Cloud Strategy

Google Launches Gemini 3.5 Flash Inside Air-Gapped Customer Sites

Sovereign AI · Cloud Strategy

Google Cloud now runs its Gemini 3.5 Flash model inside customer sites that have no connection to the public internet, the most controlled of its three sovereign AI deployments.

By Shashi Bellamkonda · November 14, 2026
3
sovereign deployments, from public regions to air-gapped sites (Google Cloud, 2026)
5
commitments in Google's sovereignty framework (Google Cloud, 2026)
12
months Google Cloud Dedicated is designed to run if its link to Google is cut (Google Cloud, 2026)
Google sells sovereign AI as three deployments, and its October 20 announcement puts Gemini 3.5 Flash in the most controlled one. Ask which model runs at the option you choose, and how its updates reach you.

Gemini 3.5 Flash is now generally available inside air-gapped Google Distributed Cloud sites, which have no connection to the public internet, according to Google's October 20 announcement. Google also made NVIDIA HGX B300 systems generally available to serve Gemini and open models inside these deployments. A Gemini search service for those sites is in preview, with the file storage, indexing and answers all kept inside the site.

That launch sits at the most controlled end of how Google Cloud handles sovereignty. Its sovereign cloud page lists five commitments (Google Cloud, 2026), including customer control of unencrypted data and an open digital future that avoids lock-in. Customers pick public cloud, a country-specific cloud or their own premises, and the page describes the AI stack as composable across Gemini, Gemma and third-party models.

Google's three sovereign deployments move operations closer to you

  • Data Boundary applies software-defined residency and access policies inside public cloud regions, with external key management and geographic limits on which Google personnel can reach your data. Google says you can build with Gemini and its Gemini platform here, with processing guaranteed to stay in region.
  • Dedicated is run by a local partner. In France, S3NS, a Thales company, operates PREMI3NS, and Google Cloud and Thales are expanding this model to Germany. Google describes GPU-accelerated AI workloads here without naming a model, and says the cloud is designed to keep running for up to 12 months if its connection to Google is severed (Google Cloud, 2026).
  • Distributed Cloud runs on commercial hardware and Kubernetes at your site, in a connected mode or an air-gapped mode. Google lists Gemini, open-weight Gemma models and sovereign enterprise search here, and says the air-gapped version cannot be remotely accessed or shut down by Google.

Google names Gemini under both Data Boundary and Distributed Cloud, and its Dedicated section describes GPU-accelerated AI workloads without specifying models. For how Google extends confidential computing to the chips that run a model, see my earlier post on Google Cloud's TPU 8tx.

S3NS lists open-weight models for its partner-operated AI services

Open-weight models make trained weights available for download and deployment under their licenses, and Google's October 20 post says Gemma and Mistral models are supported on Distributed Cloud. In France, where S3NS operates Google Cloud Dedicated, IT Social reported on November 6 that S3NS's managed AI services offer open-weight models from Mistral AI and Google, and that Gemini is absent from the list S3NS cites.

S3NS announced on December 17, 2025 that its PREMI3NS cloud earned SecNumCloud 3.2 qualification, a French government security approval (S3NS, 2025). S3NS's offer page marks its Gemini platform, Model Garden, Model Registry and online inference services as qualification in progress. A regulated buyer should verify that the qualified services cover its workload and confirm the managed AI services' qualification status before deployment.

Cirrascale describes how its appliance protects Gemini's weights

Cirrascale, a Google Cloud partner, announced an appliance in April 2026 that runs Gemini on a customer's premises. According to Cirrascale chief executive Dave Driggers's account to VentureBeat, the appliance holds the model in volatile memory, so cutting power removes it, and shuts down if its confidential-computing protections are violated. Driggers also said customers cannot extract or inspect the weights, which are the files that hold what the model has learned.

VentureBeat reported that new model versions reach the appliance through a brief private connection to Google, or by swapping the physical server for one that carries the update. In this appliance design, the customer controls where the data sits, who runs the site and whether the network connects, while the weights stay under Google's protection and updates follow a path Google sets.

A regulated buyer should verify that the qualified services cover its workload and confirm the managed AI services' qualification status before deployment.
For each workload, document the deployment and model version, then confirm the update process and applicable service qualifications.

Dedicated buyers should verify support beyond the disconnection window

Google's sovereign cloud page says an AI stack is only as sovereign as the layer you cannot control or replace, and the model is a layer that varies across these options. Open-weight models make that layer easier to move, because you can deploy their weights on hardware you control. Cirrascale's reported appliance design protects Gemini's weights through confidential computing, so buyers should confirm the protections and update procedures for their specific deployment.

Google designs Dedicated to keep running for up to 12 months if its connection to Google is severed (Google Cloud, 2026). That figure describes Dedicated only, so ask Google and your operating partner what support and updates apply beyond it. For exits, Google offers free data transfer for eligible migrations out of supported services, subject to its exit program requirements.

CIO/CTO Viability Question

For Google Cloud Dedicated, what operating support and update arrangements apply during a prolonged disconnection, including beyond the stated period of up to 12 months?

Before you commit a workload, have Google put in the contract the model and version you will run, how the selected Gemini deployment receives updates and recovers from a hardware failure while disconnected, and the terms for moving your data out.

Sources
Google Cloud. "Sovereign Cloud." Google Cloud, cloud.google.com/sovereign-cloud. Accessed 14 Nov. 2026.
Google Cloud. "New Innovations in Google Distributed Cloud." Google Cloud Blog, 20 Oct. 2026, cloud.google.com/blog/topics/hybrid-cloud/new-innovations-in-google-distributed-cloud.
Google Cloud. "Gemini Sovereign Enterprise Search on Google Distributed Cloud." Google Cloud Blog, 20 Oct. 2026, cloud.google.com/blog/topics/hybrid-cloud/managed-rag-search-inside-the-sovereign-perimeter-gemini-on-gdc-enterprise-search.
Google Cloud. "Data Transfer Fee Waiver (Exit Program)." Google Cloud, cloud.google.com/exit-cloud. Last updated 11 Sept. 2025.
Krim, Mourad. "S3NS ouvre l'IA générative sur PREMI3NS, et vend l'inférence à un tarif horaire prédéfini" [S3NS Opens Generative AI on PREMI3NS and Sells Inference at a Predefined Hourly Rate]. IT Social, 6 Nov. 2026, itsocial.fr/contenus/actualites/intelligence-artificielle-actualites-contenus/s3ns-ouvre-lia-generative-sur-premi3ns-et-vend-linference-a-un-tarif-horaire-predefini/.
S3NS. "PREMI3NS Trusted Cloud." S3NS, s3ns.io/en/offers/premi3ns-trusted-cloud. Accessed 14 Nov. 2026.
S3NS. "S3NS Announces SecNumCloud Qualification for Its PREMI3NS Trusted Cloud Offering." S3NS, 17 Dec. 2025, s3ns.io/en/news/premi3ns-secnumcloud-qualification.
VentureBeat. "Google's Gemini Can Now Run on a Single Air-Gapped Server and Vanish When You Pull the Plug." VentureBeat, 22 Apr. 2026, venturebeat.com/technology/googles-gemini-can-now-run-on-a-single-air-gapped-server-and-vanish-when-you-pull-the-plug.
Disclaimer: This blog reflects my personal views only. Content does not represent the views of my employer, Infotech. AI tools may have been used for brevity, structure, or research support. Please independently verify any information before relying on it.