Barely a week passes without another warning that the AI buildout has overshot itself. To be fair, there is plenty of cause for concern: circular financing between model developers and their infrastructure partners, eye-watering capex guidance, power generation capacity constraints, and GPU supply chasing unproven monetization assumptions. The data lends some support — as many as half of the world's data center projects slated for this year could face delays, driven by power constraints, equipment shortages, and rising opposition [1], and Q1 2026 alone saw more capacity blocked or delayed than almost any comparable period on record [2]. But delay isn't demise. Large-scale builds typically run 12 to 18 months [1], which is long enough that most "delayed" capacity is still on track, just on a tighter timeline than developers modeled. This capacity is business- and national-security-critical, and capital of that strategic weight doesn't walk away from a permitting fight.
We've been studying this market closely for the past 18 months. Our view: it's not a question of if, but where and when AI data centers will be built.
The clearest evidence this is a siting problem, not a demand problem, is geographic. Northern Virginia, still the world's largest data center market, is hitting hard limits: the average wait for a large interconnection now runs roughly seven years, pushing land costs up accordingly [3]. Capital is following available power and permitting certainty ahead of nearly every other site-selection factor, pulling development into Texas (on pace to overtake Northern Virginia by 2030) [4], the Midwest (Indiana, Ohio), Georgia, and Pennsylvania [3][4].
Set aside siting, and the demand case remains intact. Enterprise AI adoption, agentic workflows, inference at scale, sovereign AI, and defense and intelligence applications are all durable sources of compute demand independent of any single model developer's fortunes. If the bubble commentary proves right and today's leading labs face a reckoning, that changes who owns the physical assets (through distressed sales, consolidation, or repurposing) far more than it changes whether they get used. Infrastructure built ahead of a commercial inflection point has a long history of outliving the companies that first justified it. The fiber buildout of the dot-com era appeared vastly overbuilt in 2001, but became critical infrastructure by the end of the decade.
The bubble commentary often conflates the economics of frontier model developers with the physical infrastructure underneath them. There are legitimate questions around compute provisioned ahead of proven monetization, circular vendor financing, and valuations assuming today's growth persists. Questions about whether the power, electrical components, fiber, and cooling built to support it are similarly overextended are different entirely. Compute capacity can run ahead of monetized demand while the physical shell around it stays structurally undersupplied.
Beyond the model layer, the picture looks far less speculative. Across power, electrical equipment, fiber, and cooling, investment is driven by physical scarcity, not sentiment about any one lab's roadmap. Operators are no longer waiting on utilities: roughly two-thirds are now deploying or considering on-site power generation to bypass grid queues entirely [5]. Delivery times for critical electrical equipment have stretched from less than two years pre-2020 to as long as five years today [6], and rising rack densities (now averaging 27 kW per rack, up from 16 kW last year) [5] are compounding the strain on cooling at the same time. These backlogs are measured in years and ordered by a customer base spanning well beyond AI into grid modernization and industrial electrification — making this layer more durable than the narrative above it.
Risk is distributed unevenly across this stack. It's useful to think in three tiers, from most speculative to most insulated.
AI bubble concern is real but needs narrower scoping — it describes the economics of the model developers at the top of the stack, not the infrastructure being built beneath them, which is proceeding on its own timeline and geography. For clients evaluating exposure, that reframing has a direct allocation implication: exposure further down the stack offers a way to participate in AI-driven demand growth with lower correlation to today's sentiment swings.
This is where BCE adds the most value — helping clients understand data center dynamics and pipelines market by market, incorporate strategic decisions, identify priority growth opportunities, and position portfolios or supplier relationships deliberately across the three tiers.