Enterprise search moves inside disconnected environments

Google has expanded Google Distributed Cloud air-gapped with a preview of enterprise search powered by Gemini. The managed architecture can index and query text, audio and video while retrieval-augmented generation remains inside the sovereign perimeter. It targets defence, intelligence, critical infrastructure and regulated organisations that cannot send sensitive material to a public-cloud endpoint.

The alternative has often been a custom stack combining speech recognition, optical character recognition, vector search, model serving and a user interface. A packaged service can shorten deployment, but it concentrates trust in the indexing and retrieval pipeline. Customers should verify supported formats, deletion behaviour, classification boundaries and whether citations lead back to the exact local source.

Gemini and open models share the governed platform

Gemini 3.5 Flash is now generally available on GDC air-gapped for low-latency reasoning, multimodal analysis and long-context work. The platform also supports Gemma and Mistral open-weight models, allowing operators to select different models for sensitivity, performance or mission needs. Running several models under one control plane may simplify governance if their distinct licences and safety profiles remain visible.

Air-gapped deployment does not make a model inherently safe. Updates arrive less frequently, local administrators have powerful access and generated content can still leak through authorised exports. Organisations need signed artefacts, offline evaluation, model inventories and a controlled promotion process from testing to the isolated production environment.

Blackwell Ultra raises local serving capacity

NVIDIA HGX B300 systems based on Blackwell Ultra are generally available for GDC air-gapped. Google says virtual-machine networking now reaches 32Gbps, up from 25Gbps, helping move data to accelerators. The upgrade is intended to serve frontier and open models as well as agent workloads without moving computation outside the facility.

Local hardware brings capacity planning back to the customer. Demand spikes cannot automatically borrow unlimited public capacity, and large accelerators require power, cooling, spares and skilled maintenance. Buyers should size for critical service levels, decide which workloads can queue and preserve a degraded mode if a node or model service becomes unavailable.

Networking and encryption reduce operational overhead

Managed Cloud DNS is generally available and can be administered with familiar Kubernetes interfaces, avoiding a separate DNS stack inside the air gap. Hardware-offloaded IPsec protects node-to-node traffic at up to 100Gbps, according to Google, reducing the CPU cost of encryption. Elastic Security integration adds behavioural analysis, automated case management and telemetry correlation.

Those features are valuable only if they work without external dependencies. Operators should test certificate renewal, name resolution, detection content updates and incident workflows during prolonged disconnection. Security telemetry itself can contain sensitive operational details, so its storage and any later export need the same classification controls as the source workloads.

A partner marketplace broadens the local stack

Teradata, Solo.io and Mattermost are joining the GDC partner marketplace. Their software covers analytics, API and service-mesh governance, agent networking and secure collaboration. Running familiar products near local inference can reduce data movement and help teams integrate AI with existing workflows inside the perimeter.

Every additional product also adds components, administrators and update channels. Customers should map which vendor can access which layer, how vulnerabilities are delivered into the air gap and whether licences continue through disconnection. An agent gateway should enforce tool and agent reachability independently of whatever instructions a model generates.

Sovereignty needs evidence, not topology alone

Google’s release brings models, multimodal search, accelerated infrastructure and enterprise software into one isolated platform. That can make sensitive AI practical where public endpoints are unacceptable. It does not settle legal jurisdiction, personnel trust or operational resilience by itself.

A strong deployment will document the complete supply chain, verify that data never crosses the boundary, rehearse recovery and test every model against mission-specific failure modes. The architecture is most compelling when air-gapping is paired with disciplined evidence: signed software, controlled updates, observable retrieval and clear human authority over consequential actions.

Operational test

Before production use, teams should document the exact product or model version, supported region, source date and evaluation conditions. They should repeat representative tests after any model, compiler, connector, policy or infrastructure change. Published capability and performance claims are useful starting points, but confidence must come from evidence collected in the organisation’s own environment. A named owner, measurable acceptance threshold, incident path and tested fallback turn a promising announcement into a service that can be governed over time.

Evaluation should cover ordinary work, boundary cases and deliberate misuse rather than a polished demonstration alone. Results need to be sliced by language, user role, data sensitivity and consequence so a strong average cannot hide a serious failure. Teams should also estimate the human work required to review exceptions and maintain the surrounding controls. If that effort is omitted, an apparently efficient feature may simply move cost and risk into operations, security or frontline staff who were not included in the original business case.