The PJM capacity auction results landed in January with a number that made utility executives wince: $23.1 billion in costs tied to data centers for the period from June 2025 to May 2028. That figure alone tells you something about the collision course the industry is on—half of all new U.S. power demand through 2030 is expected to come from data centers, according to Axios, and the infrastructure to support it simply isn't there.
Not the grid. Not the water supply. And increasingly, not the cooling systems designed for an era when server racks drew 5 or 10 kilowatts instead of the 100 to 150 kilowatts that AI workloads now demand.
Which brings us to Madrone, a YC-backed outfit founded in 2025 by two engineers with backgrounds at Apple and QuantumScape. The company claims it can cut cooling energy and water consumption by 30% in hot, arid climates—without installing a single mechanical chiller. It's a bold pitch, especially from what has been described as a two-person operation. But in an industry where up to 40% of a facility's power budget can vanish into cooling systems, operators are listening.
"Same grid permit. More compute," goes Madrone's tagline. In Texas or Arizona, where power is constrained and water is scarcer by the year, that's a value proposition with teeth.
When the Infrastructure Can't Keep Up
The data center industry is in the midst of what JLL has labeled a "$3 trillion investment supercycle." Global capacity is projected to approach 200 gigawatts by 2030. The International Energy Agency estimates that data center electricity consumption could reach roughly 945 terawatt-hours by decade's end under base-case scenarios, with the U.S. accounting for close to 45% of that load as of 2024.
AI is the engine. Training and inference workloads generate heat densities that legacy cooling systems—designed for an earlier generation of servers—simply can't handle efficiently. And the growth is running into hard physical limits.
In May of last year, the CEO of the largest U.S. grid operator warned publicly that the system needs fundamental redesign to accommodate AI's appetite for electricity. The IEA notes that roughly one in five planned data center projects now face delays tied to grid constraints. It's not an abstract problem. Operators are waiting months, sometimes longer, for interconnection approvals that used to take weeks.
Water is proving equally stubborn. Amazon Web Services disclosed that its facilities withdrew 2.5 billion gallons in 2025, though the company pointed to a water usage effectiveness figure of 0.12 liters per kilowatt-hour—well below what it cited as an industry average of 0.84. Microsoft, at its Build conference last June, claimed that new closed-loop designs could reduce water consumption at flagship AI campuses to levels comparable to "a single restaurant" annually. Impressive, if true at scale.
But a report from around the same time found that two-thirds of some 809 planned U.S. AI data center projects are sited in drought-affected areas. Arizona paused data center sales-tax exemptions for three years last June. Illinois slowed development incentives, citing power and water concerns. Ireland has effectively frozen new data center connections in Dublin until at least 2028.
The politics are starting to catch up with the physics.
The Cooling Dilemma
Here's the core tension: AI chips run hot, racks are getting denser, and the economics of keeping everything cool are becoming harder to justify.
Madrone's website claims that every 1% improvement in cooling efficiency translates to $2.6 million in additional GPU revenue per 100 megawatts per year. The Uptime Institute's 2025 survey found that many sites still operate with a power usage effectiveness ratio above 1.5, meaning operators are losing meaningful revenue to inefficiency. Perhaps more striking, the survey showed that average PUE has remained flat for six years—constrained by legacy infrastructure and regional climate realities that no amount of software optimization can overcome.
The industry has responded with a push toward direct liquid cooling. Equinix, Digital Realty, and others have launched DLC labs and deployment programs. Vertiv announced a $50 million Ohio expansion last March to scale liquid cooling manufacturing, with operations expected to come online in the second quarter of this year. ZutaCore raised $100 million in June to scale what it calls waterless two-phase direct-to-chip cooling.
But here's the thing: DLC solves the chip-level heat problem. It doesn't eliminate the need for facility-level heat rejection. You still have to get the heat out of the building, and that's where legacy systems—mechanical chillers paired with cooling towers—eat through power and water budgets.
That's the opening Madrone and a handful of competitors are targeting.
Dew-Point Cooling: Old Idea, New Urgency
Madrone's approach centers on dew-point cooling, a variant of indirect evaporative cooling that uses alternating dry and wet channels to cool product air toward its dew point—often below the wet-bulb temperature—without adding moisture to the airstream.
The technology itself isn't new. Academic papers exploring dew-point systems for data centers date back nearly a decade, and a 2023 study in Applied Energy documented a field trial of what researchers called "super performance dew point cooling" at a live UK data center, reporting significant energy savings versus legacy vapor-compression systems. A 12-month monitoring study at a UK edge data center, published as a preprint last year, provided the first full-year real-world evaluation.
Madrone—co-founded by Akshay Trikha, who worked on machine learning at QuantumScape and materials prediction at Berkeley, and Erik Meike, who spent time on wireless power for the iPhone at Apple—positions its system as a facility-level solution that "sits outside the data center and provides cool water." The company lists prototypes ranging from 1 kilowatt to 100 kilowatts and claims it can deliver 20°C supply water in hot-arid climates without mechanical chillers.
The Texas-focused pitch suggests performance gains over traditional chiller-plus-tower baselines in desert conditions, though field data from diverse climates remains limited.
Other players are exploring adjacent territory. Seeley International markets the Climate Wizard line of dew-point class indirect evaporative coolers. Dew Point Systems, a UK spin-out backed by Cambridge Future Tech, claims tenfold efficiency gains over traditional air conditioning, though it has not published data center-specific performance metrics. Munters has deployed its Oasis indirect evaporative cooling system at facilities including DigiPlex's Fetsund site in Norway, where it handles around 10 megawatts of heat rejection, and Sabey's Intergate.Quincy facility in Washington state.
Hyperscalers, meanwhile, are taking varied paths. Google's Hamina data center in Finland uses seawater cooling without chillers, achieving very low PUE in a uniquely favorable climate. Nautilus Data Technologies operates a barge-based facility in Stockton, California, using a closed-loop system with river water for heat exchange and no cooling towers. Microsoft's closed-loop designs target minimal evaporative water use. AWS emphasizes maximizing outside-air cooling hours and shifting to closed-loop liquid where climate and workload permit.
The dew-point cooling segment, though, faces a credibility gap. Performance claims vary widely depending on climate, baseline assumptions, and system design. A UK university study might show one outcome; a Texas deployment could look very different. Madrone's 30% savings claim is specific to hot-arid climates compared to chiller-plus-tower systems—a narrower scope than it might initially appear.
What Happens Next
The trajectory for data center cooling is being shaped by three overlapping trends: the push to eliminate chillers where possible, the rise of hybrid architectures, and the growing importance of water as a constraint alongside power.
NVIDIA CEO Jensen Huang made waves at CES last year when he suggested that next-generation Rubin-class racks could operate with warm water, potentially eliminating chillers. Cooling equipment stocks dipped on the news. If AI chips can tolerate higher coolant temperatures, the economics of chiller-less cooling improve—but site climate, redundancy requirements, and the mix of IT workloads will determine how widely that approach scales. And silicon vendors have a habit of making optimistic projections.
Hybrid architectures are becoming the norm, perhaps out of necessity. High-density AI racks increasingly use direct liquid cooling at the chip level, while facility-level heat rejection may rely on a combination of dry coolers, indirect evaporative systems, adiabatic cooling, or mechanical chillers depending on climate and water availability. McKinsey's analysis of AI data center infrastructure from the past couple years has highlighted this shift toward multi-modal cooling strategies.
Water policy is tightening faster than many operators expected. The EU's data center sustainability rating scheme now includes mandatory water reporting. U.S. states and localities are experimenting with project-specific reviews, especially in drought-prone regions. Investors are pushing for site-specific water and power disclosures. AWS, Microsoft, and Google have all committed to water-positive operations by 2030, but progress is uneven and scrutiny is intensifying.
For startups like Madrone, the opportunity lies in bridging the gap between legacy chiller-based systems and the emerging chiller-less or chiller-optional architectures that hyperscalers and chip vendors are exploring. If the technology can deliver meaningful savings in real-world hot-arid deployments—and do so reliably across a range of humidity and temperature conditions—it could find a market among operators facing tight power budgets and water constraints.
The Practical Questions
The questions, of course, remain practical. How does dew-point cooling perform during extended heat waves when wet-bulb temperatures spike? What happens during dust storms or periods of low humidity? How does it scale beyond edge deployments to hyperscale facilities?
And perhaps most importantly: Can a small startup manufacturing in California compete on cost and delivery timelines with established cooling OEMs ramping production to meet surging demand? The data center buildout isn't waiting for anyone. Projects that were supposed to break ground this year are already delayed. Operators need solutions that work now, not prototypes that might work eventually.
The industry is awash in promises of efficiency gains. The winners will be the ones that can deliver them at scale, in diverse climates, with predictable economics. For Madrone and others betting on alternatives to mechanical refrigeration, the clock is ticking.
The grid constraints aren't easing. The water scrutiny isn't going away. And the AI buildout—whatever its ultimate trajectory—shows no signs of slowing down in the near term. The question is whether innovation in cooling can keep pace with the heat.
If it can't, the $3 trillion investment supercycle might hit a wall sooner than anyone in the industry wants to admit.
