Why AI Inference Pulls Data Centers Back to Tier-One Metros
Andrew Thomas explains why AI inference demand will dwarf training capacity and force data center capital back to tier-one urban hubs.
Critical Load: What’s next for data centres and AI?
With James Linacre, Andrew Thomas, Dev Gupta, Alexey Teplikhin, Jan Vesely · Sunday, September 13, 2026
In the finale of the 'Critical Load' miniseries, leading infrastructure investors evaluate data center unit economics, return profiles, and platform exit mechanics across global markets. The panel examines the shift from AI model training to low-latency inference, rising construction execution risks from inexperienced market entrants, and innovative bridging power and cooling solutions to resolve grid and water bottlenecks.
Key takeaways
- Andrew Thomas projects that AI inference compute demand will expand to several multiples the physical footprint and capacity of training clusters. Read more →
- Alexey Teplikhin warns that generalist sponsors entering data center megaprojects lack the engineering and supply chain capabilities required for mission-critical builds. Read more →
- Large hyperscale data center platforms have grown too massive for standard sponsor-to-sponsor sales, making public listings an eventual route despite friction. Read more →
- Standard grid interconnection queues now stretch beyond the 12- to 36-month delivery windows demanded by major cloud and enterprise tenants. Read more →
- Hyperscale campuses require 100 megawatts or more of capacity, with build costs running from $10 million to $15 million per megawatt in the United States. Read more →
- Direct-to-chip liquid cooling is replacing hybrid adiabatic and air-cooled setups as rack densities increase, altering unit cost profiles for high-density compute assets. Read more →
6 articles from this episode
Why Inexperienced Capital in Data Centers Will Get Burned
Alexey Teplikhin warns that generalist capital entering data center megaprojects faces major execution failure and capital impairment risks.
Why Hyperscale Data Center Platforms Face an Exit Bottleneck
Data center platforms face public market friction and heavy capex demands, driving sponsors toward private recapitalizations over IPOs.
Why Data Center Developers Are Turning to Bridge Power
Data center developers face 36-month grid queues. Andrew Thomas and Jan Vesely explain why on-site power bundling captures peak infrastructure yields.
Data Center Yields: Underwriting Hyperscale vs. Urban Colocation
Infrastructure investors target 10% yields on hyperscale campuses and mid-teens on urban colocation as global build costs hit $20M per megawatt.
Why Data Centers Are Adding Desalination and Direct-to-Chip Cooling
Dev Gupta explains how data center operators use liquid cooling, gray water, and prefabrication to beat water scarcity and permitting bottlenecks.