AI Data Centers, 2030–2035: Why Inference Will Reshape the Power Map
Blog Post | 09.06.2026 | 6 min read | Susan C. Mcleod
The next phase of AI infrastructure will not be defined by a single data-center model. It will be defined by a split architecture: massive AI training campuses on one end, and increasingly distributed inference capacity closer to users, enterprises, and major metropolitan markets on the other. From Hitachi Energy’s perspective, this is not only a compute story – it is a power infrastructure story, because the shift from centralized training to distributed inference changes how energy must be planned, delivered, and managed across the AI value chain.
From 2024 through 2030, much of the public discussion has centered on mega-sites: very large data-center campuses designed to support model training, high-density GPU clusters, advanced cooling, and enormous power requirements. That focus is understandable.
Frontier model training is capital-intensive and strategically important. But from 2030 to 2035, the more interesting shift may be the expansion of inference infrastructure – capacity designed not simply to train models, but to run AI continuously, in real time, closer to where people and businesses use it.
In practical terms, AI infrastructure will begin to look less like isolated mega-campuses and more like a layered network: mega-sites for training, regional hubs for high-volume inference, and metro-edge facilities in metropolitan (metro) regions, where latency, enterprise demand, fiber density, and power availability intersect.
The strategic shift is not that mega-sites disappear. It is that inference becomes the recurring, distributed workload that shapes where the next wave of capacity is built.
By 2030, Inference Becomes the Dominant AI Workload
McKinsey provides a useful anchor point. By 2030, it is projected that inference will surpass training as the dominant workload in AI data centers, representing more than half of all AI compute and roughly 30% to 40% of total data-center demand. Its forecast shows global data-center demand rising from 82.3 GW in 2025 to 219.0 GW by 2030, with AI inference growing from 20.9 GW to 93.3 GW and AI training growing from 23.1 GW to 62.2 GW.
The distinction matters:
- Training and inference have very different infrastructure profiles. Training can be more tolerant of distance from end users, but it requires enormous power density, specialized hardware, high-performance networking, and advanced cooling.
- Inference is different. It is driven by everyday usage: enterprise copilots, AI agents, customer service automation, fraud detection, search, personalization, software development, healthcare workflows, and consumer applications.
As AI moves from experimentation to embedded workflow, inference becomes the recurring workload. Training is episodic. Inference is persistent.
How AI infrastructure Expands, 2030 to 2035
A practical planning view is that from 2030 to 2035, incremental AI data-center expansion will likely split into four broad categories. This planning model is based on analyst direction, workload behavior, power constraints, and the operational differences between training and inference.
| Segment | Typical scale | Location logic |
|---|---|---|
| Mega training / frontier AI campuses | ~200 MW to multi-GW campuses | Power-first locations, large land parcels, grid interconnects, captive energy, and lower latency sensitivity |
| Large regional inference hubs | ~20-100 MW | Near major metros, strong fiber, cloud regions, and enterprise clusters |
| Edge inference node | <10 MW | Carrier hotels, telco central offices, colocation sites, and enterprise latency zones |
In short, the difference is one of scale, location, and workload: mega campuses are power-first environments built for frontier training, regional hubs handle the largest share of recurring inference demand near major cloud and enterprise concentrations, and metro-edge deployments extend AI delivery closer to users where latency and local access matter most.
Why Metro Markets Matter for Inference
Metro markets matter because they concentrate population, enterprise demand, network density, and regional economic activity. But they will not all play the same role in AI infrastructure.
The largest digital infrastructure markets—Northern Virginia, Dallas, Atlanta, Chicago, Phoenix, the Bay Area, Los Angeles, New York/New Jersey, and Seattle—will likely remain the primary AI hubs. But a second tier of metros, including Charlotte, Nashville, Indianapolis, Kansas City, Minneapolis, Detroit, Denver, Las Vegas, Miami, Tampa, Cleveland, Pittsburgh, and Cincinnati, is increasingly well positioned for inference growth.
The implication is clear: inference capacity will scale across a broader set of metro markets closer to users and applications, while power demand becomes more geographically distributed. That increases the importance of modular capacity, substation readiness, grid access, and faster energization across many more locations.
Power will Remain the Gating Factor
The limiting factor for AI infrastructure will continue to be “power”. The International Energy Agency projects a major increase in data-center electricity demand as AI adoption grows. Its Energy and AI analysis includes a wide range of 2035 outcomes, with global data-center electricity demand across scenarios spanning roughly 700 TWh to 1,700 TWh. That range shows that the 2035 market will be shaped not only by AI adoption, but also by efficiency gains, grid constraints, supply chains, and the ability to site capacity where power is available.
Deloitte's U.S. outlook makes the scale of the challenge more concrete. Deloitte estimates that power demand from AI data centers in the United States could grow more than thirtyfold by 2035, reaching 123 GW, up from 4 GW in 2024. That scale explains why the mega-site story remains important. It also explains why inference will decentralize if interconnection queues, land constraints, and local opposition slow development in the largest hubs.
AI Infrastructure will Ultimately be a Hybrid, Not Either/Or
There are two competing narratives in the market today. One view says AI infrastructure will keep centralizing into ever-larger mega-campuses because scale wins. Large sites can optimize power procurement, cooling, security, chip deployment, and operations. For training workloads, this view is compelling.
The other view says inference will move closer to the edge because user experience, latency, data gravity, and network cost will matter more as AI becomes embedded in everyday workflows. This view is also compelling.
The most likely answer is both. Training will remain concentrated in power-first locations. High-volume inference will sit in regional hubs. Latency-sensitive inference will move into metro-edge facilities. Specialized workloads will run on-premises or at the far edge.
Power Readiness Becomes the Differentiator
As AI infrastructure shifts from concentrated training campuses to a hybrid network of mega-sites, regional hubs, and metro inference deployments, power strategy becomes the critical differentiator. Customers will need to plan for two realities at once: very large, power-dense environments for frontier training, and a broader set of distributed deployments that must be energized faster, repeated across locations, and scaled with consistency.
That raises the importance of grid access, substation readiness, transformer capacity, power quality, and modular deployment models that improve speed, resilience, and operational control.
The strategic implication is clear: competitive advantage will increasingly go to those who can align compute growth with repeatable, energizable power architectures. In that transition, Hitachi Energy is well positioned to help customers bridge centralized and distributed AI growth through modular substations, standardized transformer configurations, and scalable power-quality solutions that support faster, more reliable expansion over time.
References
