secunet details planned AI cloud for classified work with Red Hat and deepset

secunet on October 26 detailed a planned AI cloud for organizations handling classified information, developed with Red Hat and deepset using NVIDIA technologies. The announcement adds security architecture and delivery steps to a collaboration NVIDIA had already described on October 21, rather than marking an operational service launch.

According to secunet, the design builds on SINA Cloud and uses NVIDIA BlueField-3 processors to isolate security tasks at hardware level. Existing SINA security components have German BSI authorization for classified use. That authorization can support approval of particular implementations; it does not establish that every application on the future platform is already approved.

NVIDIA's earlier account explains the partners' division of work: accelerated computing and Nemotron open models supply the AI foundation, Red Hat OpenShift AI manages workloads, Haystack orchestrates agents, and secunet operates infrastructure in German data centers. This combines a German cloud operator with technology from international suppliers. The relevant control is where and how the system operates, alongside choices about models and applications, rather than an assertion that every component originates in Europe.

The application layer matters because a language model alone does not know an organization's internal files. Haystack's documentation describes retrievers that select documents relevant to a question and pass them to another pipeline component. In a prospective agency application, that pattern could help staff search an approved collection of administrative guidance before preparing an answer. This is an illustration of retrieval-augmented generation, or RAG, rather than a named customer deployment announced by the consortium.

Finding useful material also requires measurement. Haystack's evaluator documentation includes checks comparing retrieved documents with expected results and comparing generated answers with reference answers. An organization could therefore test whether a prototype retrieves the right instructions and answers representative questions before introducing it into work. These software capabilities help explain what teams can build and evaluate; a retrieval score or answer-quality score is not evidence that a system has passed a classified-information security review.

Red Hat's deployment documentation supplies another practical distinction. OpenShift AI can be installed in restricted, disconnected networks by mirroring required software images into a private registry. Storage used by the relevant components must remain reachable inside that network, and administrators must configure authentication and other dependencies. Those documented platform options explain how locally controlled AI can be assembled. They do not establish which disconnected configuration the consortium will ultimately sell or remove the operational work of maintaining software, storage and identities.

For a government buyer or a company supporting classified work, the prospective value is a place to combine models with approved internal knowledge under a defined operating arrangement. Evaluating such an application would involve both useful answers and appropriate data access. A successful document search demonstration would address only part of that decision: the buyer would still need to understand which information a user may retrieve, which components run the application and who maintains the environment.

secunet says integration and validation of the components are the next steps, followed by required authorization and release procedures before provision. Its announcement specifies neither prices nor a general availability date. The offering therefore remains a development project whose delivery depends on those implementation and approval steps.