AI data centers face cooling, grid breaking points

By July 2026, the planning baseline for AI facilities had plainly shifted to far denser rack environments, and that reset made cooling and power design inseparable. In Analysis, planners said “we’re seeing that really kind of serious planning around that 100kW racks and in some cases designs that are pushing it, you know, 300 and beyond,” adding that “Cooling becomes a primary design system, not a retrofit,” because “In 100 to 300kW environments, the constraint is really not just access to power” but whether operators can actually deliver and remove that energy at the rack and pod level.

By August and September, reports were tying those thermal realities to broader site scarcity: longer development timelines, scarcer power, land constraints, and a shrinking pool of facilities that could support AI densities without waste. That showed up in warnings that “It’s really whether it can hit the new rack densities without leaving parts of the hall underutilized or operationally compromised,” with colo liquid adoption still “maybe 30%,” while Analysis said “rack power requirements are moving from roughly 5 to 10 kilowatts to 100, 250 kilowatts” and “Cooling is moving from air to liquid as a requirement.”

The bottleneck in AI infrastructure has moved upstream to energized power and downstream to heat rejection. Global Data Center Hub reported that “lenders deployed $121B in construction financing… The constraint is no longer capital,” while hyperscalers chased energy-rich geography, with “Amazon expanded into Louisiana’s power-abundant corridor” and its “$12B Louisiana expansion tied hyperscale growth to power-rich corridors.” At the rack level, the same publication said “next-generation AI platforms were pushing racks into a heat range that legacy facilities cannot economically support,” with deployments planned at “50–100 kW/rack and beyond,” turning cooling from an operating expense into a hard participation limit.

That physical limit is why operators are redesigning both power intake and thermal architecture instead of simply ordering more servers. Global Data Center Hub defined the break point clearly: “Once racks cross roughly the 15–25 kW threshold where conventional air approaches start to lose efficiency,” owners must “rebuild the mechanical concept,” because “Chip-level draws of 700–1,000W” create hot spots and spikes that standard air cannot absorb; meanwhile National Grid’s commercial AI-factory model in Manassas is being proven at 100 megawatts with PJM and Dominion, including the ability to “go down to 75 megawatts” on grid-operator timescales, showing that flexible, modular power-and-cooling designs are becoming prerequisites for deployment.

By late summer 2026, the evidence showed that interconnection timing alone could blow up AI data center schedules by years. In the August 15 analysis, a speaker said, “today in Urkott in Texas… If you put in a request for a hundreds of megawatts of power to build a data center today, good luck getting that power before 2031 2032,” while Enlit World reported that “Industry analysis suggests it now takes, on average, more than four years to connect a new data centre to the grid,” adding that these “lengthy grid interconnection queues… threaten to substantially delay deployment schedules.”

The delay problem extends beyond utility hookups to the clean generation many projects expect to pair with new campuses, making synchronized launches increasingly unrealistic. In June 2026, “the Federal Energy Regulatory Commission (FERC) ordered six of the country's electric grid operators to justify or revise their processes for powering large energy users… aimed largely at speeding up the connection of data centres,” while Berkeley Lab data showed “Only 13% of generation capacity that entered U.S. interconnection queues from 2000–2020 had reached commercial operation by the end of 2025,” and Enlit World noted 59 dedicated off-grid power projects for U.S. data centres as of June 2026, with more than 90% announced since early 2025.

The clearest sign that liquid cooling has gone mainstream is that the debate is no longer about efficiency gains at the margin, but about whether air can cool modern AI racks at all. In March, Akash Systems CEO Felix Jackham argued there “isn't enough power in the grid in the United States to support the kinds of compute that we're talking about here,” pitching diamond heat spreaders that cool “very very hot GPUs” without “invest[ing] in a 100 times more infrastructure”; Akash says the material cuts GPU temperatures by 10 to 15 degrees C and can bring PUE closer to 1.0, underscoring how urgently operators are searching for alternatives to conventional air handling.

But those alternatives are being judged against a harsher baseline: air cooling works only up to older power envelopes, not the rack densities AI now demands. As one July analysis put it, “there’s no reason to use liquid cooling if your rack is below 20 kilowatts,” yet rising chip density means many U.S. halls would be “half or only a third populated” because they cannot cool more with air; likewise, “Legacy data centers were typically engineered for 5-15 kW per rack. AI clusters now push toward 100 kW and beyond… Air cooling hits fundamental physical limits at these densities,” while datacenterHawk said that above 100 kilowatts “air cooling doesn't really work anymore” and with Vera Rubin-class systems “the rack density grows to 200 plus.” That is also how operators describe the shift in practice: BDC says it “started our direct to chip liquid cooling even since the CPU era” and “we co developed with our customer and we've been using direct to chip liquid cooling since 2022,” concluding that “liquid cooling” is now part of the operating model rather than a niche add-on.

Developers are increasingly treating AI infrastructure as something assembled from pre-positioned, repeatable blocks rather than built from scratch, because that is one of the few ways to compress delivery schedules. On CoinDesk Podcast Network, CleanSpark said that at its Austin County site, the first 207 megawatts can go online quickly, with room for expansion and enough gas on site to support behind-the-meter generation and a ramp to 285 megawatts, a model meant to speed capacity delivery.

That same logic is pushing the market toward modular off-grid systems that can be scaled in factory-like increments, even if the engineering is harder than many forecasts assume. Sourcery quoted Exowatt’s pitch that a customer needing 100 megawatts or a gigawatt can stack units next to each other, while BUZZ HPC Insights said Craig Tavares stresses sovereign mandates, renewable power, and modular cooling, and Latitude Media noted some AI workloads may accept sub 99.999% reliability; a July analysis said 80% of the difficulty is the power electronics in the middle, while Semi analysis projected 40 gawatts of behind-the-meter generation at US data centers by 2028, mostly gas, and another interview said backup generators have been unchanged for about 20 years even as demand drives fuel cells and jet engines.