The infrastructure built to train artificial intelligence has rewarded scale, concentrating enormous amounts of compute in a relatively small number of power-rich campuses. As AI moves from training models to serving billions of everyday requests, inference is beginning to create a very different geography for the data center industry.
Digital Realty’s latest expansion in Singapore offers a glimpse of how that geography may develop. The operator was selected in August to build another 50MW of capacity on Jurong Island, explicitly positioning the facility for AI inference alongside high-performance computing and enterprise workloads. Its reasoning is significant: inference increasingly needs to sit closer to users and the enterprise data on which applications depend, making connectivity, resilience and regional presence as important as raw computing power.
For the first phase of generative AI, the infrastructure logic was comparatively straightforward. Training ever-larger models required enormous clusters of accelerators, plentiful electricity and highly specialised cooling, encouraging development at a relatively small number of locations where hundreds of megawatts could be secured. Inference changes that calculation because the model is no longer learning; it is responding. As AI becomes embedded in financial transactions, industrial systems, healthcare, customer services and eventually autonomous applications, the distance between the compute and the user begins to matter much more.
JLL expects the transition to become visible surprisingly quickly. AI represented roughly a quarter of data center workloads in 2025, with training accounting for the majority, but it forecasts inference overtaking training as the dominant AI requirement in 2027. By 2030, AI could represent half of all data center workloads, with inference accounting for 37 per cent of the total compared with 13 per cent for training.
That does not signal the demise of the hyperscale campus. It suggests something more consequential: AI infrastructure is likely to develop additional layers.
Training and inference want different things
Training rewards concentration because thousands of accelerators need to behave almost like a single computer. The economics favour enormous facilities where operators can assemble dense clusters, deliver large quantities of power and install the networking and liquid cooling required to keep increasingly powerful hardware operating efficiently. JLL puts typical training densities at 40kW to more than 100kW per rack, levels that restrict the number of locations capable of supporting large-scale deployments.
Inference is much less uniform. Some workloads can remain inside those large campuses, particularly where response times are relatively forgiving or economies of scale outweigh latency. Others become more valuable when the computation happens closer to the customer, device, factory or application generating the request. A fraud-detection system examining a financial transaction, an industrial AI interpreting machinery data or an autonomous system making an operational decision may have very different infrastructure requirements from a consumer chatbot.
This is why the simple prediction that AI will “move to the edge” is misleading. Centralised infrastructure will remain essential, but it will increasingly form one part of a wider hierarchy. JLL expects early inference to remain heavily concentrated within existing hyperscale cloud regions before spreading towards regional hubs and, eventually, more embedded edge environments as applications become more latency-sensitive. Its forecast envisages most of the global inference initially operating within fewer than 50 major clusters even as the longer-term direction becomes considerably more distributed.
Arup reaches a similar conclusion from a different perspective. Its 2026 Data Centre Futures research describes a “ladder-shaped” network in which local edge nodes handle real-time activity, regional hubs aggregate workloads and large core sites continue performing computationally intensive tasks such as training and large-scale storage. The strategic unit therefore becomes the portfolio rather than the individual facility.
For operators, that distinction matters. The industry has spent years pursuing scale, but the next competitive advantage may come from having the right combination of scale and location.
Connectivity becomes part of the compute architecture
A distributed model only works if its component facilities can communicate quickly enough to behave as part of the same infrastructure. That puts fibre in a more strategic position than the familiar description of connectivity as another requirement alongside power and land might suggest.
The Fiber Broadband Association argues that congestion, power availability, water pressure and rising property costs in established US hubs are already encouraging developers to consider secondary markets, edge locations and rural regions. Yet those areas become viable only when reliable power is accompanied by high-capacity fibre, because moving compute away from established campuses increases rather than reduces the need to move enormous quantities of data between facilities.
That requirement becomes more demanding as AI scales. Data must move between storage, training environments, regional inference clusters, enterprise systems and eventually edge nodes, while many applications will be intolerant of unpredictable latency. The FBA estimates that the US could need to expand data-center-related fibre route mileage from around 95,000 miles to 187,000 by 2029, while highlighting the importance of diverse routes and middle-mile capacity in determining whether emerging locations can genuinely support data-intensive development.
This makes the network much more than the pipe connecting the building to the internet. In an increasingly distributed AI architecture, connectivity becomes part of the computing system itself. A regional data center with abundant electricity but poor fibre diversity may be less useful for inference than a more constrained metropolitan facility positioned inside a dense network ecosystem.
That helps explain why established hubs are unlikely simply to be replaced by cheaper locations with more power. Their concentration of carriers, exchanges, cloud platforms, enterprise customers and existing fibre remains extremely valuable, particularly for workloads dependent on rapid interaction with multiple systems.
Digital Realty’s Singapore expansion illustrates that dynamic. The attraction is not simply another 50MW of capacity but Singapore’s role as a highly connected regional hub from which workloads can reach customers and enterprise data across Asia. Digital Realty is pursuing a similar strategy in Malaysia, where its expanding Cyberjaya platform connects into Singapore and Jakarta and is being positioned for increasingly distributed and latency-sensitive AI workloads.
Geography becomes workload dependent
The implications extend beyond deciding whether to build in Northern Virginia, Frankfurt or a new secondary market. Different types of AI may increasingly create their own location logic.
Training can migrate towards areas where power is abundant and relatively inexpensive because the model does not need to be physically close to the people who will eventually use it. Inference introduces additional considerations around latency, data sovereignty, connectivity and the location of enterprise information. Some workloads may reside in large regional facilities, others in metropolitan colocation sites and a smaller proportion much closer to factories, hospitals, offices or telecommunications networks.
Arup argues that always-on inference will intensify this outward movement as AI evolves from systems that primarily generate content towards those that continuously observe, decide and act. Data generated by sensors, autonomous systems and industrial equipment increases the value of processing information closer to where decisions need to happen, while training continues to depend on large central clusters.
Sovereignty adds another dimension. As governments and regulated enterprises place tighter controls on where information can be processed and stored, the most computationally efficient location may not always be legally or strategically acceptable. JLL expects domestic AI infrastructure to expand partly because local processing requirements restrict where inference can take place, further fragmenting what was once a predominantly global cloud model.
Operators will consequently need to think less about building a collection of interchangeable facilities and more about the role each site performs within a wider infrastructure network. Large campuses remain essential for concentrated compute, while regional facilities gain importance for inference and data aggregation, and metropolitan or edge capacity addresses workloads where milliseconds, sovereignty or local data access justify placing infrastructure closer to demand.
The result will not be the wholesale decentralisation of the data center industry. Scale remains too economically powerful, and much inference will continue to run inside enormous cloud environments. What changes is the assumption that the largest possible campus in the cheapest available power market represents the answer to every AI workload.
As inference becomes the dominant way in which most organisations and consumers experience artificial intelligence, the infrastructure supporting it will need to follow the activity it serves. Power will continue to dictate where the largest clusters can be built, but networks, latency, sovereignty and proximity to data will increasingly determine where the next layer of capacity belongs.
The AI infrastructure map is therefore unlikely to become less concentrated so much as more complicated. The hyperscale campus will remain at its centre, but around it will develop regional hubs, interconnected metropolitan facilities and specialised edge environments, each performing a different part of the workload. For data center operators, the next phase of AI may be defined not simply by how much capacity they can build, but by whether they have built it in the right places.
