Two seemingly contradictory narratives circulate simultaneously in discussions of the AI data center market: that the sector is at risk of overbuilding capacity ahead of genuine demand, and that genuinely usable, power-secured capacity remains scarce and likely to remain so for years. Both narratives contain real truth, and reconciling them requires distinguishing between different categories of capacity rather than treating the market as a single undifferentiated pool.
Why Both Risks Can Be True at Once
Uptime Institute's 2026 industry predictions capture this tension directly, noting genuine uncertainty about how AI will reshape demand, even as developers face structural constraints meaning they will not, in the research's own framing, "outrun the power shortage." Aggregate announced project capacity across the industry may well exceed what underlying AI demand ultimately requires, particularly if efficiency improvements continue at the rapid pace the IEA has documented, reducing the energy and compute required per unit of AI output over time. At the same time, the specific subset of capacity that is genuinely power-secured — with confirmed interconnection, transformer and switchgear procurement, and a credible energisation timeline — remains scarce, regardless of how much capacity exists on paper in earlier-stage project pipelines.
Distinguishing Announced Capacity From Deliverable Capacity
- Industry tracking consistently shows that only a fraction of announced gigawatt-scale data center capacity is actually under active construction at any given time, with much of the remainder stalled behind grid, equipment, or permitting constraints
- This gap between announced and genuinely deliverable capacity means headline figures about total planned capacity can be deeply misleading as an indicator of actual near-term market supply
- Power-secured, fully de-risked capacity — the kind discussed throughout much of our analysis of grid interconnection and equipment procurement risk — represents a meaningfully smaller and more valuable subset of the total announced pipeline
The AI data center market is not simply oversupplied or undersupplied — it is oversupplied with speculative announcements and undersupplied with genuinely power-secured, deliverable capacity, and those are very different things to be exposed to as an investor.
What This Means for Investment Strategy
This distinction argues for investment strategies that specifically target genuinely power-secured, executionally credible projects, rather than broad exposure to the sector's full announced pipeline, much of which may never be delivered or may be delivered on a substantially different timeline than announced. It also argues for caution around speculative development that assumes demand will simply absorb whatever capacity is built, given genuine uncertainty about the pace and shape of future AI compute demand growth discussed in our analysis of scenario planning.
Demand Efficiency Gains Add a Further Layer of Uncertainty
The IEA's observation that energy use per AI task is falling at a rate unprecedented in energy history adds genuine complexity to overbuild risk assessment. If efficiency gains continue to outpace usage growth in certain segments, demand for raw compute capacity could grow more slowly than current aggressive build-out assumptions imply — even as overall AI adoption continues expanding. Investors should treat this efficiency trend as a genuine source of demand uncertainty, not dismiss it in favour of simpler, more bullish extrapolations of current growth rates.
Navigating This Balance With Disciplined Analysis
DATAPERT helps investors distinguish genuinely deliverable, power-secured opportunities from speculative pipeline exposure as part of our investment intelligence services. Explore our broader data center development advisory or start a project to discuss a disciplined approach to AI infrastructure investment.
