Google brings Gemini search and Blackwell Ultra systems to air-gapped clouds

Google’s [October 20 announcement](https://cloud.google.com/blog/topics/hybrid‑cloud/new‑innovations‑in‑google‑distributed‑cloud) packages models, retrieval, accelerated infrastructure and partner software for Google Distributed Cloud air‑gapped. The intended users are organizations that must keep sensitive data inside a disconnected or tightly controlled perimeter, including defense, public‑sector and regulated environments. The central product change is not simply another model endpoint. Google is offering a more complete local stack so teams do not have to assemble search, embeddings, model serving, GPUs, networking and security tools from separate systems. That convenience is a company claim rather than proof that a deployment meets a specific legal or mission requirement. Buyers still need to map each component to their own data classes, network zones, administrators and continuity plan.

The enterprise‑search capability is in preview. Google says it can index and retrieve text, audio and video inside the sovereign perimeter, using Gemini embedding models and retrieval‑augmented generation. The interface is meant to let analysts query and summarize repositories without sending those records to a public cloud. A local retrieval path can reduce data‑movement risk, but location alone does not establish trustworthy answers. Teams should measure citation coverage, permission filtering, stale‑document behavior and the system’s response when two authoritative records conflict. They should also record which source passages supported every consequential answer. Because the feature is a preview, availability, support and interface behavior may change. Treat it as an evaluated component rather than a finished control environment.

Google says Gemini 3.5 Flash is generally available on GDC air‑gapped, alongside support for Gemma and Mistral open models. NVIDIA HGX B300 systems based on Blackwell Ultra are also generally available. The announcement says VM networking now supports up to 32 Gbps, compared with 25 Gbps before, while hardware‑offloaded IPsec can protect node‑to‑node traffic at up to 100 Gbps. These figures describe infrastructure ceilings, not guaranteed application performance. Retrieval latency will also depend on corpus size, chunking, embedding work, storage and concurrent model traffic. Benchmark the exact mix at representative concurrency and preserve cold, warm and failure‑path results separately. The partner layer adds Teradata, Solo.io and Mattermost. Their presence broadens the software available inside the perimeter, but each integration brings its own identity, update, support and supply‑chain boundary.

Start with one bounded corpus whose authority and access rules are already understood. Confirm that indexing remains local, users only retrieve permitted records and citations resolve to the exact retained source. Then test revocation, document replacement, stale indexes and disconnected recovery. For agentic uses, keep search separate from action authority. A model may retrieve an accurate procedure without having permission to execute it. Bind every action to a named principal, narrow tool scope, current revision and idempotency key, and require human review where policy or safety demands it. The announcement gives regulated teams a clearer managed route to on‑premises multimodal AI. The practical result depends on whether those controls survive the real local deployment, not on the air‑gap label alone. *Figure by OpenTools Team; no third‑party expressive material used. It distinguishes product availability from the acceptance checks a local AI system still needs.*