NVIDIA specs every MGX rack (GB200, GB300, Vera Rubin) for 45°C inlet water. Warm-water cooling is a decade old; the default on 200 kW-class racks is what's new.
NVIDIA has put a number on the future of data center cooling: "All MGX racks are universally designed to operate with 45°C (113°F) warm-water inlet temperatures." One sentence, three product generations. The line comes from the engineering post describing the Vera Rubin POD reference architecture, and because MGX is the modular architecture underneath NVIDIA's rack-scale NVL systems, the spec covers GB200 NVL72 racks running at roughly 120 kilowatts, GB300 NVL72 racks that OEMs rate between 132 and 155 kilowatts depending on configuration, and Vera Rubin NVL72 systems expected to move into the 200-kilowatt class.
The temperature is not the news. ASHRAE's thermal guidelines have accommodated warm water since 2011, and European supercomputing centers were running hot-water loops on the same principle more than a decade ago. The news is the default: the vendor whose racks anchor the AI buildout has made 45°C the universal design point for its entire rack-scale line, and the facility ecosystem underneath now has to engineer to that number whether it was ready to or not.
The 45°C figure describes coolant at the rack inlet, delivered by coolant distribution units. NVIDIA's facility schematic in the same post shows the building's water loop feeding those CDUs at roughly 41°C; the CDUs then supply 45°C fluid to the racks. Four degrees sounds like a rounding error. It matters because ASHRAE's W-classes are defined at the facility water supply: an operator who reads "45°C" as a facility-loop temperature will size heat rejection against the wrong number.
Downstream of the inlet, the loop runs hot. In NVIDIA's June engineering post on liquid-cooled AI factories, coolant entering a fully liquid-cooled chip at 45°C exits at roughly 55°C. SemiAnalysis has modeled return temperatures approaching 65°C in some Vera Rubin configurations; NVIDIA's published figure is the lower one. Rubin is also where the design commits completely, with warm-water, single-phase direct liquid cooling and trays and networking that NVIDIA describes as fully liquid-cooled; GB200 and GB300 remain hybrid designs, shedding roughly 10 to 15 percent of their heat to air depending on configuration and whether the split is measured at the tray or the rack. Density is doing the forcing here: past the 100-kilowatt class, air cooling alone stopped being viable, and these racks left that line behind a generation ago.
ASHRAE's liquid-cooling classes tell the backstory in shorthand. The W45 class covers facility water supplied at up to 45°C, the tier at which chillers become largely unnecessary and waste heat runs hot enough to feed district heating. LRZ's SuperMUC was running IBM hot-water cooling with coolant up to 45°C in 2012; its successor, SuperMUC-NG, runs water up to 50°C and heats LRZ's offices with the waste. The temperature NVIDIA just standardized has been operating in Bavaria for fourteen years. CoolIT Systems, which builds cold plates for these same racks, put it plainly in January: "Operating servers with warm water is not new. ASHRAE standards have supported elevated water temperatures since 2011... NVIDIA's announcement simply validates this approach."
The market being pulled sits much colder than the spec. Most Blackwell deployments today design for 20-to-30°C facility water, per SemiAnalysis, and mainstream direct-liquid-cooling designs had been converging around 32°C, with Meta's OCP AI rack work targeting about 30°C. Against that baseline, the MGX spec raises the mainstream design point roughly 15 degrees in a single product generation. Dry-cooler sizing, CDU capacity, pipe diameters, pump curves, and the broader facility-engineering agenda that ORNL's new data center institute was chartered to work through: all of it now gets checked against a 45°C reference.
The point of running hot is what it does to the cooling plant. At 45°C inlet, NVIDIA says, data centers in many climates can reject heat with closed-loop dry coolers and ambient air; the Vera Rubin POD design includes isolation valves that bypass the chiller plant entirely, "running only pumps and fans." Compressors are the hungriest equipment in a conventional cooling chain. Taking them out of daily service drops PUE, the ratio of total facility power to the power that reaches the computers, and hands the savings back to the IT load.
That trade lands in a market where power rather than silicon is the scarce input; the IEA projects global data center electricity consumption roughly doubling to 945 terawatt-hours by 2030, the constraint behind the 100-gigawatt data center supercycle. NVIDIA's own modeling, with no measured PUE yet published from a production 45°C facility, says the recovered energy is "enough to allocate up to 10% additional Vera Rubin NVL72 racks... in the same power budget." But the direction follows from the physics, and the arithmetic matters just as much to a national lab or a sovereign AI program working inside a fixed power envelope, where cooling efficiency decides how much computing a megawatt actually buys.
The water numbers carry the same label. NVIDIA's June post puts evaporative cooling-tower consumption at 2.6 million gallons per megawatt per year and projects "near zero" for closed warm-water loops, up to a 100 percent reduction. Its headline economics, $4 million in annual savings on a 50-megawatt facility and "25x energy efficiency, 300x water efficiency" versus air cooling, are standing marketing figures that have appeared unchanged since an April 2025 Blackwell post. No independent measurements back them yet.
Chiller-free is where the marketing runs ahead of the thermodynamics, and NVIDIA's own cooling director supplies the caveat. In the same June post, Ali Heydari, the company's director of data center cooling, qualified the claim: dry coolers can carry the load "outside of maybe 1% of the year when we might need chillers in some climates."
The physics behind the qualifier is delta-T. A dry cooler's capacity scales with the gap between the water temperature and the ambient air, as CoolIT's January analysis lays out, and with 45°C water on a 35°C afternoon the system has ten degrees to work with. Capacity keeps falling as the air warms. One percent of a year is about 88 hours, and those hours arrive on the hottest afternoons, exactly when a dry cooler is weakest. So in most designs the mechanical plant still gets built, retained and sized for worst-case conditions; what the spec changes is how many hours a year it runs.
There is a retrofit asymmetry here too, and this reading is analysis on our part rather than an NVIDIA claim. An existing facility with a 20-to-30°C chilled-water plant can serve these racks without drama, but it only collects the efficiency dividend if it rebuilds heat rejection around dry coolers. The economics of the spec favor greenfield AI factories designed around it from the slab up.
For water-stressed regions the calculus shifts. A closed warm-water loop rejecting heat through dry coolers consumes almost no water, which makes tower-free AI factories a live design option in places where evaporative cooling is physically and politically expensive. Equinix has already committed to avoiding evaporative cooling in high-water-stress locations. For Gulf states pairing abundant power with scarce water and large sovereign compute ambitions, a design that spends somewhat more electricity on the hottest days and saves nearly all the water reads, on our analysis, as close to purpose-built.
The delta-T arithmetic does not soften for anyone, though. Summer ambients of 40°C and above erase the gradient a dry cooler needs, so year-round chiller-free operation is off the table in the Gulf; mechanical cooling stays in those designs as a seasonal workhorse rather than an 88-hour insurance policy. What the spec changes there is the water bill, and for a sovereign builder that may be the number that matters.
None of this required new physics. It required a vendor with enough market weight to settle a fifteen-year argument, and that is what the March post amounts to: dry-cooler manufacturers, CDU vendors, colocation operators, and competing accelerator programs will now spend years being measured against a 45°C reference design on 200-kilowatt-class racks. The facility engineers who spent a decade advocating warm water finally have the number written into the product line, and the rest of the industry has to build to it.
