Summary
- Google plans to bring Gemini models to Google Cloud Dedicated, its independently operated sovereign cloud offering.
- Confidential External Key Management is in preview, while new compliance frameworks will help customers document technical controls.
- France's S3NS already offers managed model inference, and Google is extending Dedicated into Germany with Thales.
Google Cloud is preparing to make its Gemini artificial intelligence models available through independently operated sovereign cloud environments in Europe, expanding the deployment options for organisations that require greater control over their infrastructure and information.
Announced at Google’s Sovereign AI Summit in Brussels on 17 November, the development will bring selected Gemini capabilities to Google Cloud Dedicated, an isolated infrastructure service operated by local partners rather than directly by Google.
The company is also introducing new encryption-key controls, regulatory audit frameworks and partnerships with European software providers. The measures address different obstacles to AI adoption in regulated industries, including restrictions on data access, operational responsibility and the ability to demonstrate compliance.
Google has not specified when Gemini will become generally available through Dedicated or which models will be supported initially. Its announcement therefore establishes a planned expansion rather than a completed deployment across its European sovereign cloud network.
Different levels of operational control
Google offers three principal approaches to sovereign cloud deployment, each involving different infrastructure and management arrangements.
Google Cloud Data Boundary provides residency, access and encryption controls within the company’s public cloud. Google Cloud Dedicated operates on isolated infrastructure managed by an independent local partner, while Google Distributed Cloud can be deployed within customers’ facilities, including environments disconnected from public networks.
These distinctions extend beyond the country in which information is stored. An organisation may retain its data in a European cloud region while continuing to depend on the provider’s technical staff, operating procedures and software update arrangements.
Partner-operated infrastructure offers additional local authority over those activities, although it continues to use technology developed by Google. The resulting degree of independence depends on the contractual and operational arrangements of the particular service.
Selected Gemini models already support commitments to process customer data within designated regions, including the EU, and Gemini capabilities are available through Google Distributed Cloud. Adding Dedicated will create another option for organisations that need advanced AI within a locally operated environment.
France’s S3NS provides an example of the model. Its PREMI3NS service, built on Google Cloud Dedicated, now offers generally available managed model serving and inference.
Customers can deploy supported open models through Model Garden or bring their own models, while S3NS operates the environment. Google says the arrangement is designed to keep deployed models and prompts inside infrastructure controlled by S3NS.
PREMI3NS has achieved SecNumCloud 3.2 qualification from French cybersecurity agency ANSSI. The qualification applies to the assessed service and its operating arrangements rather than automatically certifying every application or model subsequently deployed within it.
Google is also working with Thales to extend Dedicated into Germany, where the service remains in preview.
Encryption controls address access dependencies
Customers can already use Google’s external key management capabilities to hold encryption keys outside its infrastructure, giving them greater control over whether encrypted information can be accessed.
However, consulting an external key management system for individual operations can introduce latency and availability risks. A communication failure or interruption to that system may affect applications that require frequent access to protected information.
Confidential External Key Management, or cEKM, is intended to reduce that dependency. The new capability allows customers or designated local partners to retain control over root encryption keys outside Google Cloud while selected key operations occur within hardware-attested confidential computing environments.
Google says the arrangement can reduce latency while maintaining external control over root keys. The technology is currently in preview, and the company has not published comprehensive independent operational results covering different workloads and failure conditions.
A separate arrangement with Telefónica in Spain, also in preview, expands the options for locally managed keys.
These mechanisms address technical control over encryption and access, but they cannot independently determine whether an AI application uses information appropriately. Identity management, application permissions, monitoring and procedures for responding to incidents remain necessary.
Compliance evidence becomes more automated
Google is extending its Data Boundary frameworks to associate regulatory requirements with specific technical controls that can be monitored and audited.
A framework for Switzerland is available, while one addressing the EU’s Digital Operational Resilience Act is planned for financial services organisations.
The company’s Audit Manager allows users to start from prepared frameworks, adapt them to internal policies and generate machine-readable evidence of their compliance arrangements.
Such tools could reduce the administrative burden associated with repeated security assessments, particularly where organisations use numerous cloud services and need to demonstrate consistent controls across them.
Nevertheless, automated evidence collection cannot replace an assessment of the organisation’s actual obligations. Financial institutions remain responsible for examining outsourcing arrangements, operational resilience, recovery procedures and concentration risks associated with technology providers.
Nor does a technical control provide assurance beyond the activities it actually covers. A system demonstrating that information remains within a designated region may not establish whether every employee, supplier or AI agent has appropriate access permissions.
European software becomes part of the infrastructure
More than a dozen independent software vendors have validated products for Google Cloud Dedicated, including European businesses Memority and Tsuga.
The additional software addresses an important practical constraint. Organisations cannot run sensitive applications effectively if their identity management, monitoring and security services are unavailable in the environment selected for those workloads.
Google has also introduced an equity-free Sovereign AI accelerator for European startups operating in regulated sectors, offering engineering assistance, mentoring and access to local cloud providers.
The company’s earlier European confidential AI expansion concentrated on computing infrastructure, isolated environments and protecting information during processing. The latest changes extend those capabilities towards model availability and operational compliance.
Google says Dedicated allows local operators to block software updates in exceptional circumstances and continue operating for up to 12 months if their connection to Google is severed. These are claimed design capabilities rather than guarantees that every application can operate without external dependencies.
Customers will still need to test how workloads behave during outages, how software maintenance is coordinated and which services remain dependent on infrastructure outside the locally controlled environment.
The proposed Gemini expansion adds another deployment option for European organisations seeking advanced AI without relying entirely on centrally operated public cloud services. Its practical value will depend on the supported models, availability dates, operating contracts and evidence that local control remains effective under normal and disrupted conditions.












