Google Rolls Out Gemini On-Site

To help with search-and-rescue operations, Singapore is betting on… Madagascar cockroaches.

The first real-world deployment of the device occurred in late March, following the earthquake in India and Myanmar (magnitude 7.7; at least 5,000 dead). The insects carried infrared cameras with onboard algorithms. Their movements were controlled by electrodes.

The Ministry of Home Affairs’ science and technology agency contributed to the development of these “cyborgs,” as they call them.
In early April, its Director of Innovation referenced them, among other projects, during Google Cloud Next. “Much of all this could not be done without the cloud,” he said. And he briefly recounted the agency’s history in the public cloud. The switch began in 2019. Three years later, a sovereign-hybrid strategy had been put in place to bolster local controls (data halls, management of data flows, verified personnel…) to ensure physical control over data.

An endpoint Gemini in disconnected mode

This year, with the goal of securing AI deployments, the footprint extended to GDC (Google Distributed Cloud). This offering adapts GKE for use in on-site environments. It is offered as software-only (installable on VMware or bare metal); as well as with hardware, in either connected or air-gapped form.

For the connected version (formerly GDC Edge), the hardware provided consists of standalone servers (1U) or racks (42U) with integrated switches. The disconnected version (formerly GDC Hosted) is deliverable in racks or as an appliance (2 CPU blades, 1 GPU blade and networking).

On the AI side, GDC includes the Vertex AI stack and APIs for various Google models (speech synthesis, translation, OCR). On the occasion of Google Cloud Next, we were promised, for the third quarter of 2025, the addition of a managed Gemini endpoint. This has officially been done, now generally available on GDC air-gapped and in preview on connected GDC*.

In Google’s cloud, Gemini runs on TPUs. With GDC, it runs on NVIDIA GPUs (H200, B200, B300).

Supplementary reading:

What on-prem infrastructure is needed to carry an AI strategy?
GTC 2025: NVIDIA goes all-in on inference
Sovereign cloud: Google answers Microsoft’s data boundary
Routers, firewalls, VPNs… How to secure the edge

* We will note another preview : that of the Agentspace search service.