The numbers are almost absurd. Texas grid operator ERCOT is staring down a queue of interconnection requests that could total over 400 gigawatts—potentially reaching 450 gigawatts—more than the entire electric capacity of Germany. Many of those requests are data centers, the kind packed with AI chips that gulp electricity and throw off enough heat to melt themselves if the cooling fails.
But power, it turns out, isn't the only bottleneck. Water is becoming just as tight a constraint, particularly in a state where droughts are a recurring fact of life and aquifers are already under pressure. The dilemma facing developers is stark: use water to cool efficiently, or accept higher electric bills—and carbon emissions—to stay dry.
Enter Madrone, a two-person startup that emerged from Y Combinator's Spring 2026 batch with a bold claim: it can eliminate mechanical chillers entirely from data center cooling systems. The company says its technology can supply chilled water at roughly 20 degrees Celsius using something called dew-point cooling, cutting both power and water consumption by 30 percent compared to the traditional chiller-plus-cooling-tower setups that dominate Texas installations. If those numbers hold up in the field—and that remains an open question—it could matter a great deal in a state scrambling to build infrastructure fast enough to meet demand.
"Every 1% of cooling efficiency equals $2.6 million more GPU revenue per 100 megawatts per year," reads a line on Madrone's YC profile. It's the kind of math that gets hyperscalers' attention when their combined capital expenditure is approaching three-quarters of a trillion dollars annually.
A Market Shift Underway
The geography of American data centers is shifting. Texas and the Midwest, which accounted for about a third of U.S. hyperscale capacity at the end of last year, are now capturing more than half of new pipeline capacity, according to Synergy Research. The reasons are familiar: cheaper land, looser regulation, and—until recently—abundant power. ERCOT approved a new "Batch Zero" review process in June to study interconnection requests for loads of 75 megawatts or larger, many of them AI facilities. The grid operator consumed roughly 45 terawatt-hours servicing data centers last year, a figure that's only climbing.
But the water question is emerging as co-equal with energy, particularly as the big cloud providers race to publicize "low-water" or "zero-water" designs. In a June 24 blog post, Microsoft announced that its newest AI data centers operate with "zero water during normal operations," relying instead on closed-loop systems and air-cooled chillers. The company reported portfolio-wide water usage effectiveness of 0.27 liters per kilowatt-hour last year, down from higher historical levels, and declared itself water-positive in its most recent fiscal year. Google, meanwhile, is shifting toward air cooling in water-stressed regions like Phoenix, though two-thirds of its global footprint still uses evaporative cooling where conditions allow. Amazon reported 2.5 billion gallons of water use last year in a June 12 disclosure, with a usage effectiveness ratio of 0.12 liters per kilowatt-hour, which it says sits below industry norms.
The problem is that air-cooled chillers and dry coolers solve the water issue by burning more electricity. A study published mid-year in Scientific Reports found that rising heat and humidity are shrinking the number of hours when "air-free cooling" remains viable, especially in the tropics and the American Southeast. Texas Gulf Coast humidity is a particular headwind for pure air-side economization. You pick your poison: water or watts.
Regulatory pressure is mounting, too. The EU's Energy Efficiency Directive now mandates annual reporting of power usage effectiveness, water usage effectiveness, and renewable energy factors. In Texas, the state's Public Utility Commission is grappling with poor transparency on data center water consumption—a recent Texas Tribune investigation noted weak survey compliance. Meanwhile, a federal proposal floated earlier this year would require data centers receiving federal support to track both water and energy consumption.
Cooling accounts for somewhere between 30 and 40 percent of total facility energy depending on climate and methodology, according to multiple industry sources, and the International Energy Agency reported that data center electricity use grew 17 percent last year, now consuming about 2.6 percent of global demand. Small improvements compound quickly at that scale.
The Physics Are Old; The Field Data Are New
The underlying physics of dew-point cooling have been understood for years. Indirect evaporative systems that approach the air's dew point—often using what's known as the Maisotsenko cycle—can cool a working fluid below the wet-bulb temperature with minimal electrical input, avoiding mechanical compression entirely. Peer-reviewed work in journals like Nature and Applied Thermal Engineering has been optimizing these M-cycle devices for some time, with papers appearing as recently as late last year.
What's changed is the field validation. A year-long monitoring study of a dew-point cooling system installed on a live edge data center in the UK was posted to SSRN earlier this year, providing the first full-year real-world evidence of performance under operational conditions. A separate levelized-cost analysis published last fall in RSC Sustainability examined hybrid indirect evaporative systems paired with mechanical backup, finding that the economics pencil out in the right climate zones.
Texas is one of those zones. At least part of the year.
Madrone was founded last year by Akshay Trikha, who worked on machine learning and materials at battery startup QuantumScape and UC Berkeley, and Erik Meike, an Apple veteran who handled iPhone power and thermal engineering at Harvey Mudd. Their YC profile emphasizes "zero mechanical chillers" and that 30 percent reduction in power and water for Texas deployments. The company is backed by Y Combinator and enrolled in the NVIDIA Inception Program, though no public funding amounts have been disclosed. A pilot deployment was referenced in LinkedIn activity by Meike in June, but no third-party performance data has been published.
The 30 percent claim isn't unprecedented. Nortek's StatePoint Liquid Cooling system, developed in partnership with Meta starting in 2018, reports 20 to 30 percent less power and 35 to 50 percent less water than traditional cooling tower setups in case studies from Digital Edge in the Philippines and elsewhere. Those systems, however, often remain hybridized with mechanical backup for worst-case conditions—the hottest, most humid days when dew-point alone can't keep up. Munters, a Swedish cooling giant, has been pushing waterless thermosiphon and dry cooler systems that eliminate evaporative water use entirely, though they face efficiency tradeoffs in very hot climates. The company secured an $82 million chiller order from a U.S. data center provider last year.
The Rack-to-Room Integration Challenge
Google threw another wrinkle into the mix with its announcement of "Brazos," a rack-mounted liquid cooling system that lets operators drop liquid-cooled gear into legacy air-cooled rooms. If facility-level cooling can reliably deliver 18 to 22 degrees Celsius water to coolant distribution units—exactly the range Madrone is targeting—it becomes strategically important as a bridge technology. Research firm Omdia noted recently that gigawatt-scale campuses will require fundamental power and cooling redesigns, with targeted direct liquid cooling in high-density zones paired with more efficient air cooling elsewhere.
Perhaps more than the founders expected, the market window appears to be narrowing. ERCOT's Batch Zero notifications are targeted for late summer, and developers who can credibly promise lower water draw and lower peak demand will likely have an edge in the queue. San Antonio is already marketing its recycled water network as a data center advantage, and Google's announcement of Texas sites earlier this year emphasized minimal water use for cooling.
The Uptime Institute's survey last year found that power usage effectiveness improvements have stalled for six years straight, suggesting the industry is hungry for new approaches—or at least willing to entertain them. Jefferies estimated that hyperscaler capital expenditure is approaching $700 to $725 billion this year. Even a percentage point or two of efficiency gain translates to real money at that scale.
The Hybrid Future
The trajectory seems fairly clear, even if Madrone's specific role in it remains uncertain. Hybrid systems will dominate. Dew-point and indirect evaporative technologies will be paired with mechanical chillers or high-temperature dry coolers to handle humidity spikes and extreme heat days. A climate analysis found that Texas Gulf Coast sites can't rely on free cooling year-round, and Madrone's approach will need intelligent controls or mechanical assist for worst-case conditions—something the company hasn't detailed publicly yet.
But the market opportunity is enormous if the technology works at scale. The question is whether Madrone can move from pilot to production before the established players lock in Texas deployments. Nortek, Munters, and the hyperscalers' in-house engineering teams aren't standing still.
Dew-point cooling won't eliminate the need for liquid cooling at the rack or mechanical backup on the hottest days. But it could carve out a middle path that beats traditional chillers on both power and water—assuming, of course, that independent, third-party validation of Madrone's specific claims in a Texas deployment materializes. That remains the critical missing piece.
If the 30 percent savings prove out in the field, expect the technology to spread fast. Texas has a way of setting trends when it comes to infrastructure—for better or worse.
