There's a number that keeps data center executives up at night, and it isn't the cost of NVIDIA's latest chips. It's 1,200 watts.
That's how much heat a single next-generation AI accelerator can generate—roughly the thermal output of a household microwave running continuously, crammed into a space smaller than your hand. Multiply that across a rack of 40 or 60 GPUs, and you're looking at thermal densities that would've been considered science fiction a decade ago. The servers powering ChatGPT and its successors don't just need electricity. They need to be kept from melting.
Data centers consumed 415 terawatt-hours globally in 2024, about 1.5% of worldwide electricity demand, according to the International Energy Agency. By 2030? The IEA projects 945 TWh. The explosion is driven almost entirely by artificial intelligence workloads, which require exponentially more computational horsepower—and therefore exponentially more sophisticated cooling—than the cloud computing tasks of yesteryear.
While Meta, Amazon, and Microsoft chase multi-gigawatt nuclear power deals and sign record renewable energy contracts, a different battle is unfolding inside the facilities themselves. It's quieter, more technical, and arguably just as critical: how do you cool chips that hot without bankrupting your energy budget or overwhelming an already-strained electrical grid?
When Software Meets Thermodynamics
Enter Glacian Technologies, a Penn State spinout that's betting federally funded physics research can solve what hardware alone cannot.
The company emerged from more than a decade of work at Penn State's Sustainable Buildings & Societies Lab, where Professor Wangda Zuo has been developing what the startup now markets as "Physical AI"—essentially, software that optimizes cooling systems by blending traditional thermodynamic simulations with machine learning. The lab has pulled in over $8 million from the Department of Energy, Department of Defense, NSF, ASHRAE, and JPMorgan Chase. Zuo was named an ASHRAE Fellow this year, a recognition that carries weight in the unglamorous but essential world of HVAC engineering.
The pitch sounds almost too good: slash cooling energy consumption by as much as 74%, based on DOE case studies at facilities in Massachusetts and Florida. That's not incremental improvement. That's the difference between adding another rack of GPUs or hitting what operators grimly call "the thermal wall."
But claims like that invite skepticism, especially in an industry that's seen plenty of vaporware and overpromised efficiency gains. What makes Glacian's case more credible is the paper trail—publicly accessible DOE project documentation, peer-reviewed research, and a provenance that predates the current AI gold rush by years.
The Florida facility saw a 53% reduction in cooling energy potential through end-to-end optimization of airflow and plant controls. Massachusetts hit 74%. These aren't lab simulations. They're real-world case studies, documented and funded by the federal government long before generative AI became a boardroom obsession.
The New Math of Data Center Economics
To understand why cooling suddenly matters so much, you need to understand the shift in data center economics over the past 18 months.
Gartner projects a 16% jump in data center electricity demand just this year, with AI-optimized servers accounting for 21% of total power consumption in 2025 and 44% by 2030. Goldman Sachs sees U.S. data center power demand rising as much as 165% by 2030 compared to 2023 levels, hitting 84 gigawatts by 2027. McKinsey pegs global investment needs at $6.7 trillion through decade's end.
Zoom in on the U.S., and the numbers grow more alarming. American data centers consumed 183 TWh in 2024—more than 4% of the national electricity supply, according to Pew Research and the Department of Energy's Lawrence Berkeley National Laboratory. In Northern Virginia, the world's most concentrated data center market, facilities already account for roughly 26% of the regional grid.
The physical reality is this: next-generation GPU chips are pushing rack densities from an average of 36 kilowatts in 2023 toward 50 kW by 2027. AI clusters routinely exceed 80 to 120 kilowatts per rack. Traditional air cooling simply can't keep up at those densities—physics won't allow it.
Liquid cooling is accelerating from niche to necessity. Omdia forecasts the data center thermal management market will grow from $7.67 billion in 2023 to $16.9 billion by 2028, with liquid cooling solutions posting the fastest compound annual growth rate. Direct-to-chip systems, rear-door heat exchangers, full immersion tanks—technologies that were exotic just five years ago are now standard in RFPs for hyperscale facilities.
But here's the catch. Hardware upgrades require time, capital, and often major retrofits that can take years to design and deploy. Software optimization of existing cooling plants? That can happen in months, assuming the vendor isn't overselling.
The Constraint No One Predicted
There's a paradox at the heart of the AI infrastructure boom, and NVIDIA CEO Jensen Huang said the quiet part out loud: energy capacity is now a strategic constraint.
Next-generation AI models, Huang noted publicly, will require roughly 100 times more compute than the early ChatGPT-era systems. But you can't just drop a hundred racks of H100s into a building and call it a day. Someone has to cool them. Someone has to power that cooling. And in grid-constrained markets—Northern Virginia, Dublin, Singapore—every megawatt allocated to cooling infrastructure is a megawatt that can't run servers.
Cooling traditionally eats up 30% to 40% of a data center's non-IT energy budget. That's power that could otherwise drive revenue-generating compute. In markets like Northern Virginia, where Dominion Energy paused and then selectively resumed new connections due to transmission bottlenecks, the calculus has shifted. Freeing up cooling capacity to redirect toward compute is, in effect, adding IT capacity without waiting for utility permits or major construction projects.
This shift—from an IT power bottleneck to a facility-level thermal and power bottleneck—might be the defining infrastructure challenge of the AI era. And it's here that Glacian's offering starts to look less like academic theory and more like practical necessity.
Physics Meets Machine Learning
The technology itself is rooted in Wangda Zuo's decade-plus of research into physics-based modeling and model predictive control for building systems. The innovation Glacian is commercializing blends traditional thermodynamic simulations with machine learning—hence "Physical AI," though that's marketing shorthand for something considerably more technical.
The platform uses modular digital twins built in Modelica, a physics-based modeling language favored in aerospace and automotive engineering but less common in data center infrastructure. It combines that with a rapid airflow regression tool the lab developed called ISAT-FFD (In-Situ Adaptive Tabulation with Fast Fluid Dynamics). The result: operators can simulate "what-if" scenarios in near real-time. What happens if we raise supply air temperature by two degrees? Shift economizer settings? Adjust chilled water setpoints?
These aren't abstract questions. In a facility running tens of megawatts, a two-degree temperature adjustment can mean hundreds of thousands of dollars in annual energy costs—or the thermal headroom to add another row of GPU racks.
Glacian also benefits from NSF I-Corps customer discovery work conducted through a related Penn State venture, NexDCCool Technologies, which shares the same research lineage and involves Zuo. Through dozens of I-Corps interviews, the team learned that data center operators didn't prioritize theoretical energy savings alone. They wanted faster deployment. Simpler integration. The value proposition evolved from "save money on utility bills" to "unlock compute capacity under power constraints."
That's a subtle but critical shift in positioning, perhaps more than the founders initially expected.
Crowded Field, Different Angle
Glacian isn't the only company chasing software-driven cooling efficiency. Google's DeepMind famously reduced cooling energy by up to 40% at its own facilities starting in 2016 using reinforcement learning, though that system remains internal—a competitive advantage Google isn't licensing. Meta announced in 2024 that simulator-based RL cut supply fan energy roughly 20% and water use 4% in a pilot region, and the company is now applying those methods to new AI-optimized data center designs.
Commercial vendors are active too. Vigilent, a DOE-supported company, demonstrated multi-site cooling energy savings—2.3 million kilowatt-hours annually across eight California state data centers. EkkoSense, another software player, reports site cooling energy reductions up to 30% via digital twin-guided optimization.
On the hardware side, the established thermal management giants—Vertiv, Schneider Electric, Johnson Controls, Stulz, Trane—are all ramping up direct-to-chip and immersion cooling offerings. Schneider recently launched AI-centric prefabricated pods and NVIDIA MGX-ready racks with Motivair in-rack liquid solutions. Specialist startups include CoolIT (direct-to-chip), Iceotope (precision liquid cooling), GRC (which secured Samsung Ventures investment in 2025), and Submer (which raised $55.5 million in 2024).
So what's Glacian's differentiation, beyond the DOE pedigree?
The company focuses on optimizing the entire cooling plant—not just airflow in the server room, but chilled water systems, economizers, and multi-objective control across energy, reliability, carbon emissions, and regulatory compliance. The platform is designed to integrate with existing Data Center Infrastructure Management (DCIM) systems and cooling vendors, vendor-agnostic rather than proprietary. Crucially, it deploys on-premises, sidestepping the security concerns that come with cloud-based control systems managing mission-critical infrastructure.
Whether that's enough to carve out market share in a competitive field remains an open question.
Policy Tightens the Screws
Regulatory pressure is intensifying across major markets, which may work in Glacian's favor—or at least in favor of any vendor that can produce auditable efficiency data.
The European Union's Energy Efficiency Directive now mandates that facilities over 500 kW of IT load report to a centralized database, with annual submissions required every May 15. The EU is developing a rating scheme and considering minimum performance standards. Ireland's Commission for Regulation of Utilities finalized new rules in December 2025 that effectively end a de facto moratorium on Dublin data centers—but impose strict requirements for on-site or near-site dispatchable generation or storage for large loads. Data centers already consume roughly 21% of Ireland's national electricity, a figure that's sparked public backlash and political debate.
Singapore launched its second Data Centre – Call for Applications in December 2025, targeting at least 200 megawatts of new capacity but requiring strong green energy commitments and resilience standards. In the U.K., regulator Ofgem warned that 140 proposed data centers could require up to 50 gigawatts—more than the country's peak grid demand of roughly 46 GW. The government is prioritizing faster grid connections for AI infrastructure, but debate over affordability and net-zero commitments is intensifying, with no clear resolution in sight.
In the U.S., interconnection queues and equipment shortages are delaying projects. Dominion Energy in Northern Virginia paused new data center connections for months due to transmission constraints, only recently resuming limited capacity as new 500-kilovolt lines target completion between 2026 and 2028. Local communities are pushing back over noise from diesel backup generators, visual impacts of transmission towers, and water consumption for cooling. Some projects have been canceled or indefinitely delayed in late 2025 and early 2026—a development that would've been unthinkable two years ago.
This regulatory and grid landscape reinforces demand for demonstrable efficiency improvements. Software platforms like Glacian's—backed by DOE research, capable of producing auditable KPI data, and deployable without major construction—align with emerging compliance requirements. Multi-objective optimization becomes less of a nice-to-have and more table stakes.
The Nuclear Hedge
While cooling optimization addresses the facility-level bottleneck, hyperscalers are simultaneously pursuing aggressive—some might say desperate—clean power procurement to feed their growing appetites.
AWS signed a 17-year power purchase agreement with Talen Energy for up to 1,920 megawatts of carbon-free nuclear power from the Susquehanna plant in Pennsylvania. Meta announced deals in January 2026 with TerraPower, Oklo, and Vistra targeting 6.6 gigawatts of nuclear capacity by 2035, largely focused on Ohio. Switch and Oklo signed a non-binding framework in December 2024 for up to 12 GW through 2044 via small modular reactors—though SMRs remain largely unproven at commercial scale.
These nuclear bets reflect a broader acknowledgment that renewable energy alone won't meet the pace and scale of AI-driven demand. Even Sam Altman of OpenAI, who's personally invested in nuclear startup Oklo, conceded in February 2026 that AI energy concerns are valid, calling for faster buildouts of nuclear, wind, and solar. IEA Executive Director Fatih Birol told a major industry conference that "understanding the AI revolution is critical to understanding the future of energy."
The irony, of course, is that even if operators secure gigawatts of clean power, inefficient cooling infrastructure can waste a third of that capacity. Optimizing what's already installed buys time and headroom while the grid catches up. Maybe.
Bridge or Bet?
The trajectory seems clear enough: data center energy demand will double by 2030, AI will account for nearly half of that growth, and facility-level constraints—power, cooling, water—will shape which operators can scale and which hit a wall. Physics-based AI optimization, liquid cooling hardware, and long-term clean energy procurement will all play roles. None is sufficient alone.
Glacian's technology represents what might be called a bridge strategy. It won't replace the need for direct-to-chip cooling at 120-kW racks, but it can maximize the efficiency of existing chilled water plants and airflow systems in the near term. It won't solve interconnection queue delays, but it can free up megawatts for compute under constrained power envelopes. And it won't eliminate the need for gigawatt-scale nuclear or renewable PPAs, but it can stretch every megawatt further—a not-insignificant advantage when utilities are saying no more often than yes.
The company's DOE-backed provenance and academic rigor give it credibility in a market increasingly wary of overhyped claims. Yet commercialization is always harder than research, a lesson countless university spinouts have learned the hard way. Glacian must prove it can deliver 50% to 70% cooling energy savings in diverse, real-world facilities—not just two DOE case studies—and do so at a price point and deployment speed that beats both doing nothing and ripping out infrastructure for full liquid cooling retrofits.
The opportunity is substantial. Omdia's thermal management market forecast, regulatory mandates for transparency, and the sheer physics of AI chip power densities all point toward sustained demand for sophisticated cooling optimization. Whether Glacian can capture a meaningful share of that $16.9 billion market by 2028 depends on execution, not just innovation.
But the research foundation is there. The timing, for once, looks right. And the industry seems willing to listen to physicists as much as hardware vendors.
Which, given the thermal realities of AI infrastructure, might be the smartest bet anyone's made in a while.
