The data center construction boom, propelled by the artificial intelligence revolution, is encountering real-world obstacles. A recent dialogue between Goldman Sachs analysts and engineers from Microsoft’s former senior data center development team has revealed three key bottlenecks. Electricity supply remains the most pressing near-term constraint. Because cloud computing and AI inference workloads need to be located close to users, electricity shortages are emerging in congested markets. However, AI training workloads, which do not require proximity to users, are shifting to remote areas with ample power.
While flexible load management could potentially free up capacity, the industry’s inherent risk-averse culture hinders its widespread adoption. For instance, if data centers accepted minor load reductions, they could add load equivalent to 10% of the total peak demand in the U.S., but operators are uneasy about repeatedly powering equipment on and off. As a temporary measure, on-site power generation costs 5 to 20 times more than grid-supplied electricity. Yet, considering the profitability of large-scale AI data centers, this remains economically viable to get projects started, with the ultimate goal of connecting to the grid within three years.
Water stress is forcing the industry to transition from traditional, high-water-consumption evaporative cooling to more water-efficient but energy-intensive cooling technologies, particularly evident among hyperscale operators. This shift could cause Power Usage Effectiveness (PUE) to rise from an optimal level of 1.08 to between 1.35 and 1.40, meaning energy overhead surges from 8% to over 35%. Although new technologies like chip-level liquid cooling are emerging, colocation data centers, serving a diverse client base, may stick with chilled water system designs.
A shortage of skilled workers presents the next challenge. The specialized electrical and mechanical systems required by data centers make electricians and pipefitters critically important. According to Goldman Sachs estimates, by 2030, the U.S. will need a net increase of over 500,000 workers in manufacturing, construction, operations and maintenance, and power transmission and distribution to meet the electricity demands of all data centers. Industry groups are partnering with technical colleges to develop training programs to bridge this gap.
Meanwhile, tech giants are snapping up land at unprecedented prices, directly squeezing residential development. In Northern Virginia, the self-proclaimed “Data Center Alley” and global capital of data centers, soaring land prices are deterring residential developers. According to the Wall Street Journal, Amazon (AMZN) paid $700 million last November for a land parcel that a residential developer had acquired for just over $50 million a few years earlier, setting a record for a U.S. vacant land transaction.
In this region, rural land that once sold for tens of thousands of dollars per acre is now being quoted at over $3 million. This land grab has spread to areas near Chicago and Dallas, where one residential developer, aiming to build a data center, even purchased and demolished a 55-home subdivision. A residential land developer stated bluntly that homebuilders simply cannot compete on price.
Looking ahead, the sustainability of this construction frenzy underpins core assumptions in the macroeconomic narrative and tech stock valuations. While the investment thesis expects data center construction to translate into productivity gains and support multi-year growth, execution risks are mounting—from electricity and supply chain bottlenecks to soaring land costs—which could dash overly optimistic expectations.
