Extending 45°C Cooling Beyond NVIDIA’s Rubin AI Factories
Key Takeaways:
- NVIDIA’s Rubin‑generation AI infrastructure proves that 45°C, fully liquid‑cooled systems with no fans can dramatically cut data center cooling energy and water use in new hyperscale builds.
- Direct‑to‑chip warm‑water cooling depends on CDUs, dry coolers, and purpose‑built facility water systems, which most legacy data centers, plants, hospitals, and edge sites simply don’t have.
- Iceotope’s chassis‑level precision liquid cooling achieves similar 45°C inlet operation without CDUs or facility plumbing, making warm‑water AI cooling practical across brownfield and edge deployments, not just Rubin‑class factories
NVIDIA recently announced that their Rubin-generation AI infrastructure supports inlet coolant temperatures up to 45°C, calling it one of the biggest efficiency leaps in data center history. Why is this such a big deal? For the past few decades, colder is better has been the modus operandi for datacenter build-outs. Local water loops relied on chilled water to keep GPUs and CPUs at an optimal temperature by drawing away their heat.
Now, NVIDIA is changing the rules by informing their ecosystem that their hardware can be cooled entirely by liquid in a closed loop and without the use of fans, essentially eliminating the need for water in evaporative cooling.
We’re as excited as any company to hear this news. Iceotope has made the case for liquid-cooled systems at 45°C inlet temps for years, and we’re glad to see it becoming the norm. Now that NVIDIA is all in on liquid cooling, it makes our jobs easier.
But NVIDIA’s recent announcement only works for new datacenters specifically built-to-spec for these engineering features. Legacy datacenters, manufacturing plants, hospitals, and traditional on-prem deployments are all unable to benefit from Nvidia’s latest warm-facility-water architecture without major and costly retrofits.
What direct-to-chip cooling relies on
NVIDIA's design utilizes direct-to-chip cold plate cooling, built around a coolant distribution unit (CDU) that manages flow, pressure, and chemistry across the required facility water system (FWS). It also depends on dry coolers sized to the site and a climate where ambient air is cool enough, for most of the year, to reject heat without mechanical assistance. NVIDIA acknowledges this in their announcement. A facility in the Scottish Highlands and one in Phoenix face extremely different conditions, and chillers may have to amplify cooling once ambient temperatures climb.
That's a reasonable trade-off for a hyperscale AI factory, where the CDU plant, the dry cooler farm, and the site's climate are all known knowns.
Why direct-to-chip isn’t the only answer
Not all AI compute will live in the halls of new hyperscale builds, and most edge compute won’t require NVL144 sized racks. Like all technologies, compute naturally coalesces near data sources: in manufacturing plants, in hospital server rooms, and in telecom cabinets. These sites were built for a different era of power density, and not plumbed to support a D2C system. A CDU is a substantial piece of standalone infrastructure. It must be sized, commissioned, and maintained. Most legacy sites have no path to installing one, let alone the dry cooler capacity behind it.
While NVIDIA is adopting 45C inlet temps for future deployments, brownfield and edge deployments need not apply. Warm-inlet direct-to-chip cooling technology works fine, but the infrastructure it depends on only exists in a small number of purpose-built facilities.
Where Iceotope's approach differs
Iceotope's chassis-level directed immersion cooling, doesn't carry the same facility dependencies. Each server sits in its own sealed, rack-mounted chassis, partially filled with a small quantity of single-phase dielectric coolant. The coolant is then pumped through the chassis via manifolds to server hotspots, capturing heat through forced convection. The dielectric fluid cycles back to an in-chassis heat exchanger to pass heat to a local cooling loop and the process begins again. No water reaches the server and no facility plumbing is required. All that’s needed is an inlet and outlet pipe to a low-flow chiller or dry cooler.
Because the dielectric fluid is in contact with the whole board rather than just the chip, heat transfer is efficient enough that the facility water on the other side of that heat exchanger can run at 45°C inlet.
That difference is what lets the warm-water case extend beyond the data center. NVIDIA's architecture delivers it inside a hyperscale AI factory built to support it. Iceotope's precision liquid cooling delivers the same result in on-premises deployments and edge sites where a CDU can’t easily be installed.
What's next for the market
NVIDIA has made a strong case for warm water cooling (to be fair, 75% water and 25% propylene glycol cooling) and their support is an obvious boost for the liquid cooling industry. But NVIDIA’s direct-to-chip approach only works inside new datacenter buildouts. Iceotope's precision liquid cooling solves the same thermal bottleneck, but can do so more flexibly in more locations and with less complexity.
