaler, a SaaS vendor with an AI feature, or a local hosting firm.

What NZ firms can buy today

None of the 18 November announcements covers Australia or New Zealand. Google's post describes the sovereignty offerings in European terms only, and the named partners operate in France, Germany, Spain and Italy. An NZ business can't sign up for Gemini on a locally operated Google sovereign cloud here.

The nearest option is regional processing in Sydney. Google's Vertex AI data residency page lists Gemini 3.5 Flash and Gemini 2.5 Flash with machine learning processing in australia-southeast1. Google states that locational endpoints keep processing within the stated jurisdiction. The page lists no New Zealand region, and the Melbourne region has no stated machine learning processing commitment. Anyone in NZ using Gemini through Vertex AI can therefore keep prompts and outputs in Australia for those two models, as long as they call the regional endpoint. Calls to the global endpoint, or to models outside that list, carry no such commitment.

Australian processing satisfies plenty of NZ contracts. It won't satisfy a client, iwi partner or government agency that requires data to stay in New Zealand, or requires that no foreign operator can reach it. For those workloads, NZ organisations can choose between Google Distributed Cloud in their own facility and open-weight models on hardware they control. Google hasn't published NZ pricing or availability for running Gemini on Distributed Cloud. A smaller organisation will usually find a local model server easier to justify, and the trade-off is a smaller model with no Google support.

What the announcement changes for NZ buyers

Gemini on Dedicated widens the gap between the controls European customers can buy and the controls offered in New Zealand. Expect hyperscaler sales teams to describe "sovereign AI" here in the same language. The products behind that language here are mostly regional residency in Sydney plus customer-managed keys, without the partner operation or disconnected running that European customers will get. Buyers should ask which tier a proposal actually uses.

The announcement also gives procurement and IT teams better questions. Before October, many NZ AI contracts asked only where data is stored. A more complete contract also asks where data is processed, who operates the service, where the root keys sit, what audit evidence the supplier gives you, and how long the service keeps running if the supplier withdraws. Our AI consultancy work increasingly starts with that list, because a model choice is hard to reverse once staff workflows depend on the model.

Practical next steps

  • List every AI workload and where its data goes. Record where prompts are processed, where logs and outputs are stored, and which company operates the service. Include AI features built into SaaS tools as well as direct model APIs.
  • Check your Gemini endpoints. If you use Vertex AI, confirm that requests go to the australia-southeast1 regional endpoint, and that the model you call is on Google's list of models with in-region processing.
  • Ask every AI supplier who holds the root keys. Customer-managed keys stored with the provider differ from keys held outside the provider. Get the answer in writing.
  • Ask for audit evidence in a format your auditors can use. A residency statement in a contract is weaker than continuous logs showing where data sat and who accessed it.
  • Set a continuity target. Decide how long each critical AI workload must keep running if the supplier cuts access, and test suppliers against that number. Google's 12-month figure for Dedicated is one benchmark.
  • Keep the most sensitive data on hardware you control. For data under an NZ-only requirement, price an open-weight model on a private local AI server against waiting for a hyperscaler sovereign offering that hasn't been announced for this region.

Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.

Get my free assessment How we help
Get new articles by RSS