For years, the data center industry expected compute to decentralize.
This was the edge thesis. More distributed infrastructure. Smaller regional deployments. Less concentration in a handful of major hubs.
It didn’t happen. The market consolidated harder in Northern Virginia, Atlanta, Phoenix, Dallas, and Silicon Valley. Cloud platforms scaled faster than anticipated, network costs remained low, and centralized infrastructure was easier to operate. No workload showed up large enough to justify the complexity of going distributed.
AI inference may finally be that workload. However, the full picture is more complicated than this.
It Is Not Just AI Driving the Deficit
Goldman Sachs estimates global data centers currently consume approximately 55 gigawatts of power. AI accounted for 14% of that. Cloud computing workloads accounted for 54% of the total. Traditional enterprise IT, such as email, databases, storage, and line-of-business applications, accounted for the remaining 32%.
AI is growing faster than the other two categories, but cloud and conventional enterprise workloads already represent most demand. They are not going away. They are growing alongside AI, which makes the supply picture so difficult.
Goldman Sachs projects total data center power demand will reach 84 gigawatts by 2027, an increase of more than 50% from current levels, and will rise as much as 175% by 2030 from recent baseline levels. Newmark’s 2025 U.S. Data Center Market Outlook sharpened the supply side: projected power demands from all existing and planned U.S. data centers exceed what utilities are set to supply by approximately 50%. This gap is not a temporary condition. It is structural.
The Shift from Training to Inference
Most AI infrastructure conversations over the last two years have centered on training clusters. Bigger GPU deployments. More density. More power.
Training still matters. But the operational reality is shifting toward what happens after models move into production.
Inference is an operating expense that never turns off. Enterprise copilots, AI chat, generative AI, and customer support systems: once adopted across departments, they run continuously. AI inference is projected to grow at a 35% compound annual rate through 2030, reaching more than 40% of total global data center demand, according to McKinsey & Company. Deloitte estimates inference will account for two-thirds of all AI compute in 2026, roughly double its share just a few years prior. Inference-focused cloud infrastructure spending is estimated to reach $20.6 billion in 2026, surpassing training spend for the first time, according to Deloitte.
Agentic AI Multiplies the Problem
The shift to agentic AI, systems that plan, execute multi-step workflows, call external tools and spawn parallel subagents autonomously, introduces a compute demand multiplier that the market has not fully priced in.
A single agentic workflow can consume up to 1,000 times more compute tokens than a simple chatbot query, according to reporting by Quartz. This compounds across every department and enterprise that operationalizes agents at scale. It also reshapes the hardware mix: unlike inference, which leans on GPUs, agentic AI places significant new demand on CPUs for orchestration and workflow coordination. Morgan Stanley estimates agentic AI could add $32.5 billion to $60 billion in new data center CPU demand by 2030. TrendForce projects CPU-to-GPU ratios will shift from 1:8 toward 1:1 as agentic deployments scale.
The infrastructure deficit is not solely a GPU story. It is a full-stack compute story spanning GPUs, CPUs, memory, networking, power and physical space, all competing for the same constrained supply.
Power Became the Constraint
Site selection conversations used to prioritize network density, carrier ecosystems and tax incentives. Power availability now sits at the top of that list, not as a checkbox but as the gating factor.
Northern Virginia remains the largest data center market in the world, but the utility situation has deteriorated substantially. Dominion Energy confirmed that projects over 100 megawatts should expect power delivery timelines of four to seven years, with some sites potentially waiting 17 years, according to Data Center Dynamics. Multi-year interconnection waits have become the norm across U.S. grid regions, according to Lawrence Berkeley National Laboratory data. McKinsey projects U.S. data center power demand will grow from roughly 30 gigawatts in 2025 to more than 90 gigawatts by 2030, exceeding the entire current power consumption of California. The buildout required to meet this demand is not occurring evenly.
The density picture adds another layer. Nvidia’s GB200 NVL72, now deployed in production environments, draws 120 kilowatts per rack under its nominal specification and has been reported to draw 130 to 132 kilowatts at full load, according to HPE product specifications. That is roughly 10 times the draw of a conventional enterprise rack. Air cooling is not an option. Markets where power is available, affordable and deliverable on a realistic timeline are now genuinely scarce.
Hyperscalers Lift All Boats
When a hyperscaler commits to a region, the infrastructure benefits extend well beyond its footprint.
Hyperscalers signing the White House AI data center pledge agreed to cover the cost of all power delivery infrastructure upgrades required for their facilities, including transmission and distribution improvements that utilities would otherwise take years to fund, according to Power Magazine. The same pattern holds in fiber. CBRE’s H2 2025 North America Data Center Trends report confirmed that hyperscaler and AI development continue to fuel construction of high-capacity fiber and conduit infrastructure, with new deployments establishing medium- and long-haul pathways that improve connectivity in markets that previously lacked it.
Hyperscaler investment in a market is not just a signal of strength. It is a direct investment in the power and network infrastructure that makes the entire market more valuable for every operator.
The Markets That Were Ready
The markets drawing the most serious attention right now share a consistent profile: available utility capacity, manageable permitting, lower operating costs, affordable land and realistic construction timelines.
Boise and the broader Treasure Valley region have been building toward this moment. The area benefits from favorable hydropower resources, abundant land and low natural disaster risk, according to Mordor Intelligence. Meta’s nearly one-million-square-foot campus in nearby Kuna, chosen in part because of Idaho’s data center tax incentive program, is one of the clearest signals that major operators view the market as viable at scale. Idaho Power’s 2025 Integrated Resource Plan cited unprecedented growth driven in part by new data center development in the region. ValorC3’s Boise data center has served enterprises in the market since 2020 and is currently expanding with a new 10 MW facility designed to support AI-ready colocation and high-density workloads.
Salt Lake City and Southern Utah are among the more underappreciated infrastructure markets in the Mountain West. Utah Governor Spencer Cox launched Operation Gigawatt to double the state’s power output, focusing on nuclear and geothermal generation. Utah’s sales and use tax exemptions on qualifying data center infrastructure make the long-term economics favorable. The region combines low hazard risk, strong fiber connectivity and a growing enterprise technology base. ValorC3’s St. George data center sits at the center of this market, purpose-built for enterprises that need enterprise-grade infrastructure without the cost structure of a primary hub.
Oklahoma, and specifically the Tulsa metro, has become one of the more active development markets in the country. Tulsa leads the state with 27 data centers, according to Tulsa Flyer, with more projects announced. Google committed $9 billion to expanding cloud and AI infrastructure in Oklahoma, with campuses in Pryor, Stillwater and a newly approved Tulsa location. Meta broke ground on a $1 billion, two-million-square-foot campus at the Fair Oaks Innovation Park in East Tulsa. Oklahoma produces more energy than it consumes, a structural energy advantage the state has made central to its data center recruitment pitch. CBRE specifically cites Tulsa’s low energy costs, affordable land, tax incentives and proximity to major business hubs as core draws. ValorC3’s Oklahoma City data center has operated in this market and positions enterprises to take advantage of the state’s energy surplus and growing infrastructure ecosystem.
These markets are gaining attention not because they are new. They are getting it because the constraints everywhere else finally caught up to the advantages they have held for years.
Regional Providers Are in Position
Hyperscalers are not the only ones repositioning in Tier 2 data center markets. Regional colocation providers have been building here for years and are now seeing demand catch up to the thesis they already bet on.
CBRE’s H2 2025 report stated that inference AI is redefining demand and creating a need for more regional and distributed data centers. Cushman and Wakefield’s Americas Data Center Update reached the same conclusion: secondary and tertiary markets with abundant land and power are positioned to capture a larger share of development as primary markets remain constrained.
The colocation market is projected to reach $239 billion by 2032 at a 14.7% compound annual rate, according to GlobeNewswire. AI-focused colocation is growing faster, at a projected 27.3% compound annual rate, according to Mordor Intelligence, driven by enterprise demand for low-latency inference without the capital burden of owning facilities.
Providers with existing capacity in Boise, St. George and Oklahoma City were positioned in these markets before the constraint became obvious. The demand they were built for has arrived.
The Workloads Are Starting to Spread Out
Large training environments will remain concentrated in hyperscale markets. That part is not changing.
Inference and agentic AI will spread differently.
A manufacturer running AI agents inside production environments does not care about proximity to Ashburn, Va. They care about latency, reliability, deployment timelines and what the power bill looks like in year three. The same logic applies to healthcare systems, logistics operators, financial services firms, and enterprise software providers that operationalize AI internally.
The edge thesis arrived before the workload was ready to support it. The compute deficit created by converging cloud growth, inference scaling and agentic AI adoption is not a short-term condition. It is structural and will run for years. The markets that secured power and land before the constraints became obvious are positioned to matter. The ones still waiting on Dominion are not.