Nvidia physical agents enter urban facilities: shared computing and local operation of sensitive data

The management of facilities and cities is moving from monitoring screens and aggregating camera data to delegating field intervention to physical AI agents. The operational launch announced by Nvidia in partnership with its collaborators, based on initiatives from the GTC Berlin conference and the Smart Cities Forum in Barcelona, rests on a new economic and regulatory equation: providing shared computing infrastructure that gives municipalities, local government agencies, and developers a unified operating base, instead of requiring each institution to purchase and manage independent data centers, while keeping the processing of sensitive records and monitoring flows under the local control of cities.

This path starts with daily resident services and extends to visual perception agents, without the need for massive infrastructure; workstation equipped with RTX processors or small local servers suffice to run these models. In Spain, the company One Million Bot integrated high-end open models trained at the Barcelona Supercomputing Center with Nvidia’s NeMoTron models to run municipal service agents for Madrid Digital and Open Catalonia across more than 150 deployment sites, monitoring administrative procedures and urban planning notifications.The core of this experiment lies in linking conversation to the direct execution of municipal transactions within government systems.In the San Francisco Bay Area, Live X employed Cosmos world models, open NeMoTron models, and OmniVerse and Metropolis libraries on RTX Pro processors to build virtual hosts that speak local languages and provide interactive services in airports, hospitals, and public facilities.

At the road and public safety level, cities rely on vision agents to transform existing camera streams into immediate responses for crowd and incident management. In the Polish city of Wroclaw, K2K built a shared accelerated computing infrastructure that uses Cosmos inference models and the Metropolis planner to summarize video and search within it, serving multiple municipal sectors with the same equipment. In Malaysia, IT Max offers long-term managed service packages for traffic management, expanding the scope of the Samurai V2 vision agents on RTX Pro 6000 processors to cover more than 17,000 cameras by 2027, and integrating the open TerraSim platform from Laplace Intelligence to test signal timing.This reduced travel times in congested corridors by up to 26 percent.

This pattern extends to critical infrastructure and environmental planning through digital twins, where Hanyuel uses NeMoTron models in the Dangote refinery in Nigeria’s Lekki region to understand operational data, and creates a digital twin via the OmniVerse platform and Del Power Edge servers to predict equipment bottlenecks and train operational crews. In Taiwan, Linker Vision and Dutch architecture firm MVRDV integrated simulation tools to evaluate 15 years of vegetation cover scenarios in just ten minutes, accelerating design from two weeks to less than an hour. Ceris Technologies also expands a sovereign computing accelerator program at the Port of Tangier Med in Morocco, adding an AI-driven safety and congestion analysis plant for workers and cargo routing, alongside a sovereign computing accelerator program at the Barcelona startup hub for agriculture, medical imaging, and audio processing.

This model places municipal officials, port operators, and transport managers in the Gulf, Egypt, and the Levant before different asset-management accounts.The actual transformation no longer requires draining budgets to build isolated central computing complexes, but to operate a shared local processing infrastructure that protects data confidentiality.If you lead a technical team in a transport authority or an economic city, the change that imposes itself is moving from slow human monitoring of thousands of screens to running inference agents that test traffic and logistics decisions in simulation environments before field deployment, turning the engineering demand from merely dealing with software conversational interfaces to tuning world models and live digital twins.