NVIDIA Collaborates with Partners to Introduce Innovative AI and Smart City Solutions in Dublin, Ho Chi Minh City, Raleigh, and Beyond

NVIDIA Blueprint Unifies Digital Twins, VLMs, and Edge Vision AI for City-Scale Deployment

Urbanization is accelerating, and the pressure on public services is sharply rising. The United Nations projects that two-thirds of humanity will live in cities by 2050, adding roughly 2.5 billion people to urban areas. That shift intensifies the need for responsive transportation, resilient infrastructure, and data-driven safety measures. Against this backdrop, NVIDIA and an expanding partner ecosystem are deploying physical AI—AI that perceives, reasons, and acts in the world—to help cities operate with real-time intelligence from curb to cloud.

At the Smart City Expo World Congress in Barcelona, the company’s latest advances coalesce in the NVIDIA Blueprint for smart city AI. The framework integrates high-fidelity digital twins via NVIDIA Omniverse libraries, synthetic data generation, and vision language models (VLMs) with video analytics agents built using the Blueprint for Video Search and Summarization (VSS) on NVIDIA Metropolis. New NVIDIA Cosmos world foundation models and VLMs unlock photorealistic data and physical reasoning, while updated cookbooks—Cosmos Predict, Cosmos Transfer, and Cosmos Reason—provide step-by-step recipes for intelligent traffic and safety workflows.

Why does this matter now? The smart traffic management segment alone is projected to reach $20 billion by 2027. But traffic is just the start; the same stack orchestrates energy optimization, disaster response, and multimodal mobility. A fictional composite of city operators—call her Lina, a control room lead—illustrates the day-to-day shift: instead of monitoring dozens of camera walls, Lina consults an AI agent that summarizes live feeds, flags anomalies, and proposes actions that align with policy and safety thresholds.

Building confidence in such systems requires transparent modeling and rigorous validation. NVIDIA’s approach pairs physically accurate simulation with scalable deployment at the edge, so models can be stress-tested under rare events—fog, glare, flooding—before turning on in the field. For deeper context on open-world simulation and foundational modeling for physical AI, see this exploration of synthetic environments and the Omniverse. The industry’s momentum is reinforced by domain-specific breakthroughs—such as AI-accelerated physics for engineering—that are now translating into city-scale planning and operations.

Security and governance are equally crucial. City CIOs are prioritizing data minimization, risk scoring, and red-teaming of AI agents. Techniques like automated failure attribution support root-cause analysis when edge systems behave unexpectedly, while evolving cyber practices—see a brief on AI-powered browser cybersecurity—help reduce the attack surface across thousands of cameras and IoT endpoints. As foundation models evolve, the field also tracks comparative safety research across providers; an overview like OpenAI vs. Anthropic in 2025 helps decision-makers calibrate expectations for generalized assistants that might co-pilot municipal workflows.

What the unified stack enables

In practice, successful deployments hinge on coordination across cloud, edge, and network layers. Vendors including Amazon Web Services, Microsoft, Google, Dell Technologies, and Cisco are aligning compute, storage, and networking with NVIDIA GPUs and SDKs. Operational tech leaders—Siemens for mobility systems, IBM for data governance, Bosch for sensors and security, and Qualcomm for edge AI—are vital to interoperable architectures.

• 🌆 City-scale digital twins simulate construction detours, weather extremes, and crowd flows with Cosmos Predict and Omniverse.
• 🎥 VLM-powered agents summarize live video, reducing operator fatigue and filtering false alarms.
• 🚦 Intelligent traffic controls optimize signal timing, emergency routing, and event logistics.
• 🔒 Policy-aware compute enforces privacy zones and retention windows by design.
• ⚡ Edge-to-cloud orchestration balances latency, cost, and resilience across sites.

With stakeholders aligned and a simulation-train-deploy loop in place, cities can move from pilots to production more swiftly—turning physical AI into an everyday utility for safer, smarter streets.

Geospatial Intelligence in Raleigh: Esri and NVIDIA Metropolis Orchestrate Real-Time Operations

Raleigh, North Carolina, is piloting a new class of geospatial AI agents that ingest live camera and sensor feeds and project them onto an interactive city map. Built with Esri and NVIDIA’s Blueprint, the system translates raw video streams into actionable overlays—congestion levels, queue lengths, incident markers—so operations teams can coordinate interventions across traffic engineering, public works, and emergency services.

The solution taps NVIDIA Metropolis for vision AI and the VSS blueprint for search and summarization, then fuses the results within ArcGIS for location-aware decisions. Cosmos Reason VLMs contextualize anomalies (“blocked lane near school zone”) and propose playbooks (“re-time signals 7–12 on MLK Blvd, dispatch tow”). The city architecture—representative of many U.S. municipalities—blends GPU-accelerated servers from Dell Technologies in a regional hub, Cisco SD-WAN for backhaul, and hybrid workloads on Amazon Web Services, Microsoft Azure, and Google Cloud for scale-out analytics.

Raleigh’s team is particularly focused on reducing operator overload. By having AI pre-triage alarms and summarize trends, staff can spend more time resolving problems and less time searching video. Studies in similar environments indicate that AI-assisted triage can cut false alarms by double-digit percentages. Milestone Systems’ work in this space—featured later—shows how compliant training data paired with VLMs can reduce alarm fatigue by up to 30%.

Pipeline from live video to action

Instead of staring at multiple dashboards, operators consume a single geospatial interface. Under the hood, the flow is modular, auditable, and policy-aware.

• 🛰️ Ingest: Edge cameras and sensors stream to Metropolis pipelines with on-device redaction.
• 🧠 Understand: VLMs classify events, tag objects, and score severity using Cosmos Reason.
• 🗺️ Localize: Events are placed on a live map with lane, curb, and asset context from Esri.
• 📣 Act: Suggested responses appear as playbooks and can auto-trigger signal plans.
• 📚 Learn: Outcomes feed back to simulation for continual improvement.

As model sizes grow and training becomes more accessible—see a primer on affordable model training—cities can fine-tune agents for local rules and context. Some municipalities also watch the broader LLM landscape to evaluate co-pilots for administrative tasks; note the comparative perspective in ChatGPT vs. Claude in 2025, which features trends in reasoning ability and tooling integration relevant to civic operations.

With a single pane of glass, Raleigh is building a pattern others can emulate: connect sensors, constrain the data, reason locally, visualize globally, and always loop learnings back into simulation.

Ho Chi Minh City and Danang: Linker Vision Scales Physical AI with Omniverse

Vietnam’s fast-growing metros—Ho Chi Minh City and Danang—are deploying Linker Vision’s end-to-end implementation of the NVIDIA Blueprint. The program builds on a successful rollout in Kaohsiung City, Taiwan, where vision AI reportedly cut incident response times by up to 80%. In Vietnam, the focus expands to construction oversight, urban mobility, and safety analytics at city scale, with an emphasis on simulation-first development.

At the heart of the effort are simulation-ready 3D digital twins from AVES Reality, fused in NVIDIA Omniverse. This enables teams to generate photorealistic scenarios for work zones, lane closures, and extreme weather—using Cosmos Predict and Cosmos Transfer to synthesize rare but critical edge cases. The Cosmos Reason VLMs then evaluate the scenarios, flagging potential conflicts or hazards. When models meet performance bars in sim, they deploy to the field on ruggedized edge nodes, bringing physical AI closer to the curb where latency matters most.

Program governance is built around transparency and measurable outcomes. City leaders define KPIs—on-time project delivery, reduced congestion near construction, improved pedestrian safety—and align them with agent goals. This is where globally shared learnings become essential. For instance, methods for attributing model failures help operations teams triage drift or sensor anomalies, while developments in self-improving AI systems point toward agents that can boost performance through structured feedback without compromising governance.

From detours to dashboards: a day in the life

Consider the scenario of a new light-rail extension. Digital twins simulate excavation phases, truck routes, and pedestrian diversions across weeks of construction. The AI agent predicts queue formation on feeder roads and tests alternative signal plans in silico. Once live, it monitors compliance, detects early signs of gridlock, and recommends micro-adjustments to keep traffic moving while minimizing safety risk.

• 🚧 Construction monitoring: Cameras and LIDAR ensure work zones adhere to approved plans.
• 🚦 Traffic orchestration: Agents balance throughput and safety minute by minute.
• 🌧️ Weather-aware control: Synthetic data trains robust models for rain and low light.
• 🧭 Wayfinding updates: Digital signs change based on real-time occupancy and flows.
• 🔁 Continuous learning: Field feedback retrains models for local nuance.

Vietnam’s blueprint-aligned program reflects a broader trend: cities want solutions they can simulate, verify, and scale. As global collaborations deepen—highlighted by initiatives like cross-border AI partnerships announced at APEC—urban innovators gain access to shared components, best practices, and reference architectures. The result is faster time to value and higher confidence in mission-critical AI.

Dublin’s Micromobility and Road Safety: Bentley, Cesium, VivaCity, and NVIDIA Jetson

Dublin’s strategy emphasizes people-first streets. Through the Smart Dublin program, the city is combining Bentley Systems and Cesium 3D geospatial platforms with NVIDIA Omniverse for real-time visualization, enabling planners to see how walking, cycling, scooters, and vehicles interact in space and time. AI transportation specialist VivaCity contributes computer vision sensors built on NVIDIA Jetson and Metropolis, which provide accurate multimodal counts and behavioral insights.

One early analysis revealed a counterintuitive pattern: when rainfall was layered onto micromobility data in a Cesium-powered digital twin, cycling volumes remained resilient. Planners can use that insight to justify protected lanes that function year-round. Meanwhile, Bentley’s Blyncsy leverages NVIDIA Cosmos and Metropolis to generate synthetic data for road-condition analytics, helping maintenance teams prioritize resurfacing and hazard mitigation.

Privacy, compliance, and interoperability are foundational. Dublin’s architecture employs on-sensor redaction, edge inference, and secure transport—mechanisms that can be independently audited. Supply chain and integration partners also matter: Cisco networking, Dell Technologies compute, Amazon Web Services cloud services, and data governance patterns informed by IBM help ensure the system scales responsibly. Street hardware from Bosch and edge acceleration from Qualcomm support dense deployments, while Siemens expertise connects insights to signal control and mobility infrastructure.

Micromobility intelligence in practice

By understanding multimodal behavior at the block level, the city can optimize for safety and flow without overbuilding. The digital twin allows scenario testing before concrete is poured or lanes are re-striped, reducing costly rework and minimizing disruption to residents and businesses.

• 🚴 Mode share visibility: High-accuracy counts inform investment in bike and bus-priority corridors.
• 🛣️ Conflict detection: AI flags risky interactions at junctions for targeted redesign.
• 🌬️ Weather resilience: Synthetic data trains models robust to rain, fog, and glare.
• ⚙️ Maintenance planning: Blyncsy analytics guide proactive interventions.
• 🔐 Data minimization: On-device processing limits personally identifiable exposure.

For broader context on omniverse-scale modeling and its implications for city systems, this overview of open-world foundation models shows how synthetic environments accelerate robust policy testing. Complementary infrastructure investments—such as new regional data facilities like the Michigan AI data center initiative—signal how compute availability is expanding for public sector workloads worldwide.

Dublin’s blueprint-driven micromobility approach demonstrates a pragmatic path: synthesize, simulate, and scale, while earning public trust through privacy and performance.

From Control Rooms to the Edge: Milestone, Deloitte, and a Global Mesh of Smart City Partners

Ramping from pilot to production demands repeatable patterns and robust hardware. Across solution showcases, AAEON, Advantech, Aetina, Dell Technologies, HPE, OpenZeka, and YUAN High Technologies are demonstrating physical AI pipelines on NVIDIA RTX PRO Servers, NVIDIA DGX Spark—described as the world’s smallest AI supercomputer—and NVIDIA Jetson Thor modules for energy-efficient edge inference. Integrators align these platforms with municipal standards for reliability, cybersecurity, and lifecycle management.

Milestone Systems is introducing generative AI in its XProtect video management platform, enabling users to extract analyses from video libraries, review alerts, and auto-generate reports. The capability rides on Cosmos Reason VLMs that Milestone post-trained with 75,000 hours of compliant traffic footage per region (EU and U.S.). Early adopters in Dubuque, Iowa, and Genoa, Italy, plan to evaluate how these features cut operator alarm fatigue by up to 30% through automatic review and false-alarm reduction. Milestone will also offer these specialized VLMs-as-a-service, opening a pathway for developers to build domain-specific applications.

Deloitte is applying the Cosmos suite to automate street inspections across thousands of crosswalks. Cosmos Predict converts static scenes into photorealistic, physically accurate videos; Cosmos Transfer augments them with weather and lighting variations; and Cosmos Reason evaluates the outputs to prioritize improvements. This pipeline compresses previously manual processes into repeatable, machine-aided workflows—an approach resonant with the broader industry’s shift to explainable, auditable AI and to operational comparisons like 2025 assistant model evaluations that inform procurement and governance.

Operational resilience remains a top priority. Cities are exploring distributed designs that keep critical services running during network outages. Partnerships with hyperscalers (Amazon Web Services, Microsoft, Google) combine cloud elasticity with on-prem reliability. Meanwhile, research and product updates—surveyed in resources such as a 2025 review of assistant capabilities—inform how co-pilots support human operators without over-automation. Cultural and social considerations matter too; for example, a piece on digital well-being signals underscores the care needed when deploying AI in public spaces.

Hardware-software choreography in real deployments

The path from lab to street often hinges on cohesive integration across vendors and agencies. Experience suggests that running a simulation-first playbook, aligning to open data standards, and defining measurable, citizen-centered outcomes are non-negotiable for success.

• 🧩 Composable architecture: Mix-and-match edge nodes, networks, and agents to fit site constraints.
• 🛡️ Defense-in-depth: Segment networks, encrypt data, and apply policy at every layer.
• 📏 Outcome contracts: Tie SLAs to travel-time savings, safety gains, or emissions cuts.
• 🧪 Pilot with purpose: Use digital twins to test enforcement and equity implications.
• 🌐 Ecosystem cadence: Share playbooks across cities through standards bodies and forums.

Global AI infrastructure is expanding—see trends like the Midwest data center buildouts—and cities stand to benefit as compute becomes more accessible. As training costs fall, indicated by coverage of efficient model training, municipalities can customize agents for local policy, language, and infrastructure. The practical takeaway: anchor innovation in simulation and governance, and deploy where latency, privacy, and resilience requirements demand it—often, at the edge.

Showcase Cities, Shared Patterns: Dublin, Ho Chi Minh City, Raleigh, and the Road Ahead

Across featured cities, a consistent pattern emerges: simulate first, deploy incrementally, and measure relentlessly. While local needs differ—micromobility in Dublin, construction-aware traffic in Ho Chi Minh City, geospatial command-and-control in Raleigh—the NVIDIA Blueprint provides a common backbone. When paired with strong partners and an outcomes mindset, cities can rapidly adapt to surges in tourism, new transit lines, or climate-induced disruptions.

Success depends not only on algorithms but also on interagency coordination and public communication. Control-room operators gain co-pilots that summarize and prioritize; planners get sandboxes to test redesigns; and residents receive quicker responses with fewer disruptions. Transparency remains essential: cities should publish metrics, explain how data is handled, and invite feedback. In doing so, they build legitimacy alongside capability.

Operational technology vendors and cloud providers all play roles. Siemens ties insights to traffic signals, Bosch contributes robust sensors, Qualcomm powers low-energy edge AI, and IBM supports governance and lifecycle controls. At the infrastructure layer, Amazon Web Services, Microsoft, and Google scale the data plane and model training, while Dell Technologies and Cisco anchor compute and networks on-prem. The net effect is a resilient mesh that can sustain public services even under duress.

Field-tested patterns that travel well

The following repeatable practices have surfaced across pilots and scale-ups, and they map directly to the urban challenges of the next decade.

• 📊 Metric-first design: Start with congestion minutes saved, crash-rate deltas, and service-level targets.
• 🧪 Equity testing: Use twins to check that signal plans don’t disadvantage vulnerable users.
• 🏗️ Construction-aware timing: Align lane closures, detours, and transit signal priority in sim.
• 🔐 Privacy by default: Use on-sensor redaction and minimize retention windows.
• 🔁 Closed-loop learning: Continuously fine-tune VLMs with local feedback and governed datasets.

As cities iterate, it helps to keep a pulse on broader AI themes that can influence policy or procurement. For example, capability shifts in generative agents—captured in analyses such as a 2025 assistant review—may change expectations for human-in-the-loop operations. And while out of scope for public deployments, trend pieces like emerging AI content categories occasionally foreshadow general advances in model alignment and filtering that later benefit civic systems. The common thread is disciplined adoption: simulate thoroughly, validate ethically, and deploy where the public good is clearest.

These showcase cities point to a scalable future: a single, iterative loop—simulate, train, deploy—connecting design studios, control rooms, and curbsides. With that loop in place, urban innovation becomes an ongoing service rather than a sporadic project.

Engagement at SCEWC: Platforms, Playbooks, and Where to See Physical AI Live

At the Smart City Expo World Congress, the NVIDIA partner ecosystem is demonstrating how blueprints translate into street-level outcomes. Many demos run on NVIDIA RTX PRO Servers for centralized workloads, NVIDIA DGX Spark for compact high-performance training, and NVIDIA Jetson Thor for efficient edge inference. Hardware partners—AAEON, Advantech, Aetina, Dell Technologies, HPE, OpenZeka, YUAN High Technologies—show turnkey kits that cities can prototype within weeks.

Visitors can explore Akila’s physical AI in Monaco-Monte-Carlo train stations and its digital twin deployment at the University Mohammed VI Polytechnic in Morocco, illustrating the power of unified operations data models. K2K is demonstrating real-time roadway safety analytics using Cosmos Reason and the VSS blueprint, with additional use cases in waste management. These live systems help cities imagine how a curated set of agents—traffic, safety, maintenance—can work in concert on shared infrastructure.

Beyond the expo floor, policy and procurement leaders are discussing vendor-neutral procurement, open data standards, and shared governance frameworks. For a broader lens on the competitive and collaborative landscape shaping AI in 2025, resources such as market comparisons of assistant providers and commentary like cloud-native, latency-sensitive apps can help illustrate why edge-heavy architectures are gaining favor among cities. Resilience considerations also extend to workforce tools and communications; even guides such as understanding queued email behavior can matter during incident response where reliable messaging is critical.

What to look for on the show floor

Practitioners evaluating demos can apply a few simple criteria: does the system operate within defined privacy bounds, does it reduce time-to-action, and can it be simulated and audited end to end?

• 🧪 Simulation parity: Demos that mirror field conditions with measurable fidelity earn trust.
• ⏱️ Latency transparency: Clear edge vs. cloud paths enable predictable response times.
• 🔍 Explainability: Human-readable summaries and logs support oversight.
• 🧰 Interoperability: APIs align with open standards to avoid lock-in.
• 📈 KPI dashboards: Real metrics—not vanity—prove readiness for scale.

For adjacent perspectives on how AI ecosystems evolve, see this look at conversation-driven product features, which mirrors how interface design can make complex AI accessible. As cities scale physical AI, the lesson from consumer AI applies: small UX choices have outsized operational consequences.

How does the NVIDIA Blueprint reduce risk for city deployments?

By pairing high-fidelity digital twins with Cosmos-based synthetic data and VLM reasoning, cities can test policies and edge cases before field rollout. The simulation-train-deploy loop creates auditable evidence that systems will behave within safety and privacy bounds.

Which partners are critical for end-to-end smart city solutions?

Beyond NVIDIA’s Omniverse, Cosmos, Metropolis, and VSS, cities rely on cloud providers (Amazon Web Services, Microsoft, Google), infrastructure vendors (Dell Technologies, Cisco), and domain specialists (Siemens, IBM, Bosch, Qualcomm, Esri, Bentley, VivaCity, Milestone, Deloitte) to ensure interoperability, governance, and performance.

What outcomes can cities expect in the first year?

Typical targets include reduced incident response times, double-digit false-alarm reductions, measurable travel-time savings on key corridors, and faster inspection cycles. Programs that start with simulation and metric-first design show the fastest time to value.

How are privacy and compliance handled?

Architectures emphasize on-device redaction, data minimization, and policy-aware compute. Governance frameworks—supported by vendors like IBM—help enforce retention windows, access controls, and transparent auditing.

Where can the public see live demonstrations?

At Smart City Expo World Congress, partners showcase deployments running on NVIDIA RTX PRO Servers, DGX Spark, and Jetson Thor. Additional case studies highlight projects in Dublin, Ho Chi Minh City, Danang, Raleigh, Monaco-Monte-Carlo, and Morocco’s UM6P campus.

Rachel Tanaka

Rachel has spent the last decade analyzing LLMs and generative AI. She writes with surgical precision and a deep technical foundation, yet never loses sight of the bigger picture: how AI is reshaping human creativity, business, and ethics.