There's a moment in every infrastructure buildout when the physics stop cooperating. For data centers, that moment arrived sometime around 2023, when NVIDIA's latest chips started pulling so much power that the decades-old playbook—rows of servers, walls of fans, frigid air pumped in by the ton—simply broke down.
The numbers tell the story, though perhaps not the whole one. By 2030, according to the International Energy Agency, data centers worldwide will devour roughly 945 terawatt-hours of electricity annually. That's more than double what they consume today. AI-optimized facilities? They're on track for a fourfold jump. In the United States alone, data centers are projected to account for half of all new electricity demand growth through decade's end—a statistic that's already rattling utility executives and state regulators from Virginia to Oregon.
And the hardware roadmap? It shows no mercy. NVIDIA's Blackwell B200 GPU already draws around 1,000 watts. The company's future Rubin-era chips are expected to hit 1,800 watts per socket. Air cooling—the workhorse that kept server farms humming for thirty years—is running out of headroom. Fast.
Which brings us to an unlikely savior: two-phase, direct-to-chip cooling. It's a waterless, dielectric-based system that some vendors claim can slash data center cooling energy by 40 to 50 percent while supporting rack densities that seemed like science fiction just two years ago. Companies like Accelsius, a 2022 startup that licensed its core technology from Nokia Bell Labs, are now demonstrating platforms that handle 250 kilowatts per rack using refrigerants that boil at the chip surface and condense in a closed loop. No water in the white space. No massive air handlers. Just heat, vapor, and thermodynamics.
Meta unveiled a 140-kilowatt liquid-cooled AI rack at the Open Compute Project Summit in 2024. Google has been running more than 2,000 liquid-cooled TPU pods for years, claiming 99.999 percent uptime. The question for the rest of the industry isn't whether liquid cooling will dominate AI infrastructure. It's how quickly operators can retool—and which approach will win.
When Air Stops Working
Traditional data centers typically burn 20 to 40 percent of their total power budget just keeping the place cool: chillers, computer room air handlers, pumps, and fans spinning endlessly to hold server inlet temperatures within guidelines set by ASHRAE, the industry's standards body. That model worked fine when racks drew 5 to 10 kilowatts.
Today's AI clusters are a different beast entirely. Hyperscale deployments are pushing 120 to 140 kilowatts per rack. Research targets at OCP and Google now contemplate 600-kilowatt racks—even one-megawatt configurations. At those densities, air simply can't carry enough heat away. Doesn't matter how many fans you spin.
The market is scrambling to catch up. Dell'Oro Group projects the data center liquid cooling sector will approach $7 billion by 2029, with direct liquid cooling revenues roughly doubling in 2025 to around $3 billion and quintupling by decade's end. TrendForce estimates liquid cooling penetration hit 10 percent in 2024 and will exceed 20 percent this year, driven largely by Blackwell GPU shipments. An Uptime Institute survey from 2024 found that 22 percent of data center operators already use some form of direct liquid cooling, with another 61 percent actively considering it. Most of those deployments remain small pilots, but the trajectory is unmistakable.
Colocation giants, meanwhile, are racing to make liquid cooling a standard offering. Equinix announced plans to enable direct-to-chip cooling in more than 100 facilities across 45-plus metros. Digital Realty launched direct liquid cooling and rear-door heat exchanger services across 170 sites, supporting up to 150 kilowatts per rack. The infrastructure is being built because the demand is already here—enterprise AI workloads that can't afford to wait.
The Physics Problem
Air has a low specific heat capacity and limited mass flow through server chassis. Translation: you need enormous volumes of it—and substantial fan power—to remove high chip heat loads. Single-phase water cooling, where liquid stays liquid throughout the loop, improves thermal conductivity by a factor of roughly 1,000 over air. But even that approach struggles as GPU thermal design power climbs past 1,000 watts. Cold plates designed for water need low inlet temperatures and face thermal resistance limits that become problematic at the heat fluxes modern accelerators generate.
NVIDIA's own technical notes on the Blackwell GB200 NVL72 architecture suggest that liquid-cooling designs can cut annual energy consumption by around 25 percent in reference builds, with 300 times better water efficiency in certain configurations compared to evaporative cooling. That efficiency gain matters not just for operating costs but for regulatory survival.
Ireland's data centers already consume 20 to 24 percent of the country's electricity—a figure that prompted a moratorium on new power connections, only recently evolved into a renewable-energy mandate framework. Amsterdam paused new data center construction entirely, with a re-evaluation slated for 2030. Northern Virginia's "Data Center Alley" faces grid pressures, land constraints, and community pushback. Operators in those markets need every efficiency lever they can find, or they risk being shut out entirely.
Air cooling locks you into colder facility water temperatures—typically 7 to 12 degrees Celsius at the chiller—to maintain acceptable server inlet temps. That chiller work consumes substantial energy, especially in warm climates. Warm-water liquid cooling can operate with facility water at 35 to 45 degrees Celsius, dramatically reducing or eliminating mechanical cooling loads during much of the year. Some studies suggest each one-degree Celsius increase in facility water temperature yields a 4 to 5 percent reduction in cooling energy. Compounded over a year, that's real money.
The Frontier exascale supercomputer at Oak Ridge National Laboratory, which uses warm-water liquid cooling, achieves a power usage effectiveness ratio of approximately 1.03. That's a benchmark nearly impossible to hit with conventional air systems at high densities. (For context, a PUE of 1.0 would mean zero overhead; everything goes to computation. Most air-cooled facilities run closer to 1.5 or 1.6.)
The Vapor Play
Two-phase cooling leverages the latent heat of vaporization—the large amount of energy absorbed when a liquid turns to vapor. Instead of heating water or a single-phase coolant by a few degrees as it flows through a cold plate, a two-phase system uses a refrigerant engineered to boil right at the chip surface.
In Accelsius's NeuCool platform, this happens in what the company calls a "vaporator"—essentially a cold plate where the dielectric refrigerant absorbs heat and transitions to vapor. That vapor travels through low-pressure lines to a condenser (in the company's in-rack IR80 system or row-based MR250 coolant distribution unit), where it gives up the heat and condenses back to liquid. The cycle repeats in a closed loop, with no water in the white space and minimal pumping power because the refrigerant moves at low pressure.
The result: very high heat-removal capacity per unit of fluid flow. Accelsius claims its system supports more than 4,500 watts per socket—a figure the company describes as the highest in direct-to-chip cooling. Thermal resistance at the cold plate runs around 0.020 degrees Celsius per watt at 700-watt thermal design power, the level of NVIDIA's H100 GPU. The IR80 in-rack unit delivers 80 kilowatts of direct-to-chip cooling per rack; the newer MR250 row-based CDU can handle 250 kilowatts per rack in configurations supporting one 250-kilowatt rack or two 125-kilowatt racks.
At the OCP Global Summit in San Jose this past October, Accelsius demonstrated a full 250-kilowatt AI rack being cooled, with GPU temperatures held below throttle thresholds even when facility water was running at 40 degrees Celsius—a temperature that would cripple most air-cooled systems.
The refrigerants matter, too. Accelsius uses dielectrics like R1233zd(E), an HCFO with a global warming potential near 1 to 4, zero ozone depletion potential, and A1 (non-flammable) safety classification. Those characteristics position the technology favorably against tightening European Union F-gas regulations and the broader scrutiny around PFAS—per- and polyfluoroalkyl substances—which has hit immersion cooling fluids hard after 3M announced it would exit PFAS manufacturing by the end of 2025. Low-GWP, non-flammable refrigerants like R1233zd(E) and R-515B align with sustainability mandates while sidestepping some of the supply-chain disruption facing legacy immersion chemistries.
The Energy Calculus
Accelsius's headline claim—40 to 50 percent savings in data center cooling energy versus conventional air cooling, with zero water consumption—rests on several mechanisms. First, by removing most of the heat directly at the chip, the system reduces the load on facility-level cooling infrastructure. Second, the ability to operate with facility water 6 to 8 degrees Celsius warmer than other liquid technologies translates to more than 25 percent cooling-energy savings in those facility systems alone, according to the company's data. Third, the closed-loop dielectric design eliminates evaporative water consumption entirely—a non-trivial advantage in water-stressed regions or under zero-water mandates.
Accelsius also cites 35 percent operational expenditure savings compared to single-phase direct-to-chip systems and 8 to 17 percent total cost of ownership savings over a five-year horizon in reference designs developed with engineering firm Jacobs. That study, unveiled in 2024, suggested two-phase cooling could deliver comparable capital expenditure to single-phase systems while offering substantially lower OpEx due to reduced pumping power (the system uses less than one kilowatt of system power for an 80-kilowatt rack) and warmer-water operation. At scale, the company projects potential billions in cumulative energy savings across the industry—though those figures are extrapolations based on adoption assumptions, not hard contracts.
Worth noting: most of these savings figures are vendor-reported, derived from lab tests, or based on study assumptions rather than peer-reviewed, independent measurement and verification in production customer sites over extended periods. The Jacobs reference design is a modeling exercise. Accelsius's deployments at sites like Telehouse London's Liquid Cooling Lab, Equinix's Co-Innovation Facility in Ashburn, Park Place Technologies in Cleveland, and iM Data Centers in Miami remain largely demonstration or pilot scale as of early 2025.
Operators evaluating the technology should probably request site-specific energy audits and performance data to validate claimed savings in their own climate, architecture, and workload contexts. Then again, the warm-water operational principle is well-established elsewhere—Oak Ridge's Frontier achieves that 1.03 PUE, and immersion cooling vendors like LiquidStack report partial PUE figures of 1.02 to 1.03 and energy reductions of roughly 41 percent versus air. The physics check out. The question is how much of that theoretical advantage translates to real savings after accounting for integration complexity, redundancy requirements, and the inevitable learning curves.
Proof Points and Partnerships
Accelsius's commercialization strategy has centered on high-visibility labs and partnerships to build credibility. At Telehouse London's Liquid Cooling Lab, the company deployed its thermal simulation rack—a configurable load bank that can simulate 30 to 50 kilowatts standard, up to 250 kilowatts—to demonstrate NeuCool performance alongside other vendors' technologies. The multi-vendor lab also features JetCool's microjet direct-to-chip system (single-phase), Legrand rear-door heat exchangers, and EkkoSense analytics, giving operators a side-by-side comparison. Similar units have been placed at Park Place Technologies in Cleveland and in Global Switch's showroom in London, with expansion underway in the European Union and Canada.
The Equinix Co-Innovation Facility in Ashburn, Virginia—a key testing ground for colocation and enterprise infrastructure—hosts Accelsius alongside Flex/JetCool and other liquid cooling providers. Equinix's broader commitment to enable direct-to-chip cooling in more than 100 data centers across 45-plus metros signals that colos see liquid as a market necessity, not a niche. Digital Realty's rollout of direct liquid cooling and rear-door heat exchanger options across 170 sites reinforces the trend. These aren't speculative bets—they're infrastructure investments driven by customer demand for GPU clusters that simply can't be deployed with air cooling.
In October 2025, Johnson Controls announced a strategic investment in Accelsius, aiming to scale the technology across its global data center infrastructure channels. Johnson Controls brings deep expertise in building mechanical systems and a vast installed base of chiller and controls infrastructure, which could accelerate adoption. Separately, Accelsius secured a $24 million Series A in early 2024, joined NVIDIA's Inception program, and was selected as a contributor to the U.S. Department of Energy's ARPA-E COOLERCHIPS project led by the University of Texas at Arlington, which focuses on hybrid cooling for next-generation data centers. The Innventure parent company, which commercialized the Bell Labs-origin technology, reported a sales pipeline exceeding $1 billion in indications in 2024 SEC filings—a figure that reflects interest, not bookings, but suggests genuine market traction.
Real-world production data is still emerging, though. The iM Data Centers deployment in Miami, announced as a collaboration for HPC and AI workloads in a hot, humid environment, is positioned as a path to production following thermal simulation rack validation. But as of early 2025, most public case studies remain in pilot or demonstration phases. That's not unusual for infrastructure this new, but it does mean the industry is still collecting long-term reliability, operational complexity, and true TCO data. Buyers, understandably, want to see more.
A Crowded Field
Accelsius is far from alone. Single-phase direct-to-chip cooling—where water or a water-glycol mix circulates through cold plates—dominates near-term deployments. CoolIT Systems, acquired by private equity firm KKR in 2023 for $270 million, is a market leader. Vertiv offers its CoolPhase product line and a broad CDU portfolio. Schneider Electric acquired Motivair in 2025 and now markets an integrated liquid cooling suite. Asetek and Lenovo's Neptune platform provide OEM-integrated options. These systems are well-understood, widely deployed, and supported by mature supply chains.
Two-phase direct-to-chip remains a smaller segment but is growing. ZutaCore's HyperCool platform, also waterless and dielectric-based, claims 35 percent-plus power use reduction and supports 1,500 to 2,800 watts per socket. JetCool, now a Flex company, uses microjet single-phase technology—spraying coolant directly onto chips—and reports 15 percent IT power savings in demonstrations, with CDUs rated up to 300 kilowatts per rack and 4-kilowatt-per-socket support. At the Equinix Co-Innovation Facility, Flex demonstrated up to 50 percent cooling-power savings and 15 percent IT power savings in a rack-level deployment using JetCool's system.
Immersion cooling—where servers are submerged in dielectric fluid—offers yet another path. LiquidStack's two-phase immersion systems report 41 percent energy reduction versus air and partial PUE around 1.02 to 1.03. Submer, GRC, Iceotope, and Vertiv all offer immersion variants. Immersion simplifies some thermal management challenges (no cold plates to maintain), but it complicates serviceability. Pulling components out of a tank is messier than swapping a GPU in a rack. And the approach faces supply-chain headwinds from PFAS restrictions and 3M's exit from fluorinated heat-transfer fluids.
The competitive dynamic is less about outright winners and more about use-case fit. Single-phase direct-to-chip is pragmatic for organizations upgrading existing infrastructure incrementally and wanting to preserve some air-cooled flexibility. Two-phase direct-to-chip appeals to operators targeting the highest densities and lowest cooling energy, willing to adopt newer technology. Immersion suits dense, homogeneous deployments—mining operations, HPC clusters—where serviceability trade-offs are acceptable. Hybrid approaches are common, too: direct-to-chip for GPUs and CPUs, rear-door heat exchangers for the rest of the rack, particularly in retrofit scenarios.
Dell'Oro Group notes that single-phase direct-to-chip will dominate near-term growth, but two-phase adoption will accelerate as chip thermal design power and heat flux climb. That timeline maps neatly to the NVIDIA roadmap: Blackwell at ~1,000 watts is manageable with single-phase. Rubin at 1,800 watts starts to strain those systems. Whatever comes after that may require two-phase as table stakes. Accelsius's positioning at the high end—4,500 watts per socket, 250 kilowatts per rack—targets that future explicitly.
Regulation as Catalyst
Regulation is quietly reshaping the liquid cooling landscape. The European Union's revised F-gas regulation, which took effect in 2024, phases down HFC consumption to zero by 2050 with escalating quotas and market bans on high-GWP refrigerants. That's accelerating adoption of low-GWP alternatives like HFOs, HCFOs, ammonia, and CO2 in chillers and cooling systems. Data center operators in the EU increasingly favor refrigerants like R1233zd(E) and R-515B—which Accelsius uses—because they satisfy both GWP targets and non-flammability safety requirements.
PFAS scrutiny adds pressure from another angle. The broad category of per- and polyfluoroalkyl substances includes many fluorinated heat-transfer fluids historically used in immersion cooling. 3M's announcement that it would exit PFAS manufacturing by December 31, 2025—including its Novec and Fluorinert product lines—has sent immersion cooling vendors scrambling for alternative chemistries and suppliers. While Accelsius's two-phase direct-to-chip systems use HCFO/HFO refrigerants (not PFAS-classified in the same regulatory way), the broader environment around fluorinated compounds creates supply-chain uncertainty that operators are factoring into technology selection. Vendors that can demonstrate low-GWP, non-PFAS, and non-flammable profiles have a clearer path through evolving compliance landscapes.
Water constraints may be the most immediate regulatory driver. Data center moratoria or stricter permitting standards in Ireland, Amsterdam, and Singapore (which mandates PUE of 1.3 or better for new facilities) steer designs toward higher efficiency and waterless white-space operation. Microsoft announced in August 2024 that it would pilot "zero-water" evaporative alternatives and transition new designs fleetwide. Accelsius's zero-water claim—closed-loop dielectric with no evaporative cooling in the white space—aligns with that shift. For operators building in water-stressed regions or facing political pressure around water use, eliminating evaporative cooling isn't just an efficiency play. It's a permitting advantage.
ASHRAE has responded to the liquid cooling surge with updated guidance. Technical Committee 9.9 published new liquid cooling templates and intensified focus on CDU design, redundancy, transient modeling, and dew point management. CIBSE issued similar directives in its 2025 liquid cooling CPD modules. These standards help de-risk adoption by codifying best practices, but they also raise the bar for vendors—operators now expect rigorous documentation, not marketing brochures.
The Billion-Dollar Question
So where does this all land? Two-phase cooling solves real problems: it handles absurd power densities, slashes cooling energy, eliminates water consumption, and positions operators favorably against tightening regulations. But it's also newer, less proven at scale, and requires operators to trust physics they may not fully understand. Serviceability concerns linger. What happens when a vaporator fails in the middle of a training run?
The industry is placing its bets anyway. Accelsius's $1 billion-plus sales pipeline
