How AI Infrastructure Is Reshaping the Future City
- Jun 11
- 11 min read
What did the city of the future look like in your childhood imagination?
Maybe it was robots collecting trash in apartment complexes, self-driving vehicles gliding through traffic-free streets, or an AI assistant like Iron Man's J.A.R.V.I.S. handling your every need. In WALL-E, humans drift through life on floating chairs inside a giant spaceship, living almost entirely inside a virtual world.
These visions differ wildly in their details, but they share one constant: AI is woven into everything. How you move, how you communicate, how you work — AI is the invisible layer underneath it all. In that sense, the future city is less a place than a massive operating system, running countless AI services simultaneously. That idea is now driving a new wave of urban concepts: the AI City, the Agentic City.
But here is a question most of these visions skip over. Once that future city is actually running — where does all the AI compute happen, and who pays for it?
"Can we actually keep these services running sustainably at city scale?"
AI requires compute. Compute costs money. In an agentic era where millions of AI calls fire every day across a single city, the infrastructure bill is enormous. The best AI services in the world cannot scale to an entire city if the operating cost is too high to sustain. The future city, it turns out, will be determined less by the intelligence of its AI models than by the infrastructure that runs them. So what does that infrastructure actually need to look like?
1. The Center of Gravity in AI Infrastructure Is Shifting
The way AI infrastructure is designed has gone through three distinct phases since ChatGPT's emergence in 2022.
The first was the Training Era. The priority was building large language models by feeding them massive datasets. This required dense clusters of high-end GPUs running for weeks at a time. Centralized hyperscale data centers — purpose-built to concentrate thousands of GPUs in one place — were the obvious fit.
The second phase was the Inference Era. The emphasis shifted from building models to running them in production at scale. As user counts grew, inference calls accumulated continuously, and operating costs grew with them. The chart above captures the inflection point: around 2025, inference compute overtook training compute for the first time. Getting infrastructure right for production serving — not model training — became the central engineering problem.
Now we are in the third phase: the Agentic Era. Where a traditional chatbot handles one question with one response, an AI agent handles one request by searching for information, calling multiple tools, and verifying results across many steps. A single user request can trigger dozens or hundreds of downstream inference calls. Inference demand does not just grow — it compounds. The problem is that the centralized data center model was not designed for this. Hyperscale facilities concentrate high-end GPUs (H100s, H200s) in single locations that require enormous upfront capital, dedicated power infrastructure, and industrial-scale cooling systems. Every dollar spent building and running that infrastructure ultimately flows through to the user's bill.
There is a further dimension ahead. AI is expected to expand into the physical world — sensors embedded across cities, autonomous vehicles, robots, smart devices — what researchers call Physical AI. In that environment, the bottleneck is not just raw compute volume. It is latency. Data generated at the edge of a city cannot always wait to travel to a distant data center and back. Processing must happen faster, in more places, with lower round-trip delay. Concentrating everything in a single facility becomes structurally harder to justify as that requirement intensifies.
2. The Bottleneck Inside Today's Smart Cities
This is not a future problem. It is already playing out in the cities that have moved furthest toward the AI City vision: smart cities.
Smart cities use ICT and AI to address urban challenges in real time — traffic flow, energy management, environmental monitoring, public safety, and civic administration. In a sense, the smart city is the earliest version of the AI City concept in production.
But most smart cities today carry a structural constraint. Only a fraction of urban data has been migrated to the cloud. The rest lives in isolated silos — traffic data in the traffic management system, environmental data in the environmental monitoring system, public safety data in a separate operations center. Because data is fragmented across organizations and departments, analyzing the city as a unified intelligent system in real time remains structurally difficult. This is a key reason edge computing has not spread through large urban environments as quickly as expected.
The data processing model creates a second problem. Most smart city services today follow a collect-and-centralize architecture: data flows from edge sensors to a central data center, which processes it and sends instructions back. As cities grow, sensor-generated data from traffic, environment, and safety systems grows exponentially. Routing everything to a central point creates network congestion and increases latency. Add AI agents and real-time analytics services on top of that, and the inference load on the central facility scales rapidly alongside operational costs. Smart cities face two compounding challenges: the data silo problem and the inference demand problem.
The 2024 Hyper-Connected Intelligent City Research Report published by Korea's Korea Agency for Infrastructure Technology Advancement (KAIA) identifies multi-distributed cloud as the next-generation core technology for addressing these problems. Rather than routing all data through a single central facility, a multi-distributed cloud connects computing resources distributed across the city into one unified cloud-like layer. Traffic data can be processed close to where traffic is generated; environmental data can be analyzed near its sensors. The result is lower network congestion, better real-time performance, higher scalability, and the ability to utilize idle compute resources at the local level. In short: treating the entire city like one giant distributed computer. This architecture makes city-scale edge computing genuinely achievable — faster processing of real-time traffic and environmental data, more efficient use of urban IT infrastructure, and the flexibility to absorb growing inference demand as AI services multiply across the city. The city cannot become a unified intelligent system if only the AI gets smarter. The infrastructure itself must be distributed.
3. The World Has Already Started
This is not theoretical. The experiments are already running.
In the UK, Heata installs compact servers next to residential hot water tanks. As the server runs, the heat it generates warms the water. The homeowner gets lower heating bills; Heata sells the compute capacity to enterprise customers. Heata is currently running a 100-home pilot in partnership with British Gas and has been recognized as an energy efficiency improvement technology eligible for home energy performance certification. By converting server waste heat into usable domestic energy, it functions as an environmentally friendly infrastructure model as well.
In the United States, smart home energy company SPAN is working with NVIDIA and homebuilder PulteGroup on the XFRA project. Liquid-cooled GPU cabinets — each containing 16 NVIDIA RTX Pro 6000 Blackwell GPUs and 4 CPUs — are mounted on the exterior walls of newly built homes. Homeowners provide wall space and electrical headroom in exchange for a discounted electricity rate. The key insight is that the average home uses only about 40% of its available electrical capacity. SPAN's smart panels identify that unused headroom and direct it toward AI compute workloads. The fact that the world's largest semiconductor company is now designing residential homes as first-class AI infrastructure nodes signals that this direction has moved well beyond experiment.
These cases share a clear pattern. The home is becoming a basic unit of AI computing infrastructure. Existing spaces and existing resources are being converted into compute capacity, and the value generated is shared with the space provider. AI infrastructure no longer lives exclusively inside massive data centers. It is expanding into homes, buildings, and local communities across cities.
At AIEEV, we believe cases like these will multiply. But for the Heata and SPAN models to scale to the city level, one core problem must be solved: how to unify and operate compute resources that exist in many different locations as a single coherent cloud.
That is the problem AIEEV has been building toward. Air Cloud is a distributed AI cloud service built on technology that connects GPUs, NPUs, and other compute resources across heterogeneous environments into a single unified cloud. Our approach is different from a simple GPU marketplace. Rather than just connecting disparate resources, we focus on managing and operating distributed compute so that real AI services can run on top of it reliably.
Air Cloud can connect personal PCs, GPU server farms, enterprise servers, and regional data centers. We also operate our own AI-optimized compute node (the white device pictured above), designed to be installed wherever power and network connectivity exist — homes, apartment complexes, office buildings, industrial parks. The goal is not to concentrate AI infrastructure in any one place, but to make it expandable into whatever environments need it. Where Heata uses a home's hot water system and SPAN uses residential electrical capacity, we are building the underlying technology that connects and operates these distributed resources as a unified cloud. The natural question that follows: can this distributed AI infrastructure model actually work at the city level? As it turns out, one country is particularly well-positioned to find out.
4. Why Korea Is an Ideal Testing Ground
South Korea's National AI Strategy Committee this year outlined a 'K-AI City' roadmap, calling for the transformation of existing smart cities into what they describe as "hyper-connected cities operated by AI." The Ministry of Land, Infrastructure and Transport has designated pilot AI cities on a regional basis and is actively expanding city-scale AI deployment models. The notable shift in policy direction: away from centrally planned infrastructure build-outs toward regional-level AI services and infrastructure that interconnect. This aligns with the distributed cloud philosophy described above.
Distributed cloud infrastructure carries more than technical significance here.
Large-scale data centers carry real costs for the communities around them: massive power consumption, cooling infrastructure, and concentrated strain on local utilities. In South Korea, this has produced a pattern of recurring conflicts. Earlier this year in Geumcheon-gu, Seoul — a densely populated urban district — over 100 residents gathered to protest data center construction. 2,675 people signed a petition opposing the project. Their concerns centered on heat island effects from waste heat, groundwater depletion from cooling systems, and fire risks from large battery installations. The local government issued a construction suspension order. Similar conflicts have repeated in multiple locations across the Seoul metropolitan area.
The underlying cause is structural. South Korea is a small, densely populated country where the land area required by a hyperscale data center — and the power infrastructure to support it — is increasingly hard to accommodate, especially in urban cores. Distributed AI infrastructure approaches this problem differently. Instead of a single massive facility, small compute nodes are distributed across the city. Thermal load and power consumption are spread rather than concentrated. The city's existing idle capacity becomes usable rather than wasted.
This is where a distinctly Korean advantage emerges. South Korea has one of the highest apartment residency rates in the world — the majority of the population lives in apartment complexes. High-density residential clusters with shared infrastructure are structurally well-suited for deploying and operating distributed compute nodes. The trend is reinforcing itself: major construction companies are now actively investing in AI-integrated apartments, smart home systems, and complex-wide digital platforms. The network and digital infrastructure required to host compute nodes is already being built into new residential construction as a baseline feature. South Korea, in short, offers a compelling environment for testing the urban distributed AI infrastructure model that smart city research points toward.
5. From Residential Complex to City, From City to Network
What would this actually look like in practice?
The most realistic starting point is the spaces where people already live and work. Residential complexes, office buildings, and industrial parks host small AI nodes, which connect at the district level to form a distributed computing network. At this stage, the complex stops being just a living space and becomes a micro data center — a node that produces compute capacity the way a solar panel produces electricity. An apartment complex might host a Home Node, an office building an Office Node, an industrial park an Industrial Node. These nodes handle AI inference for local services during normal operation, and connect into the broader city network when shared capacity is needed.
District-Level AI Infrastructure
As these complexes interconnect, the city gains a new class of infrastructure. Traffic, environmental, safety, and energy data can be processed locally rather than routed to a central facility. A specific intersection's traffic analysis runs on the edge node in that district. Anomaly detection from surveillance feeds is handled by the nearest available compute. This is precisely the direction KAIA's multi-distributed cloud framework points toward: process data near where it is generated, reduce network cost, and improve real-time performance. The city is no longer just a space that generates data. It becomes a space that processes it.
The City as AI Infrastructure Producer
One step further: the city itself becomes a producer of AI infrastructure. Today, cities are primarily consumers of AI services — they pay cloud providers for compute. In a distributed computing environment, a city's idle power capacity, unused servers, and underutilized GPUs can supply compute to the network. One district may have surplus power during daylight hours; another may have idle compute overnight. Connected as a network, these resources allow the city to simultaneously consume AI services and supply AI infrastructure to others. This mirrors the concept of the energy prosumer that is already reshaping electricity markets — households that once only consumed power now generate and sell it via solar panels. The future city may hold an analogous position in compute markets.
City-to-City AI Networks
At the broadest scale, cities can share compute across geographic boundaries. If a large event in one city temporarily spikes AI inference demand, idle resources from a neighboring region can absorb the overflow. Industrial zones with surplus compute can supply capacity to cities with high demand and earn revenue in return. This is not just cloud usage — it is a potential model for inter-city AI infrastructure cooperation. To be clear, none of this is fully operational today. But when you place K-AI City's public AI infrastructure ambitions next to the multi-distributed cloud framework KAIA has outlined, this scenario is worth serious consideration. The smart city of the future may treat its computing network the way it treats roads and power grids — as public infrastructure the city itself owns and operates.
Closing
We tend to measure AI-era competitiveness in terms of bigger models, more GPUs, and larger data centers. But as AI expands across entire cities, a different question may matter more:
"How efficiently can we connect and use the resources that already exist inside the city?"
Heata found the answer in residential hot water systems. SPAN found it in unused residential electrical capacity. Smart city research is pointing toward multi-distributed cloud as the architectural answer at urban scale. South Korea's constraints — limited land, high population density — are not obstacles to this model. They may be exactly the conditions that make the distributed infrastructure experiment not just viable but necessary. Ultimately, the competitive advantage of the future city is less likely to come from how much AI it owns than from how sustainably it can keep that AI running.
At AIEEV, we are thinking through how to connect compute resources distributed across cities and regions, and how to make those resources usable as real AI infrastructure. We will continue sharing our thinking on sustainable AI infrastructure and the possibilities of distributed computing.
References
Hyper-Connected Intelligent City Research Report (KAIA, 2024)
Electronic Times, "AI-operated hyper-connected city... National AI Committee discusses K-AI City roadmap" (May 15, 2026)
Smart City Korea Official Portal (smartcity.go.kr) / AI City Task Force launch (Sep 5, 2025)
Busan Eco Delta Smart City official site / K-water KHARN interview (2024)
Heata official site (heata.co) / Data Center Dynamics, "British Gas teams up with Heata"
BlockMedia, "NVIDIA pursues home AI server installation with Span and PulteGroup" (blockmedia.co.kr/archives/1100855)
SPAN official blog, "SPAN Announces XFRA" (span.io/blog)
