The AI race is running into an unexpected wall-and it's not chips or capital. It's water, power, and local resistance. As Amazon.com, Inc.
AI Growth Meets Real-World Limits
According to a Reuters report, multiple hyperscalers have already scrapped or delayed billion-dollar data center projects in the U.S. amid community pushback over water usage, land, and energy demand. Shareholders are now pressing for greater transparency, particularly around water consumption-an issue that's quickly moving from ESG talking points to operational risk.
The scale is hard to ignore. North American data centers consumed nearly 1 trillion liters of water in 2025, roughly equivalent to New York City's annual usage. At the same time, emissions are moving in the wrong direction-Google's emissions have risen sharply since its 2020 targets, raising fresh questions about how AI expansion aligns with climate goals.
Even where companies are improving efficiency, disclosure remains uneven. Investors, including Trillium Asset Management, are demanding site-level clarity, arguing that without it, the true cost-and risk-of AI scaling remains opaque.
China's Radical Workaround
While U.S. tech giants grapple with constraints on land, China is testing a different model-moving compute offshore. A new underwater data center off Shanghai, developed by Shanghai HiCloud Technology, sits 10 meters below the surface and uses natural seawater cooling to reduce energy intensity.
The concept isn't entirely new-Microsoft previously explored subsea data centers through Project Natick-but China's deployment marks a shift from experimentation to commercialization.
Designed for AI workloads, these facilities leverage stable ocean temperatures, modular expansion, and proximity to coastal infrastructure to sidestep some of the bottlenecks facing land-based builds.
The Shift Investors May Be Missing
The implication is subtle but important: the AI trade may be moving beyond chips and into infrastructure constraints. Water access, permitting, and community acceptance are starting to dictate how fast-and where-AI can scale.
If that's the case, the winners may not just be those building the most powerful models, but those finding ways around the physical limits of powering them.
