Liquid Cooling, and the End of Air

Google reports liquid cooling at gigawatt scale across more than 2,000 TPU pods over seven years from TPU v3 through Ironwood, with fleet-wide uptime consistently about 99.999% and CDU availability about 99.999% since 2020, achieved through redundant pump and heat-exchanger units on UPS — six per rack, of which five suffice.

The CDU exists to demarcate the Facility Water System from the Technology Cooling System — ASHRAE's stated recommendation — coupling them only through a heat exchanger, with pumps, filtration and sensors on the server side; its key metric, approach temperature, is the gap between facility-supplied and server-supplied liquid.

Google's fifth-generation Project Deschutes CDU design was opened through the Open Compute Project, and manufacturers building to it state 2 MW of cooling capacity, a 3 degrees Celsius approach temperature difference, 80 PSI available pressure, 0.2 micron side-stream filtration and fully redundant power feeds per pump circuit.

Leak handling is layered rather than clever: wireless spot detectors at high-risk points plus rope detectors around cabinet perimeters and vertical pipe runs, coupled to automatic shutoff valves, with equipment shipping built-in leak detection and pressure relief and certified to UL/CSA/IEC 62368-1.

ASHRAE states chip power is in uncharted territory and loss of cooling can be catastrophic, naming throttling and rapid-temperature-spike damage as its two concerns, and recommending thermal inertia, active redundancy through changeover, transient modelling, coolant quality monitoring and load migration inside the minimum server time-to-throttle.

Direct-to-chip cooling runs a facility primary loop and a server secondary loop that never mix, exchanging heat through a plate heat exchanger in the CDU, with rack manifolds, sensors and control valves ending at a quick-disconnect coupling — the natural operator-customer demarcation, except that the coolant crosses it, so neither party owns fluid chemistry.

Liquid's efficiency case is that warmer coolant still works, letting facilities run warmer water and cut or drop mechanical chillers, with pump power under 5% of equivalent fan power; NVIDIA's 300x water and 25x energy ratios carry no stated baseline and are recorded as claims, not findings.

3M announced on 20 December 2022 that it would exit all PFAS manufacturing by end-2025, covering all fluorinated fluids, against about $1.3bn of annual net sales and $1.3-2.3bn of pre-tax charges; trade reporting ties this and EPA hazardous-substance classification to two-phase immersion cooling specifically, single-phase being able to use mineral or plant-based oils.

LiquidStack's chief executive argued the 3M exit mainly hits semiconductor etch rather than cooling and that better alternative fluids already exist — but the reporter noted one named replacement, Solvay's Galden, is itself a PFAS.

In-chip microfluidics was demonstrated by IBM in 2008 and funded by DARPA from 2013 without reaching a shipping data-centre product, and the open questions — yield, defect rates from extra fabrication steps, liability for post-assembly leaks, no fleet reliability data, no interface standards, no retrofit path, no production timeline — remain unanswered.

Microsoft and Corintis etched hair-width, AI-designed leaf-vein coolant channels into the silicon itself, reporting up to 3x better heat removal than cold plates and up to a 65 percent reduction in peak silicon temperature rise in a GPU, with 13 percent lower intake fluid temperature and 55 percent lower pressure drop than parallel-microchannel cold plates.

Google's Brazos is a rack-mounted closed-loop liquid-to-air system supporting 60 kW per rack across three 11-OU modules in an OCP ORv3 rack, running deionised water or 25% propylene glycol on a 40-60 V DC busbar input, installable one rack at a time with no facility chilled-water retrofit.

An operator states that retrofitting air-cooled data centres is extremely complex and costly and that designing for liquid from the outset is typically faster, cheaper and less risky, because liquid cooling requires customer, equipment maker and facility operator to agree on temperature, pressure and flow — mismatches can prevent deployment.

Water carries approximately 4,000 times more heat per unit volume than air with roughly 30 times the thermal conductivity; against halls designed for 5-10 kW racks and an air ceiling put at about 15 kW, direct-to-chip liquid cooling reaches around 145 kW and GPU racks now exceed 200 kW, trending toward a stated 1 MW target for 2030.

GPU power rose from approximately 300 W in 2018 to 1,200 W in 2024, with accelerators now exceeding 1,000 W against a historical 100 W, and localised hotspot power densities predicted to rise by up to 2.5 times between 2018 and 2023 — so the heat problem is concentration as well as total.

Computing equipment has been liquid-cooled by technologies developed by IBM and others since the 1960s; the early-1990s bipolar-to-CMOS shift cut power density enough to make air cooling economical for about a decade before densities returned to pre-1990s levels in the early 2000s — making air the interlude rather than the baseline.

Vertiv's filed second-quarter 2026 results show net sales of $3,274 million, up 24% year on year with 18% organic, and full-year guidance at a $14,000 million midpoint — but the filing carries no liquid-cooling line item, so none of that growth can be attributed to liquid cooling.

At least two unrelated manufacturers built to Google's opened CDU specification — Boyd showing a conforming 2 MW unit in October 2025 and Nidec prototyping another at 2 MW class and 80 PSI for SC25 in November 2025 — turning a hyperscaler's internal design into a multi-sourced purchasable one.

Schneider Electric agreed on 17 October 2024 to acquire a 75% controlling interest in the liquid-cooling specialist Motivair for $850 million in cash, with the remaining 25% planned for 2028 — a price paid in what the reporting describes as a nascent market.