Cities drown in data and starve for action — and NVIDIA now wants agentic AI to close that gap. The company has gathered partners across Spain, Poland, Malaysia, Morocco, Nigeria and Taiwan to run resident-service agents, vision agents and digital twins directly on city infrastructure, with the computing kept local so governments keep control of sensitive data.
The announcement lands this week as the Smart City Expo World Congress opens in Barcelona. It builds on initiatives first shown at GTC Berlin and pairs two ideas NVIDIA calls physical AI — cameras, sensors and robots in the real world — with agentic AI: language models that can plan, decide and carry out tasks. Here is what is actually deployed, who operates it, and where the watchful parts — surveillance, procurement opacity and vendor lock-in — start to get uncomfortable.
- Why Cities Are Turning to Agentic AI Now
- Agentic AI for Residents: 1MillionBot and Spain’s 150 Deployments
- LiveX’s AI City Brings Conversational Agents to Public Spaces
- City Vision Agents Are Rewatching Camera Feeds
- Digital Twins: From a Nigerian Refinery to Taiwan’s Trees
- Sovereign AI Factories Keep City Data at Home
- The Shared-Infrastructure Economics Behind Agentic AI for Cities
- What’s Next: Agentic AI Takes Over Smart City Expo
Why Cities Are Turning to Agentic AI Now
From sensor spray to coordinated action
Municipal governments collect more data than almost any other institution: traffic cameras, air-quality sensors, transit telemetry, 311 calls, utility meters, permit applications. Yet most of that data sits in departmental silos, reviewed by humans only after something has gone wrong. The mismatch between what cities know and what cities do is precisely the gap NVIDIA is targeting with agentic AI — autonomous language-model systems that monitor inputs, reason about goals, and execute multi-step tasks without waiting for a clerk.
The company frames the move as an extension of physical AI: sensors, cameras and eventually robots feeding shared on-premise compute, where agentic AI layers judgment on top. The pitch to mayors is political as much as technical. Because the infrastructure is installed in the city or its sovereign cloud, sensitive footage and resident records never leave local control — a direct answer to the data-residency objections that stalled a decade of “smart city” proposals. The plan spans Europe, Southeast Asia, Africa and Latin America, and leans on local partners to operate it.
| Region | Partner | Core stack | Mission |
|---|---|---|---|
| Spain | 1MillionBot + BSC | ALIA open models + Nemotron | Resident-service agents for government portals |
| San Francisco Bay | LiveX | Cosmos, Omniverse, Metropolis on RTX PRO | Conversational “AI City” hosts in public venues |
| Wrocław, Poland | K2K | Cosmos + Metropolis Blueprint for VSS | Shared AI infrastructure, camera search and summarization |
| Malaysia | ITMAX | SamurAI V2 vision agents on RTX PRO 6000 | Traffic management across 17,000+ cameras by 2027 |
| Lekki, Nigeria | Dangote + Honeywell | Forge Plant Cognition, Nemotron, Omniverse | Refinery digital twin and optimization agents |
| Taiwan | Linker Vision + MVRDV | Omniverse + Nemotron ESG agents | Urban-forest canopy simulation for planners |
| Tangier, Morocco | Cires Technologies / Tanger Med | Sovereign AI factory on NVIDIA compute | Port safety, congestion, cargo routing |
What the deployment map shows
Every entry shares one shape: a local operator, NVIDIA silicon and libraries, and a job a resident can actually feel — a permit answered faster, a corridor that drains at rush hour, a forest plan drawn in hours instead of weeks. That matters more than the agentic AI branding. Our analysis of visual AI in cities found the last wave of municipal AI failed exactly here: sensors without a loop back into action. Agentic AI is the loop.
Agentic AI for Residents: 1MillionBot and Spain's 150 Deployments
Open-source ALIA models meet Nemotron
Spain is the deepest live test of agentic AI in the public service. The 1MillionBot project combines ALIA — the open-source language model family built at the Barcelona Supercomputing Center — with NVIDIA’s Nemotron models, and wraps them in resident-service agents that handle real administrative work. The agents answer citizen inquiries, guide people through municipal procedures, book appointments, and triage urban-planning incidents.
Two large public bodies already run them: Madrid Digital, the technology arm of the Comunidad de Madrid, and Administració Oberta de Catalunya, the open-government platform serving Catalan municipalities. Together they account for more than 150 deployments across Spain — not pilots, production services with residents on the other end of every request.
What makes the deployment instructive is the licensing. The open-source core gives the Spanish state something a commercial API never offers: the right to read, audit and modify its own chatbot. For a category of software that mediates legal entitlements — benefits, permits, appeals — auditability is not a nice-to-have, and this is increasingly how European procurement writes the requirement for agentic AI systems.
LiveX's AI City Brings Conversational Agents to Public Spaces
A city as an interface
In the San Francisco Bay Area, the developer LiveX is building what it calls an “AI City” — and it is running on NVIDIA’s full rack. The agentic AI stack pairs Cosmos world foundation models (which predict how physical scenes evolve) with Nemotron language models, Omniverse 3D environments and Metropolis vision libraries, all executing on RTX PRO GPUs installed on site.
The visible layer is conversational: AI hosts staff airports, hospitals and event venues, speaking residents’ and visitors’ local languages rather than defaulting to English. Behind the kiosk sits the harder prize. An agentic AI layer that can see, simulate and answer becomes the operating system for the building it stands in — wayfinding, capacity, safety handoffs. LiveX is betting that conversational front desks normalize the deeper infrastructure behind them: once an airport trusts an agent to answer a passenger, it starts asking the same models to reroute the queue. This is the consumer-facing edge of NVIDIA’s urban strategy, and a useful preview of what other cities eventually procure.
City Vision Agents Are Rewatching Camera Feeds
Wrocław: one shared brain for many cameras
Wrocław is rebuilding its surveillance estate around shared AI infrastructure operated with local partner K2K. Instead of each agency buying separate analytics, the city funnels existing camera streams through Cosmos world reasoning models and the Metropolis Blueprint for Video Search and Summarization (VSS). Operators query feeds in plain language — find every intersection where a bus was blocked this week — and receive summaries instead of hours of tape. Mobility management, public-safety response and event handling all draw on the same video understanding, which is exactly what makes the economics work and exactly what civil-liberties advocates worry about.
Malaysia: 17,000 cameras and a traffic simulator
In Malaysia, ITMAX runs managed services that include traffic management, deploying SamurAI V2 vision agents on RTX PRO 6000 GPUs with a rollout of more than 17,000 cameras by 2027. The differentiator is TeraSim, the open-source traffic simulation engine from Laplace Intelligence: the vision agents feed live conditions into a calibrated digital replica of the corridor, and agentic AI tests signal plans inside the simulation before the city commits to them on the street. ITMAX reports peak-hour corridor travel times cut by up to 26%.
The surveillance question agentic AI doesn’t dodge
Using agentic AI to summarize a city’s camera network collapses the friction that once kept mass surveillance expensive: who asks, how long they may look, which feed they’re allowed to see. We covered the academic case for and against this trade-off in depth in our review of Visual AI in Cities: The Promise and Peril of Using Visual AI to Study Cities; the Wrocław deployment is that book’s argument walking. The mitigations worth demanding: query logging with named requesters, retention limits written into procurement, and an independent auditor with read access to the VSS indexes.
Digital Twins: From a Nigerian Refinery to Taiwan's Trees
The Lekki refinery: agents on the plant floor
The Dangote refinery at Lekki, Nigeria — one of the largest single-train refineries ever built — runs Honeywell Forge Plant Cognition on NVIDIA infrastructure, with Nemotron-powered agents watching process data for bottlenecks, equipment degradation and optimization openings. Compute is a mix of Dell PowerEdge servers and RTX PRO GPUs.
A full industrial digital twin on Omniverse is planned, so engineers can walk the simulated plant before touching the real one. For a city the size of the refinery’s footprint, the distinction between plant optimization and municipal infrastructure planning is blurring: the same twin toolchain that shaves an hour off a crude unit also schedules the power, water and roads serving it.
Taiwan: fifteen years of forest growth in ten minutes
In Taiwan, Linker Vision and the Dutch-Rotterdam architecture firm MVRDV use Omniverse together with Nemotron-powered ESG agents to simulate urban forestry. The system grows a fifteen-year canopy in about ten minutes of simulation, letting planners compare street-tree strategies before a sapling is ordered. Design iteration that once took roughly two weeks now completes in under an hour.
The pairing is telling: an ESG agent that can argue for trees, and a physics-grade twin that can prove where they will cast shade in 2041. Planning offices that adopt it stop debating canopy aesthetics slide-by-slide and start comparing measured cooling per street. NVIDIA’s simulation stack also powers the broader trend we covered in AI-powered VR city planning, where immersive twins put the same models in front of non-expert stakeholders.
Both projects illustrate why NVIDIA leads twin work with physical AI rather than language alone: the twin has to respect geometry, materials and time, while the agents on top bring the judgment. When the two finally fuse — agentic AI acting inside the twin to propose changes to a city — procurement documents will need a whole new vocabulary of controls.
Sovereign AI Factories Keep City Data at Home
Tanger Med: a port that owns its inference
Morocco’s Tanger Med port — Africa’s largest container hub — is working with Cires Technologies on a sovereign AI factory: dedicated NVIDIA compute physically hosted under Moroccan control, running worker-safety monitoring, congestion detection, cargo-routing optimization and port-operations analytics. The factory model deliberately inverts the cloud pattern. Rather than renting intelligence from a foreign hyperscaler, the port owns the machine, and an ecosystem of app providers builds services on top of it. Every vendor that wants to sell the port a model must ship weights to be run inside, where the port can inspect and govern them.
The Barcelona accelerator behind it
The same philosophy drives the Barcelona Supercomputing Center’s AI Factory Accelerator Programme, which gives startups access to sovereign computing capacity and has already seeded the open-source ALIA models underpinning Spain’s resident-service agents. Focus areas include agriculture, medical imaging, knowledge retrieval and speech — the model categories a public body needs and a general-purpose API rarely serves well.
The regional pattern recurs from Barcelona to Singapore: sovereign AI factories are how middle powers buy into agentic AI without outsourcing their institutions’ data. NVIDIA’s own Southeast Asia rollout rode the same wave, as we covered in AI Day Singapore, and the robotics arm of the strategy continues in agentic systems like Isaac ROS 5.0 — cities are simply where the two lines now meet.
The Shared-Infrastructure Economics Behind Agentic AI for Cities
Why shared beats per-agency
The business pattern is as important as the technology. Instead of transport buying a vision system, police buying another, and parks a third, NVIDIA’s city program has one operator — K2K in Wrocław, ITMAX in Malaysia, Cires in Morocco — stand up shared computing infrastructure that every department, municipality and local developer can draw from. Entry costs stay modest: an RTX PRO workstation or a small GPU system is enough to start, with datacenter capacity added only once workloads prove out. That “small front door, shared backbone” design is how vendors convert a single agentic AI pilot into a multi-year municipal platform.
| Stage | Hardware | Who pays | What it buys |
|---|---|---|---|
| Pilot | One RTX PRO workstation | Single department | One workflow, e.g. camera search on a few feeds |
| Shared backbone | Small GPU cluster, on-prem | City or regional operator | Many departments; sovereignty and reuse |
| City-wide | Datacenter-scale RTX PRO / PowerEdge | Operator + service contracts | Vision agents across thousands of cameras; local developer ecosystem |
The three risks procurement must price in
Shared infrastructure concentrates risk exactly where it delivers value. One queryable camera index means one breach away from a city-wide surveillance map; one operator means one renewal clause with leverage over every department that now depends on it. The mitigation column below is not marketing — it is the short list of contract terms an oversight board should require before signature for any agentic AI deployment.
| Risk | How it bites | Mitigation |
|---|---|---|
| Surveillance creep | Natural-language queries erase practical friction against mass monitoring | Named-requester query logs, retention caps, independent audits |
| Procurement opacity | Platform renewal quietly becomes a monopoly | Open interfaces (VSS blueprints), export rights for indexes and weights |
| Vendor lock-in | Proprietary runtime raises the cost of leaving | Open models (ALIA, TeraSim), local hosting, data-portability clauses |
| Data protection | Resident records processed outside legal jurisdiction | Sovereign hosting; the reason on-prem compute leads every deal |
The ethics layer still missing
Shared agentic systems will increasingly make resource trade-offs — which corridor gets green-wave priority, which neighborhood gets shade first. Our piece on LLMs as ethical advisers in urban design argued those decisions need documented human ownership, not just model cards. NVIDIA’s announcements describe compute and models; none yet describe a governance API for agentic AI in the public sector. Cities signing now should ask for one in writing.
What's Next: Agentic AI Takes Over Smart City Expo
Barcelona, November 3–5
The announcements arrive as Barcelona hosts the Smart City Expo World Congress, where NVIDIA’s EVP for EMEA, Paolo Guglielmini, delivers a keynote and joins a panel moderated by Charbel Aoun. Expect the resident-service agents, VSS camera deployments and sovereign AI factories above to be presented as a finished playbook — because commercially, they are.
What to watch after the keynote
Three signals will tell whether this wave of agentic AI sticks. First, procurement language: whether renewal and data-portability clauses appear in the first generation of contracts. Second, open-model share: whether ALIA-style open weights keep winning public tenders or give way to proprietary endpoints. Third, governance: whether any city publishes its query logs for a video-search index — the single most informative transparency act available in this domain. The physical-AI side of the story keeps moving too — our report on humanoid robots finally learning to grab shows how quickly the actuator half of “physical plus agentic” is maturing.
For a decade, smart-city pilots died in the space between insight and action. Agentic AI is that missing muscle — and NVIDIA has now bolted it to city halls on four continents in a single news cycle. The cities that thrive will be the ones that write the off-switch into the contract while the vendor is still smiling. The rest will discover that the agent they bought for the queue also runs the queue, the camera, and eventually the budget.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
