> ## Content Index > Fetch the complete content index at: https://www.insidedeeptech.com/llms.txt > Use this file to discover other available public pages before exploring further. # Data Center Liquid Cooling: A Full Guide to Direct-to-Chip, CDUs, and Immersion for AI - URL: https://www.insidedeeptech.com/data-center-liquid-cooling-ai-full-guide/ - Published: 2026-08-17T12:14:42.000Z - Updated: 2026-08-17T12:14:42.000Z - Description: Everything you need to understand why AI racks in 2026 need water, not just air: direct-to-chip cold plates, coolant distribution units, rear-door heat exchangers, immersion, and the facility-water classes that decide whether a hall can host a 120 kW rack. - Author: Editorial Team - Tags: Energy, AI, Hardware Air cooling did not fail. The racks outgrew it. A typical enterprise cabinet still sits in the high single digits of kilowatts. An NVIDIA GB200 NVL72 is a liquid-cooled, 72-GPU rack that NVIDIA and its partners size at about 120 kW of power and heat — an order of magnitude above the air-cooled halls most of the industry spent twenty years optimizing. That is not a preference. Water carries heat roughly a thousand times more densely than air. Once you pack 72 Blackwell GPUs and 36 Grace CPUs into one NVLink domain, the alternative is either freezing the room or blowing a gale through it. Neither is a facilities plan. What changed in 2026 is that liquid cooling stopped being an HPC specialty and became the default thermal path for frontier AI. Direct-to-chip cold plates, coolant distribution units, and a facility water loop are now part of the same procurement as the silicon. Rear-door heat exchangers and immersion remain in the mix, but they solve different problems. The useful question is which loop actually removes the heat, at what temperature, and what fails when the pumps stop. ### Key takeaways - **A 120 kW AI rack is a plumbing problem wearing a GPU costume.** NVIDIA specifies the [GB200 NVL72](https://www.nvidia.com/en-us/data-center/gb200-nvl72/?ref=insidedeeptech.com) as a rack-scale, liquid-cooled design. The company's own hardware guide puts rack power at approximately 120 kW. Partner SKUs (Supermicro's SuperCluster, Schneider's published GB200 architecture) land in the 125–135 kW operating band. Air cannot move that heat at a tolerable temperature or noise level. - **Direct-to-chip is the 2026 production method. Immersion is not.** Cold plates on GPUs, CPUs, and NVLink switches capture most of the heat into a closed technology cooling loop. The rest still leaves as air. Rear-door heat exchangers exist to catch that residual. Full immersion — single-phase or two-phase — is real, deployed, and still a minority architecture for hyperscale AI. - **The CDU is the real product, not the pretty hose.** A coolant distribution unit isolates dirty, variable facility water (the FWS) from the clean, tightly specified loop that touches silicon (the TCS). ASHRAE now publishes two temperature ladders for that reason: W-classes for the building, S-classes for the IT inlet. - **Warm water is the efficiency prize. Most AI silicon still wants it cooler than the brochure.** Higher ASHRAE W- and S-classes let you skip chillers and reject heat with dry coolers or cooling towers. High-TDP GPUs often force a lower S-class in the dense zone, which brings compressors back into the design. That trade-off, not the existence of liquid, decides PUE. - **Cooling is now on the critical path with power, HBM, and lithography.** [Small modular reactors](https://www.insidedeeptech.com/small-modular-reactors-nuclear-power-ai-data-centers/) and gas turbines decide whether megawatts exist. Liquid cooling decides whether those megawatts can be turned into tokens without throttling. A campus that cannot plumb 120 kW cabinets will sit on allocated power it cannot use. - **Leaks, transients, and chemistry are the unsolved operations problems.** ASHRAE's 2024 resiliency bulletin is blunt: loss of cooling at these chip powers causes throttling first and hardware damage second. Redundant pumps, thermal inertia, fluid quality, and load-migration software are not optional extras. ## Liquid cooling at a glance | Attribute | Detail | | ------------------------ | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | What it is | A set of heat-removal methods that move waste heat from chips into a liquid loop, then to the building and the atmosphere | | What it is not | A replacement for electricity, a 2026 immersion-only industry, or a way to ignore facility water | | Core trick | Capture heat at or near the die, where the temperature difference is useful, instead of dumping it into room air | | Key hardware | Cold plates, manifolds, quick disconnects, CDUs, rear-door coils, (optionally) immersion tanks, dry coolers or cooling towers | | Transmission medium | Water, water-glycol, or a dielectric fluid in a closed loop; air still handles residual heat in most deployments | | Main applications | AI training and inference racks, HPC, any cabinet above the air-cooling ceiling of the hall | | Deployed today | Direct-to-chip on GB200/GB300-class racks at hyperscalers and neo-clouds; rear doors as retrofit and residual-heat catchers; immersion in selected HPC and colocation sites | | Biggest unsolved problem | Building, staffing, and operating a water plant that survives pump loss, chemistry drift, and 50% load steps without cooking a $3 million rack | ## What liquid cooling actually is Start with the thing it is not. It is not a magic fluid that makes GPUs free. It is not a substitute for [firm megawatts](https://www.insidedeeptech.com/small-modular-reactors-nuclear-power-ai-data-centers/). And "liquid cooled" on a product page does not tell you whether the heat left the building, or merely moved from the GPU into a CDU that still needs a chiller. Every watt that enters a data center becomes heat. Fans, CRAC units, and chillers have been the standard way to reject that heat since the mainframe era. They still work — for racks that look like 2019\. Uptime Institute's recent industry averages still sit in the high single-digit kilowatts per cabinet. NVIDIA's Blackwell-generation NVL72 does not. The company quotes a full-rack power density of 120 kW and, in its [water-efficiency note](https://blogs.nvidia.com/blog/blackwell-platform-water-efficiency-liquid-cooling-data-centers-ai-factories/?ref=insidedeeptech.com), says traditional air would need near-freezing supply air or near-gale velocities to keep up. The physics is ordinary. Liquid water has a volumetric heat capacity on the order of 4,000 kJ/m³·K. Air at room conditions is about 1.2 kJ/m³·K. NVIDIA's "nearly 1,000x" density claim is that ratio, rounded. You can move the same joules with a much smaller volume, at a higher temperature, with less fan work. That is the entire product. What the industry then argues about is *where* the liquid touches the heat. ### The problem the chips created An AI accelerator is a heater with a memory problem. The logic die, the [HBM](https://www.insidedeeptech.com/high-bandwidth-memory-hbm-full-guide/) stacks sitting a few millimeters away, and the NVLink switches all dump heat into a package that [advanced packaging](https://www.insidedeeptech.com/advanced-packaging-in-chips-explained-a-comprehensive-guide/) has made smaller and hotter at the same time. Heat flux at the die, not rack kilowatts, is why a heatsink and a raised floor run out of headroom before the busbar does. The same constraint shows up in the [inference TCO](https://www.insidedeeptech.com/the-real-cost-of-running-ai-in-2026-inference-chips-compared/) conversation. A throttled GPU is an expensive idle. Tokens per second, not nameplate TFLOPS, is what the customer bought. If the cold plate cannot hold junction temperature through an all-reduce spike, the model slows down and the megawatt-hour was wasted. This is also why cooling belongs next to power in [the 2026 infrastructure stack](https://www.insidedeeptech.com/the-state-of-ai-infrastructure-2026-compute-power-and-constraints/). Interconnection queues and transformers decide whether the campus gets electricity. Liquid loops decide whether that electricity can be packed at 120 kW per footprint. A hall with 40 MW of utility feed and 20 kW of cooling per cabinet has 40 MW of stranded power. ## How the heat actually leaves the rack A working liquid-cooled AI hall is four loops stacked on top of each other: the cold plate (or tank), the technology cooling system, the facility water system, and the heat-rejection plant outdoors. Mix them up and you get leaks, corrosion, or a chiller you did not budget. ### Direct-to-chip cold plates This is the production architecture for GB200- and GB300-class racks. A metal cold plate, usually copper, sits on the GPU, the CPU, and often the NVLink switch, with thermal interface material in between. Coolant — typically a treated water or water-glycol mix, not a dielectric — runs through microchannels in the plate. Heat conducts into the metal, then into the fluid. Manifolds at the rear of the rack feed every tray through blind-mate quick disconnects. NVIDIA's OCP contribution describes exactly that: an enhanced blind-mate manifold and a floating tray connection sized for 120 kW of rack cooling capacity. The important percentage: cold plates do not capture everything. Power supplies, NICs, DIMMs, and fans still dump heat into air. Integrators and facilities engineers commonly put the liquid fraction at 70–80% of server heat, with 20–30% left as exhaust. That residual is why hybrid racks still have fans, and why a rear-door coil is often the second product in the bill of materials. Supermicro's published GB200 NVL72 SuperCluster is a useful concrete SKU. It lists 8 × 33 kW power shelves (264 kW of PSU capacity with redundancy) against 125–135 kW operating power and 132 kW of "total power," plus a 4U in-rack CDU rated at 250 kW, a 1.3 MW in-row CDU option, and 180/240 kW liquid-to-air skids for sites with no facility water. Those are vendor ratings, not a heat-balance audit. They do tell you the shape of the product: the CDU is sized with margin; the rack is not air-cooled even as a fallback. ### Coolant distribution units The CDU is a pump, a heat exchanger, a filter, and a set of sensors in a box. Its job is political as much as thermal. Facility water is large-volume, chemically dirty, and allowed to swing. The loop that touches a $40,000 GPU is not. ASHRAE TC 9.9's liquid-cooling rewrite in the Datacom Encyclopedia is built around that split. The facility water system (FWS) is specified in W-classes. The technology cooling system (TCS) that enters the IT equipment is specified in S-classes. A CDU's approach temperature — often a few degrees, sometimes more than 10 °C on a tired plate exchanger — is the conversion between them. A W32 plant with a 5 °C approach is an S37 inlet, not a W32 inlet. Operators who skip that arithmetic buy silicon they cannot cool. The same committee's September 2024 resiliency bulletin is the operations document that belongs next to the product sheet. At these chip powers, two failures matter. Throttling: the GPU clocks down when the die overheats, and you lose the job you paid for. Rapid temperature rise: if flow stops, the thermal mass of a high-TDP package is small enough that damage, not just throttling, is on the table. The recommended mitigations are prosaic: a CDU as the FWS/TCS demarcation, more fluid volume for inertia, active pump redundancy, transient modeling, coolant quality and filtration, and software that does not step the load from 10% to 100% without warning. ### Rear-door heat exchangers A rear-door heat exchanger (RDHx) is a water coil on the back of the cabinet. Server fans push exhaust through it. The coil absorbs the heat. The air that re-enters the aisle is closer to room-neutral. Passive doors rely on the servers' own fans; active doors add their own. This is the least invasive way to put water in a hall that was designed for air. You do not open the server. You do not requalify a cold plate. You tap facility water, hang a door, and raise the rack density until the coil or the PDU runs out. Vertiv's Liebert DCD line has long been sold at 30 kW and 50 kW. In July 2026 Supermicro expanded its own RDHx catalog to ten models, 10–120 kW at the door and up to 240 kW at the rack, including as a partner to direct-to-chip rather than a replacement for it. That last use is the 2026 one that matters. On a 100 kW direct-to-chip rack, 20–30 kW of residual air is still a serious CRAC load if you let it into the room. A rear door puts that remainder into the same water plant. It is not how you cool a 1,000 W GPU by itself. It is how you keep the aisle from becoming the heat sink. ### Immersion Immersion puts the board in a dielectric fluid. Single-phase systems circulate a mineral oil or engineered fluid and dump heat through a heat exchanger; the fluid never boils. Two-phase systems boil the fluid on the components and condense it in a coil at the top of the tank. Heat transfer coefficients are excellent. Capture fractions approach 100%. Tanks replace racks. The costs are operational. You cannot service a GPU the way a hyperscale technician services a tray. Dielectric inventory, fluid loss, material compatibility, and a forklift-scale maintenance model are the reasons immersion remains a serious HPC and colocation product and a niche hyperscale one. ARPA-E's COOLERCHIPS program funded two-phase immersion work, including an Intel Federal project that was later cancelled. That is a research signal, not a 2026 SKU. Treat vendor "100% heat capture" claims as a tank specification, not as a statement that your existing hall can be filled with tanks next quarter. ## The three architectures, compared | Architecture | What the liquid touches | Typical heat capture | Where it wins | Where it loses | 2026 role | | --------------------------- | ------------------------------------------ | -------------------------------------------------- | --------------------------------------------------------- | ------------------------------------------------------------------ | -------------------------------------------------- | | Rear-door heat exchanger | Exhaust air, via a coil on the cabinet | The air leaving the rack; none of the die directly | Retrofits, residual heat, mixed air/liquid halls | Cannot, alone, hold a 120 kW NVL72 | Companion to D2C; primary only at moderate density | | Direct-to-chip (cold plate) | GPU/CPU/switch lids, through a metal plate | \~70–80% of server heat into liquid; rest is air | Production AI racks; serviceable trays; OCP/MGX ecosystem | Needs a CDU, manifolds, leak detection, residual-air plan | Default for GB200/GB300-class deployments | | Single-phase immersion | Entire board, in dielectric | Near-total, if the tank is the unit of service | Extreme density, uniform heat capture, few fans | Service model, fluid handling, tank logistics | Selective HPC / colo; not the hyperscale default | | Two-phase immersion | Entire board; fluid boils on the package | Near-total, very high heat flux | Highest flux research and some production tanks | Complexity, fluid management, cancelled as well as active programs | Still a specialist path | The honest reading of 2026 is that direct-to-chip plus a CDU plus (often) a rear door is the shipping stack. Immersion is not dead. It is not what NVIDIA contributed to the Open Compute Project. ## Facility water classes: the numbers that actually matter ASHRAE's liquid-cooling classes are how you avoid buying a chiller you do not need — or skipping one you do. **W-classes describe facility water supply temperature**, the water the building makes. The current ladder is named for its upper limit: W17 (17 °C), W27, W32, W40, W45, and W+ above 45 °C. Lower classes typically imply chillers and cooling towers. W32 and W40 are the "often chiller-free" band in many climates. W45 and W+ are the heat-reuse and dry-cooler band. Minimum water temperature for all W-classes is 2 °C. **S-classes describe TCS supply temperature at the IT equipment**, the water the cold plate actually sees. Published guidance puts common S-classes in a 30–50 °C band (roughly S30 through S50), with S20 and S25 later added for silicon that cannot tolerate the warm-water story. Minimum TCS temperature is specified relative to room dew point, because a cold manifold in a humid aisle is a condensation machine. The efficiency claim everyone wants — skip the compressor, reject heat with a dry cooler, maybe sell the waste heat to a district loop — only holds if the S-class is high enough. High-TDP GPUs push the other way. ASHRAE's own resiliency note says chip-power trends may require a *lower* S-class in the dense zone. That is the opposite of the marketing slide. A 120 kW row that needs S25 inlet in Phoenix in August still has a chiller. Liquid cooling made the chiller smaller. It did not delete thermodynamics. The U.S. Department of Energy's 2024 [Best Practices Guide for Energy-Efficient Data Center Design](https://www.energy.gov/sites/default/files/2024-07/best-practice-guide-data-center-design.pdf?ref=insidedeeptech.com) is aligned with that hierarchy: capture heat in liquid, raise temperatures, reuse what you can (energy reuse effectiveness), then reject the rest dry to save water (water usage effectiveness). NREL's own HPC facility is the existence proof the authors cite — on the order of 6% of energy spent on cooling equipment versus a much fatter fraction in a typical air hall. That is a well-run, purpose-built plant, not a converted 2012 colocation cage. ## Who is actually shipping the plant The silicon vendors set the thermal envelope. The cooling vendors fill it. NVIDIA made liquid a rack feature, not an accessory, by shipping NVL72 as a liquid-cooled domain and contributing the manifold and tray designs to OCP in 2024\. The [OCP write-up](https://developer.nvidia.com/blog/nvidia-contributes-nvidia-gb200-nvl72-designs-to-open-compute-project/?ref=insidedeeptech.com) is specific: 120 kW of cooling capacity, blind-mate manifolds, floating tray connections, a deeper 1,400 A busbar, and 1RU liquid-cooled compute and switch trays. GB300 NVL72 continues the same fully liquid-cooled rack idea for Blackwell Ultra. Those are architecture decisions. They force the rest of the bill of materials. The CDU and cold-plate names that show up next to those racks are a short list repeating: Vertiv (including a joint 7 MW GB200 reference architecture NVIDIA says can cut implementation time by up to 50% — a vendor figure), CoolIT (CHx-class liquid-to-liquid CDUs; the CHx2000 is marketed at 2 MW and a 5 °C approach for GB300-class rows), Boyd, nVent/Stulz-class rear doors, Schneider Electric (publicly associated with 132 kW/rack GB200 infrastructure), Motivair, Cooler Master, and a long tail of manifold, quick-disconnect, and cold-plate specialists NVIDIA itself enumerated at OCP. System integrators — Supermicro, Dell, HPE, GIGABYTE, QCT, Wiwynn — ship the tray. None of that is a commodity catalog the way a 5 kW CRAC is a commodity catalog. Lead times on CDUs, CDUs' own heat exchangers, and qualified quick disconnects are now part of cluster scheduling, the way CoWoS slots and [EUV](https://www.insidedeeptech.com/euv-lithography-asml-machines-full-guide/) scanner time are part of chip scheduling. ## The bottlenecks that will slip the dates ### The hall was not poured for this A 120 kW, 1.3-tonne liquid-cooled rack wants floor loading, overhead or underfloor piping, a leak-detection grid, a different PDU, and a different maintenance aisle. Uptime Institute has long reported that a majority of older enterprise rooms cannot host this density. That is not a surprise. It is why the 2026 buildout is new halls, retrofitted "AI zones," and factory-built modules, not a software update to 2015 raised floors. ### Chemistry and leaks are availability problems A glycol loop that grows bacteria, a filter that clogs, or a drip on a busbar will take down more than one GPU. ASHRAE's operational list — fluid quality, filtration, hydraulic and thermal commissioning — is the unglamorous half of "liquid cooling." Sites that treat the CDU as a set-and-forget appliance will learn this the expensive way. ### Transients are nastier than steady state AI training and inference already swing load by large fractions of rated power in about a second. The IEA has flagged that as a grid problem. It is also a cooling problem. A loop with too little thermal inertia, or a redundant pump that takes too long to spin, sees a temperature spike the cold plate cannot hide. The resiliency bulletin's insistence on transient modeling is there because empirical data from 30 kW racks does not transfer. ### Water use is a political number, not just an engineering one Evaporative cooling towers are efficient and thirsty. Dry coolers are the reverse. NVIDIA's Blackwell water-efficiency claims (including a 300× figure versus traditional air-cooled architectures) are company-stated comparisons, not a metered WUE from a named campus. They point in a real direction: warmer liquid loops enable dry rejection. They do not license a press release that says AI factories use no water. Local permits will keep asking. ### Research cooling is not production cooling ARPA-E's [COOLERCHIPS](https://arpa-e.energy.gov/programs-and-initiatives/view-all-programs/coolerchips?ref=insidedeeptech.com) program is the right long-term benchmark and the wrong 2026 shopping list. The target is cooling energy below 5% of IT load, anywhere in the United States, at >80 kW/m³ (about >3 kW per server, or >126 kW in a 42U rack), with chip-to-coolant differences under 10 °C. DOE put about $40 million into the first cohort. Some projects are active (HP microfluidic cooling, HRL graphite microchannels). Some were cancelled. Treat the program as the efficiency ceiling the industry is aiming at, not as a catalog of parts you can order for a Q4 cluster. ## A realistic timeline | Horizon | What is actually available | Confidence | | ----------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------- | | Now (2026) | Direct-to-chip NVL72/NVL36 racks with in-rack or in-row CDUs; rear doors as retrofit and residual-heat catchers; immersion in selected HPC/colo | Deployed at hyperscalers and specialists; still scarce as a turnkey colo SKU | | Near term (2027–2028) | GB300-class and successor racks at still-higher TDP; more hybrid D2C+RDHx; warmer S-classes only where silicon allows | High that density rises; medium that chiller-free designs win in hot climates | | Medium term (2028–2031) | COOLERCHIPS-class components if they clear reliability; more heat reuse in Europe and some U.S. districts; possible two-phase cold plates | Plausible, vendor- and permit-limited | | Long term (2030s) | Cooling energy in the low single-digit percent of IT load as a normal design point, not a lab demo | Speculative on schedule; directionally consistent with DOE targets | One useful heuristic: ask for the S-class at the IT inlet, the CDU approach temperature, the residual air fraction, and the time-to-throttle on a pump loss. Do not ask whether the rack is "liquid cooled." Everything serious is. ## The bottom line 2026 is the year air stopped being the default answer for frontier AI cabinets. The physics is old. Water moves heat better than air. The new work is industrial: a cold plate on every high-TDP package, a CDU that keeps facility water away from the die, a rear door or a remaining CRAC plant for the leftover 20%, and a heat-rejection scheme that does not bankrupt the water permit. NVIDIA's 120 kW NVL72 made that stack mandatory for anyone buying the leading rack. It did not make the stack simple. For anyone making decisions today, the split is clean. If the question is how to land a GB200- or GB300-class cluster this year, you are buying direct-to-chip, a CDU, and a facility loop — and you should budget a residual-air plan. If the question is whether a vendor's cooling story is real, ask for the W-class and S-class, the capture fraction, the redundant-pump transient, and a hall that can actually hold the weight and the pipe. Immersion remains a legitimate architecture. It is not the one the 2026 AI buildout is standardizing on. Power still has to show up. Liquid cooling is how you spend it. ## Frequently asked questions ### What is data center liquid cooling? Any design that moves a server's waste heat into a liquid — water, water-glycol, or a dielectric — instead of relying only on room air and computer-room air handlers. In 2026 AI practice that usually means cold plates on the chips, a coolant distribution unit, and a facility water loop that rejects heat outdoors. ### Why can't AI racks still use air? They can, at lower density. An 8-GPU air-cooled HGX box in a 40–60 kW cabinet is still a product. A 72-GPU NVLink domain at \~120 kW is not an air product. The volumetric heat capacity of air is too low, so you would need extreme flow or extreme supply temperatures. NVIDIA's published NVL72 architecture is liquid-cooled by design. ### What is a CDU? A coolant distribution unit. It pumps and conditions the technology cooling loop, exchanges heat with facility water through a plate heat exchanger, and keeps the two chemistries apart. It is the device ASHRAE wants as the demarcation between the building and the IT. ### Direct-to-chip vs immersion: which is better? Different products. Direct-to-chip keeps a familiar tray-and-rack service model and is what GB200/GB300 racks ship with. Immersion captures more of the heat and handles higher flux, at the cost of tanks, dielectric fluid, and a slower maintenance model. Hyperscale AI in 2026 is standardizing on direct-to-chip. Immersion remains important in HPC and some colocation. ### Do I still need air cooling if I have liquid? In a direct-to-chip hall, yes, for the 20–30% of heat that never enters the cold plate — unless a rear-door exchanger or a fully enclosed chimney takes that remainder. Immersion tanks are the exception, and they still reject heat through a water or refrigerant plant. ### What are ASHRAE W-classes and S-classes? W-classes (W17, W27, W32, W40, W45, W+) specify facility water supply temperature. S-classes specify the coolant temperature that enters the IT equipment. The CDU's approach temperature sits between them. Warm classes enable chiller-free and heat-reuse designs; high-TDP GPUs often force a cooler S-class in the AI zone. ### Does liquid cooling save water? It can, if the warmer loop lets you reject heat with dry coolers instead of evaporative towers. It can also use more water if you keep the towers and raise density. NVIDIA's 300× water-efficiency comparison for Blackwell versus traditional air architectures is a company-stated figure. Metered WUE at a named site is the number that matters for a permit. ### Will cooling constrain AI the way power does? It already does, locally. A campus can have utility megawatts and still be unable to host 120 kW cabinets without a water plant, a floor, and a CDU fleet. Globally, electricity generation and interconnection remain the larger ceiling. Cooling is how that electricity becomes usable density. ## Methodology This guide is based on primary vendor, standards, and agency disclosures available as of 17 August 2026: NVIDIA's GB200 NVL72 product page (rack-scale liquid-cooled design, 36 Grace CPUs + 72 Blackwell GPUs); NVIDIA's DGX GB200 hardware documentation (approximately 120 kW rack power; 8 × 33 kW power shelves); NVIDIA's October 2024 Open Compute Project contribution (120 kW cooling capacity, blind-mate manifolds); NVIDIA's Blackwell water-efficiency blog (120 kW density; historical cooling share up to 40% of electricity; 300× water-efficiency and 25× energy-efficiency claims, labeled here as company-stated); ASHRAE TC 9.9 Datacom Encyclopedia / liquid-cooling class notes (W17–W45/W+ facility water; S-classes for TCS inlet, including later S20/S25); ASHRAE's September 2024 liquid-cooling resiliency bulletin (CDU demarcation, thermal inertia, throttling vs. damage); the DOE/FEMP 2024 Best Practices Guide for Energy-Efficient Data Center Design; and ARPA-E COOLERCHIPS program materials (<5% cooling energy / IT load, >80 kW/m³, \~$40 million first cohort). Supermicro GB200 SuperCluster power and CDU ratings (125–135 kW operating, 250 kW in-rack CDU, 1.3 MW in-row) and partner figures from Vertiv, Schneider, and CoolIT are vendor-stated and are not independently heat-balanced here. Direct-to-chip capture fractions of \~70–80% are the commonly cited engineering range used by integrators (including nVent hybrid designs and Cisco/Vertiv residual-heat notes), not a single ASHRAE-mandated constant.