Google’s Gemini arrives on-premises for regulated industries

Google’s most advanced large language model has just crossed a critical threshold for enterprises that can’t compromise on data privacy. In a strategic move timed to align with Google Cloud Next 2026, Cirrascale Cloud Services unveiled a first-of-its-kind deployment model: a fully air-gapped appliance running the complete Gemini model without ever connecting to the internet or Google’s public cloud.

The announcement arrives at a pivotal moment for industries where regulatory scrutiny and data sovereignty demands have historically clashed with generative AI’s promise. By packaging the full version of Gemini into a hardened hardware platform, the offering eliminates a long-standing tradeoff between model capability and operational control.

A Hardware Vault for the AI Crown Jewel

The solution combines Google Distributed Cloud’s orchestration layer with Cirrascale’s custom appliance design, built on Dell hardware and certified to run the complete set of model weights. Unlike conventional on-premises AI deployments that merely host a stripped-down variant, this approach delivers the full Gemini experience while ensuring zero exposure beyond the customer’s own facility.

Each appliance ships with eight Nvidia GPUs and layers confidential computing protections that treat the model itself as an ephemeral asset. According to Cirrascale CEO Dave Driggers, the architecture was designed to address what he calls “the trust collapse” in public cloud AI services.

“Companies reached a point where they realized prompts, inputs, and outputs all belong to the hyperscaler by default. That was the inflection moment when secure, on-premises AI became non-negotiable.”

Zero-Tolerance Security Through Ephemeral Design

Security isn’t just layered; it’s fundamentally embedded into the appliance’s behavior. The model lives entirely in volatile memory, meaning power loss or termination of a user session immediately purges all associated data. Driggers emphasized that this extends to tamper detection: any attempt to breach the confidential computing envelope triggers a hardware lockdown so severe that the unit must be returned to Dell or Google for forensic review.

“If confidential compute is violated, the machine doesn’t just shut down—it self-destructs its model state and leaves an audit trail. It’s effectively a time bomb designed to protect Google’s intellectual property.”

For customers with the strictest compliance mandates, even firmware updates are handled through physical swaps. A new appliance arrives preloaded with the latest model version, while the old unit undergoes a cryptographic purge before being decommissioned.

Who Needs Air-Gapped AI—and Why Now

The demand surge isn’t theoretical. Driggers points to three sectors driving adoption at scale:

• Financial services: Institutions handling market-sensitive data face regulatory frameworks that forbid third-party data residency.
• Healthcare and life sciences: Drug discovery pipelines and patient records require watertight isolation from external networks.
• Defense and government: Classified workflows demand air-gapped environments where even metadata leakage is unacceptable.

These organizations have long relied on proprietary or open-source models that lacked the sophistication of frontier AI. Now, they can harness the full power of Gemini without surrendering control, prompting Driggers to call this deployment “the next evolution of the partnership” with Google.

What Comes Next for On-Premises AI

The preview launch signals only the beginning of a broader shift toward hardware-enforced AI sovereignty. As model weights grow larger and regulatory pressures intensify, the appetite for completely isolated AI infrastructure will likely accelerate. Cirrascale and Google are already exploring ways to streamline offline updates while preserving the air-gapped promise.

For enterprises caught between innovation and compliance, the message is clear: the future of enterprise AI won’t be confined to the public cloud. It will be built on hardware that disappears when it’s no longer needed—and that’s exactly the point.