FIELD NOTE · SUPPLEMENTAL SET
The facility water gets colder: W45 today, W27 eventually
ISSUEDJULY 11, 2026DRAWN BY THE NIGHTLY SWEEP
ASHRAE renamed its liquid cooling water classes so the number is the maximum facility water supply temperature in Celsius, and its own committee expects that number to fall. W45 is a 45C ceiling on supply water, W32 is 32C, W27 is 27C, and all six classes, W17 through W+, carry the same 2C lower limit. The direction of travel is the finding.
TC 9.9's 2021 white paper on liquid cooling frames it as a replacement cycle rather than a forecast. Direct liquid cooled products that today "might have the ability to operate on W45 (W4)" get replaced by "newer, higher-powered products only capable of W32 (W3) operation, and eventually cooling requirements may drive facility water temperatures down into the W27 (W2) range. Facilities should plan accordingly."
ASHRAE is careful about which part of that it stands behind. A page earlier the same section says the exact numbers are somewhat speculative, because IT manufacturers have detail on only one to two future chip generations. The direction is the committee's position; the three temperatures are its best estimate.
So the free cooling case that sells liquid cooling in a hot dry climate is running against the committee that wrote the classes. Warmer supply water, more waterside economizer hours, fewer compressor hours: that is the pitch, and the paper has the supply temperature going the other way as heat flux at the socket rises.
Which horizon a source is describing matters here. LBNL's 2024 United States data center energy usage report models AI liquid cooled space at W45 across a horizon running to 2028, and cites the 2014 class edition for that assumption. ASHRAE projects the other way over the same years. Both are credible and they disagree, and a plant sized on either one should know which it is betting on.
ASHRAE TC 9.9, Emergence and Expansion of Liquid Cooling in Mainstream Data Centers, 2021.
The second constraint is geometry. The paper cites Datacom Book 13 for best of breed raised floor delivery at 1900 cfm per floor tile. Against that it sets servers "that require 100 cfm or more per U," and notes that relying solely on a single tile "would translate into only occupying 19U within a rack."
So the airflow argument arrives in a unit a designer already owns. Nineteen rack units out of a 42U or 48U cabinet, taking the conventional cabinet heights as the assumption rather than as a sourced figure, and the rest of the cabinet has to come from somewhere other than the floor, which is where overhead, in-row and rear-door coils enter the drawing.
The numbers the trade quotes instead are 15 kW, 20 kW, 25 kW and 30 kW, and none of them traces to a published limit. Title 24's containment trigger, 10 kW of ITE design load per room, is what the Energy Commission summarizes out of Section 140.9(a). It decides whether you need barriers, not whether air still works, and its own exception for racks under 1 kW is the only per-rack number in the section.
That 5 kW is not a mechanical load. Server fans ride the same UPS as the servers, so ASHRAE works the trade off on the electrical side: going from 2 percent to 10 percent fan power "equates to reducing the data center UPS capacity by 8%."
The air side got colder too. The 5th edition Thermal Guidelines added class H1 for high density air cooled equipment, allowable 15C to 25C and recommended 18C to 22C. ASHRAE also records that many platforms once rated A4, good to 45C, have "slipped back to only supporting the A3 or even A2 class."
NVIDIA's own hardware documentation for the DGX GB rack scale systems says "The rack power consumption is approximately 120kW." The figure is stated generically across GB200 and GB300 rather than split by model, so it belongs to a DGX GB NVL72 rack. Divide 120 by 9 and the reference rack sits more than thirteen times out from where the fleet actually is.
It is also not fully liquid cooled. The same page says the compute trays are cooled by liquid running through manifolds to cold plates on the CPUs and GPUs. Then the caveat: "The rest of the components like networking and storage devices are air cooled, which is pushed through the system by the fans." A designer who believes the marketing and provides no room air path has built a rack that cannot cool its own switches.
What fills the densest racks is mostly not AI. Uptime's respondents name enterprise applications at 61 percent against 20 percent for generative AI training and 14 percent for inference, and the report puts it plainly: it is "only above 60 kW per rack that these computationally intense workloads take primacy over enterprise software."
So a generation request and a load request move on separate tracks under separate tariffs, and generation is the one with the ISO leading, which makes where a project can physically go a different problem from getting served. The mechanical designer controls neither document.
The paper carries the Energy Commission's 2025 demand forecast for data center load in the ISO Balancing Authority area: 1.8 GW added by 2030, 4.9 GW by 2040, against CPUC resource planning that expects more than 30 GW of new peak demand by the mid-2030s. That is the ISO grid rather than the state. The Energy Information Administration's balancing authority reference tables list Los Angeles Department of Water and Power and Imperial Irrigation District as active balancing authorities in their own right, directly interconnected with CAISO rather than inside it. So load in their territories is not in that 1.8 GW.
Supply has been arriving on that grid. Roughly 31 GW of nameplate new generation and storage has come online to serve ISO load since 2020, more than 16 GW of it storage. In 2025 California brought more than 6 GW of new nameplate capacity online in ISO territory, for the second consecutive year.
That 6 GW is all resource types rather than storage alone. The transmission CAISO names as approved for data center load is in the San Jose region, and the paper names no Southern California equivalent.
Answering the water objection moves the number the wrong way. LBNL states the trade off without hedging: water cooled chillers and other evaporation based cooling systems "are generally more energy efficient than an air-cooled chiller or other waterless systems. While air-cooled chillers use no water, they use more energy." Water is its own argument with a supply side build out running in parallel, and it is not the argument the interconnection queue is answering.
Which leaves load as the only term the design team still gets to argue about, and the case for managing a load down instead of upsizing the service is the same case at a smaller scale. A plant laid out on the assumption that supply water gets warmer as the racks densify has been sized against the committee's own projection. The correction is mechanical cooling, and mechanical cooling is load, and the load connection belongs to the participating transmission owner on its tariff and its calendar.
TC 9.9's 2021 white paper on liquid cooling frames it as a replacement cycle rather than a forecast. Direct liquid cooled products that today "might have the ability to operate on W45 (W4)" get replaced by "newer, higher-powered products only capable of W32 (W3) operation, and eventually cooling requirements may drive facility water temperatures down into the W27 (W2) range. Facilities should plan accordingly."
ASHRAE is careful about which part of that it stands behind. A page earlier the same section says the exact numbers are somewhat speculative, because IT manufacturers have detail on only one to two future chip generations. The direction is the committee's position; the three temperatures are its best estimate.
So the free cooling case that sells liquid cooling in a hot dry climate is running against the committee that wrote the classes. Warmer supply water, more waterside economizer hours, fewer compressor hours: that is the pitch, and the paper has the supply temperature going the other way as heat flux at the socket rises.
Which horizon a source is describing matters here. LBNL's 2024 United States data center energy usage report models AI liquid cooled space at W45 across a horizon running to 2028, and cites the 2014 class edition for that assumption. ASHRAE projects the other way over the same years. Both are credible and they disagree, and a plant sized on either one should know which it is betting on.
ASHRAE TC 9.9, Emergence and Expansion of Liquid Cooling in Mainstream Data Centers, 2021.
The transition is a socket number, not a rack number
ASHRAE puts the break in the density curve at around 2018, when a decade of large performance gains at modest power increases ended, and it states the air to liquid transition in socket watts. "As socket power moves through 300 W toward 400 W, standard 1U and 2U servers become more difficult from an air-cooled standpoint." That is a chip level threshold, and no rack level rule of thumb can see it.The second constraint is geometry. The paper cites Datacom Book 13 for best of breed raised floor delivery at 1900 cfm per floor tile. Against that it sets servers "that require 100 cfm or more per U," and notes that relying solely on a single tile "would translate into only occupying 19U within a rack."
So the airflow argument arrives in a unit a designer already owns. Nineteen rack units out of a 42U or 48U cabinet, taking the conventional cabinet heights as the assumption rather than as a sourced figure, and the rest of the cabinet has to come from somewhere other than the floor, which is where overhead, in-row and rear-door coils enter the drawing.
The numbers the trade quotes instead are 15 kW, 20 kW, 25 kW and 30 kW, and none of them traces to a published limit. Title 24's containment trigger, 10 kW of ITE design load per room, is what the Energy Commission summarizes out of Section 140.9(a). It decides whether you need barriers, not whether air still works, and its own exception for racks under 1 kW is the only per-rack number in the section.
Fan power fell below 2 percent and is coming back
Server fan power ran as high as 20 percent of server power, fell below 2 percent through the multicore period, and is climbing again on the densest hardware. ASHRAE: "A fan power percentage of 10% to 20% is not uncommon for some of the denser servers... In a 50 kW rack, the fan power translates to be at least 5 kW."That 5 kW is not a mechanical load. Server fans ride the same UPS as the servers, so ASHRAE works the trade off on the electrical side: going from 2 percent to 10 percent fan power "equates to reducing the data center UPS capacity by 8%."
The air side got colder too. The 5th edition Thermal Guidelines added class H1 for high density air cooled equipment, allowable 15C to 25C and recommended 18C to 22C. ASHRAE also records that many platforms once rated A4, good to 45C, have "slipped back to only supporting the A3 or even A2 class."
The fleet is at 9 kW and the reference rack is 120 kW
Uptime Institute's 2025 global survey puts the average of modal rack densities at almost 9 kW, up from 8.3 kW the year before, and at 7.5 kW once the small number of facilities whose typical rack is already 30 kW or above are taken out. The mode of the distribution is 4 to 5 kW, at 30 percent of respondents. More than 80 percent of operators report no racks above 30 kW anywhere in their facility. Uptime also puts the industry weighted average annual PUE at 1.54 in 2025, the sixth consecutive year effectively flat, against 1.48 for facilities commissioned within five years and 1.44 for facilities of 20 MW and above.NVIDIA's own hardware documentation for the DGX GB rack scale systems says "The rack power consumption is approximately 120kW." The figure is stated generically across GB200 and GB300 rather than split by model, so it belongs to a DGX GB NVL72 rack. Divide 120 by 9 and the reference rack sits more than thirteen times out from where the fleet actually is.
It is also not fully liquid cooled. The same page says the compute trays are cooled by liquid running through manifolds to cold plates on the CPUs and GPUs. Then the caveat: "The rest of the components like networking and storage devices are air cooled, which is pushed through the system by the fans." A designer who believes the marketing and provides no room air path has built a rack that cannot cool its own switches.
What fills the densest racks is mostly not AI. Uptime's respondents name enterprise applications at 61 percent against 20 percent for generative AI training and 14 percent for inference, and the report puts it plainly: it is "only above 60 kW per rack that these computationally intense workloads take primacy over enterprise software."
The ISO takes the generator, the utility takes the load
Colder facility water means more compressor hours, and compressor hours are kW, and the kW is somebody else's decision. CAISO's Large Load Considerations issue paper of January 30, 2026 sets out the split: the current assignment "gives lead responsibilities for resource (generator and storage) interconnection and transmission planning to the ISO, and load connections to the utilities, specifically the participating transmission owners (PTOs)." The paper's own diagram annotates the rest: study costs, retail rates and potential cost recovery are "governed by local utility tariff, approved by Local Regulatory Authority (in most cases the CPUC)," and a footnote adds that the CPUC calls load interconnection energization.So a generation request and a load request move on separate tracks under separate tariffs, and generation is the one with the ISO leading, which makes where a project can physically go a different problem from getting served. The mechanical designer controls neither document.
The paper carries the Energy Commission's 2025 demand forecast for data center load in the ISO Balancing Authority area: 1.8 GW added by 2030, 4.9 GW by 2040, against CPUC resource planning that expects more than 30 GW of new peak demand by the mid-2030s. That is the ISO grid rather than the state. The Energy Information Administration's balancing authority reference tables list Los Angeles Department of Water and Power and Imperial Irrigation District as active balancing authorities in their own right, directly interconnected with CAISO rather than inside it. So load in their territories is not in that 1.8 GW.
Supply has been arriving on that grid. Roughly 31 GW of nameplate new generation and storage has come online to serve ISO load since 2020, more than 16 GW of it storage. In 2025 California brought more than 6 GW of new nameplate capacity online in ISO territory, for the second consecutive year.
That 6 GW is all resource types rather than storage alone. The transmission CAISO names as approved for data center load is in the San Jose region, and the paper names no Southern California equivalent.
Answering the water objection moves the number the wrong way. LBNL states the trade off without hedging: water cooled chillers and other evaporation based cooling systems "are generally more energy efficient than an air-cooled chiller or other waterless systems. While air-cooled chillers use no water, they use more energy." Water is its own argument with a supply side build out running in parallel, and it is not the argument the interconnection queue is answering.
Which leaves load as the only term the design team still gets to argue about, and the case for managing a load down instead of upsizing the service is the same case at a smaller scale. A plant laid out on the assumption that supply water gets warmer as the racks densify has been sized against the committee's own projection. The correction is mechanical cooling, and mechanical cooling is load, and the load connection belongs to the participating transmission owner on its tariff and its calendar.