Google announced AI additions for its air-gapped Google Distributed Cloud offering. It describes general availability for Gemini 3.5 Flash and NVIDIA HGX B300, while listing Gemini-powered enterprise search as a preview. Google Cloud official announcement
Keeping data off the public internet does not remove model deployment, patching, import validation, or recovery work. An evaluation that stops at connectivity misses those dependencies. Match the product’s supported update process to your organization’s import controls.
On this page
Separate the infrastructure, model, and data layers
An air-gapped environment is separated from external networks. Running AI inside it still requires hardware, model-serving software, and a data layer connecting organizational records. Installed GPUs do not make document search available if model imports and indexing are unfinished. Read the announcement as changes across these layers, each with its own implementation and operating requirements.
General availability and preview status affect the deployment decision. For a preview, review support terms, expected changes, and the incident escalation path alongside the feature itself. A generally available model or hardware platform does not automatically give the search layer identical support conditions. Keep those entries separate in procurement material and in the operational readiness review.
Retrieval must preserve the permissions of the source
Retrieval-augmented generation, or RAG, finds relevant records and supplies them as context for a model’s answer. This is distinct from retraining the model on internal documents. Documents may be divided into chunks and indexed for retrieval. Access controls must therefore cover those chunks and their indexes as well as the files from which they originated.
Consider an HR employee and a general employee asking the same question. They should receive evidence appropriate to their permissions. Blocking the original file is insufficient if an indexed fragment remains available. Test document deletion and permission revocation against new searches and existing conversations to understand what information remains accessible inside the disconnected environment.
Review an actual update workflow
Make the operating plan concrete by walking through the arrival of a new model or software version. An organizational workflow might obtain files and signatures outside the boundary, complete validation and approval, and import them into an internal repository. Record the importer and approved version so the configuration can be reconstructed. This is a planning example; confirm GDC-specific import mechanics in product documentation and support agreements.
After the update, compare document permissions, evidence returned for representative questions, and recovery from backup. Better generated answers can coexist with regressions in indexing or authorization. Readiness depends on repeatedly operating updates and permission changes, rather than treating disconnection itself as sufficient proof that the complete AI service is ready.
Evidence to request before an air-gapped deployment
- Check availability and support separately for models, search, and accelerators.
- Document update import, signature verification, and backup and recovery procedures.
- Test search permissions and index changes using representative documents.
The announcement broadens air-gapped AI choices. Keep preview capabilities separate from generally supported components in procurement and operational planning.
Announcement reference date: 2026-10-21 · Checked: October 24, 2026
