Hyperscalers face cooling, grid limits

The AI boom has a very physical problem. Hyperscalers can still buy GPUs, raise capital, and announce giant campuses. But getting enough electricity to the site, getting enough heat out of the building, and getting local approval to build the thing are starting to bite harder than the servers themselves. That became clearer over the last few weeks as Meta’s giant Richland Parish project in Louisiana turned into a power story as much as a tech story. Entergy said it needs 10 gas plants to support the site, with more than 7 gigawatts of capacity tied to the buildout. At the same time, developers and analysts have been pointing to a broader pattern — more projects getting delayed, resized, or canceled because the grid, cooling systems, and local politics are not moving at AI speed. ### Why is power the first hard limit? A modern AI campus is not just a bigger cloud data center. Training clusters pack dense accelerators into racks that pull enormous power and throw off enormous heat. That means operators need not just megawatts on paper, but firm delivery, substations, transformers, backup systems, and often new generators for key electrical gear. ### Why is cooling suddenly just as important? AI racks are getting so dense that old air-cooling assumptions stop working. Once you push toward the newest accelerator-heavy designs, cooling becomes part of the site-selection problem, not just a facilities detail. Liquid cooling is moving from nice-to-have to necessary for many, right when many of the hottest data-center markets are already resource-constrained. ### What does Meta’s Louisiana project show? It shows the scale of the workaround. Instead of waiting for the grid to magically catch up, the answer is increasingly dedicated generation and custom utility arrangements. Entergy boosted its capital plan to $57 billion, largely because of the Meta load, and around that customer. ### Are projects actually getting blocked? Yes — and not just slowed by engineering. Utility Dive cited Baird analyst Justin Hauke saying canceled data-center builds rose to 25 in 2025 from six in 2024. Local opposition is now a real variable, especially where residents worry about water use, noise, transmission buildout, diesel backup, to new hyperscale proposals. ### So why are companies still spending so much? Because the demand signal is still huge. Goldman Sachs has been framing the AI buildout in the hundreds of billions, with the five biggest U.S. hyperscalers projected for a combined $736 billion of capex across 2025 and 2026 in one estimate. But that spending no longer guarantees fast deployment. Land with power access, cooling headroom, and political permission now carries a scarcity premium. ### What changes next? Site selection gets more strategic and less glamorous. Operators will chase places with faster interconnection, cheaper power, and fewer permitting fights. They will sign long-term utility deals, add more on-site generation or storage, and design around liquid cooling from day one. Some of the AI race will be won by whoever can secure electrons and remove heat fastest — not whoever makes the splashiest model announcement. ### Bottom line? The constraint has moved down the stack. AI used to look chip-limited. Now it looks infrastructure-limited. Hyperscalers can promise trillion-parameter futures all day, but the real gate is a much more old-economy trio: power, pipes, and permits.