Cities Worldwide Build and Scale AI Services With NVIDIA Physical and Agentic AI

Ahead of Smart City Expo World Congress in Barcelona, NVIDIA rounded up city deployments built on its models and hardware: resident-service agents across Spain, vision agents reading city camera feeds in Poland and Malaysia, and digital twins for planning in Taiwan and industrial sites in Nigeria.

Some claims are specific. ITMAX in Malaysia is scaling vision agents toward more than 17,000 cameras by 2027 and says traffic simulation cut peak-hour corridor travel times by up to 26%. Linker Vision and MVRDV say they can simulate 15 years of urban canopy growth in 10 minutes.

The message to city governments is that getting started doesn't take a huge investment. That lowers the barrier for useful services and for surveillance alike.

We see real upside here. Resident-service agents that handle appointments and routine municipal procedures in local languages can save people time, and Spain's 1MillionBot deployments, using open ALIA models from the Barcelona Supercomputing Center, are a good sign that this can be built with local and open tools. Planning tools that shrink a two-week design iteration to under an hour could make public consultation faster and better informed.

The vision agents are a different matter. Turning existing camera streams into searchable, summarized intelligence for mobility, safety and event response is a major expansion of what a city can know about people in public space. The cameras might already be there, but the capability isn't the same. A feed nobody watches is very different from one an agent can search and summarize on demand.

NVIDIA's framing stresses that local governments stay in control of how sensitive data is processed. That's a useful principle, but control isn't the same as limits. We'd want each city to publish what the agents look for, how long data is kept, who can query it, and whether it's shared with police or other agencies. The post doesn't address any of that, and it shouldn't be left to vendors.

Note the 26% figure is an upper bound from a partner, not an audited citywide result. Vendor blogs are a fair place to announce deployments and a poor place to verify them.

If this scales the way NVIDIA hopes, shared city AI infrastructure becomes normal municipal plumbing. That can be good. It needs the same public oversight as any other piece of civic infrastructure that watches residents.

NVIDIA's post builds on initiatives announced at GTC Berlin and promotes its Cosmos world models, Nemotron open models, Omniverse and Metropolis libraries, and RTX PRO GPUs. Smart City Expo World Congress runs Nov. 3–5 in Barcelona.