Share with your CTO
The global AI data center pipeline carries a $2.3 trillion headline, but only $492 billion of it is actually under construction, and the gap between those two numbers is where most capital allocation mistakes will happen. A cross-analysis of 14 major industry reports from JLL, CBRE, Goldman Sachs, Structure Research, and others finds that gross capacity announcements of roughly 223 GW will net out to about 150 GW of funded, energized AI infrastructure by 2030. Construction costs are up 21% per MW since late 2024. Atlanta has passed Northern Virginia for capacity under construction, and 77% of North American builds are now in power-rich frontier markets rather than traditional hubs.
What this means for your business
Where you sit on the delivery pipeline determines whether these numbers are an opportunity or a liability. Organizations consuming compute at scale are insulated from pipeline noise if their workloads are already on contracted infrastructure, but CTOs sourcing capacity through 2026 and 2027 are negotiating in a market where 80.4% of capacity under construction is already pre-leased and primary-market vacancy sits at 1.4%. The question isn’t whether demand is real. It’s whether the specific capacity you’re being sold has power, financing, and a realistic commissioning date behind it.
The piece introduces a seven-step ladder from “announced” to “accepted and earning revenue,” and the analytical value here is genuine. Most headline pipeline figures live at steps one through three, where a sponsor has land or a letter of intent but no contracted power or construction financing. Steps four and above, contracted power plus take-or-pay lease terms plus construction underway, are where supply actually becomes usable capacity. CTOs evaluating colocation contracts or long-term capacity commitments should be asking providers to demonstrate which rung they’re on, not accepting megawatt totals as evidence. High-yield bond markets raised $27.8 billion for data centers in H1 2026, with lenders extending 70-80% loan-to-cost on projects backed by non-investment-grade tenants at spreads 200-300 basis points wider than top-credit deals. That’s the market pricing construction and tenant risk correctly. Your contract terms should do the same.
The power constraint is the durable one. PwC projects equipment rising from 70% to 93% of AI infrastructure investment by 2050, which means the building itself becomes a smaller and smaller fraction of the value chain over time. The organizations that capture lasting advantage aren’t the ones that own the most announced megawatts. They’re the ones that secured firm power delivery dates early, because power availability is now the actual gating resource, not capital or land. If your infrastructure roadmap still treats location prestige, say, proximity to a primary market, as a proxy for delivery certainty, the data says that model broke somewhere around 2024.
Concept deep-dive: Take-or-pay clauses
A take-or-pay clause obligates a tenant to pay for a fixed share of contracted capacity whether they consume it or not, typically 60-85% of their allocated power in current data center deals. Think of it as a minimum monthly bill that runs regardless of usage, similar to a cell plan with a large base charge. For CTOs signing long-term capacity agreements, these clauses shift demand-forecast risk from the developer to the tenant, making accurate workload projection a direct financial exposure rather than an operational planning problem.
Based on reporting from The AI Infrastructure Pipeline Is Not the Investable Market, originally published 2026-11-03 21:01:00.
