In short: Oregon State University's College of Earth, Ocean, and Atmospheric Sciences retrofitted ZutaCore® HyperCool® waterless two-phase direct-to-chip liquid cooling into an existing data center without replacing a single server. In-house technicians converted each machine in 15 to 20 minutes with no downtime, and the university reclaimed roughly 300 kVA that would otherwise have gone to four more CRAC units.
The Numbers!
| Servers replaced: | Zero — existing Dell, HP and Supermicro retained |
| Install time: | 15–20 minutes per machine, in-house |
| Power reclaimed: | ~300 kVA (four CRAC units at 75 kVA each avoided) |
| Cooling load avoided: | 240 additional tons projected under air cooling |
| Processor temperatures: | ~72°C to 66°C; up to 30% lower |
| Compute performance: | 20% improvement |
| Downtime: | None |
| Time to full deployment: | ~1 year from first conversation |
Most liquid cooling conversations assume a new build — a greenfield site, a clean sheet, a design that has the cooling in it from the start. That isn't the situation most people are in.
On DCD's Keeping It Cool broadcast, Chris Sullivan, Director of Research and Academic Computing at Oregon State University's College of Earth, Ocean, and Atmospheric Sciences, described the other situation. An existing data center due for a rebuild. Hardware already bought. Grant money already spent. And science running around the clock that couldn't be paused while anyone rethought the thermals.
What follows is his account of getting liquid cooling into that environment — including the parts that don't appear on a datasheet.
Why did Oregon State's cooling have to change?
Oregon State's cooling had to change because expanding air cooling would have swallowed roughly 300 kVA — power the university wanted for compute, not for moving air.
Sullivan's framing is worth borrowing: he has never met a data center without cooling limits. What changed was where those limits started to bite. The College had 70 tons of cooling in place and projections for another 240 tons to handle what was coming, and each CRAC unit he switched on drew 75 kVA. He was looking at adding four more. That is the calculation that turns cooling from an operational line item into a strategic constraint. Once he'd done it, the question stopped being how to cool more and became how to get that power back.
There was a second signal, and it was audible. Sullivan had started hanging headphones inside the data center so administrators could work in there; pushed hard, the machines were loud enough to hear from outside the room. Stacking in more fans was a diminishing return. Both problems pointed at the same root cause — heat being moved by air, a long way from where it was generated.
Can you retrofit liquid cooling without replacing your servers?
Yes — Oregon State University kept its existing Dell, HP and Supermicro fleet and replaced none of it. That single constraint shaped the entire evaluation, because it ruled out most of the market before any technology was assessed.
Sullivan's compute is bought with research grants, and once a grant is spent, it's spent. The money can't be re-allocated, and the hardware has to keep answering research questions for three to five years, sometimes considerably longer. The Ocean Observatory Initiative alone is a $220 million programme streaming minute-level data continuously.
So any cooling approach that required new servers wasn't really a cooling decision. It was a decision to buy the hardware twice, and there was no budget for that.
That's what took immersion off the table — capable, but it meant replacing equipment he already owned, from a relatively small field of vendors, for an investment expected to last ten to fifteen years. Single-phase water stayed in contention longer, until visits to other data centers turned up water sitting on the floor. Running inside a chilled water loop, he needed the water to stay in the loop.
Two-phase direct-to-chip answered both questions at once. Because a dielectric fluid evaporates at the chip rather than circulating volumes of water, the flow rates are dramatically lower — so he wouldn't simply be trading a CRAC unit for a pump system — and the technology retrofits onto existing racks, processors and chilled water systems.
What convinced the facilities team?
The facilities team signed off because the connection came down to two hoses and low flow rates.
Every infrastructure project has a gatekeeper, and it's rarely the person who chose the technology. Sullivan's facilities group had to approve this one, because they'd own the water flow, the manifold, and everything up to the CDU. They worked through the options and kept finding reasons to say no: this manifold looks hard to manage, that flow rate is a problem, this will be difficult for us.
Then they got to HyperCool.
"Two hoses. We're great. Where do you want those?"
That was effectively the entire negotiation — and it opened the door to the test that actually mattered. Because the main manifold wasn't built out yet, the pilot rack went into someone's lab, tapped into the building's chilled water loop, with people still working in the room. Turn on a rack of air-cooled machines in an occupied lab and people leave. That rack ran for close to a year at full performance, and nobody complained.
The install was equally undramatic. Sullivan's grants had no line item for external labour, so the real question was whether his own people could do the work. They could: 15 to 20 minutes per machine, every unit checking out, with a learning curve on sequence rather than skill. The team later disconnected the whole deployment and moved it room to room unaided. None of it required taking the data center down while that minute-level data kept arriving. From first conversation to fully deployed data center took about a year.
What is the hardest part of a liquid cooling retrofit?
At Oregon State, the hardest part wasn't the cooling at all. It was rack width.
Wider cabinets mean four legacy racks come out for every three that go in, which forces a re-plan of hot and cold aisle alignment in sets of four. Anyone expecting a one-for-one swap gets caught out. Oregon State folded the change into a rebuild it was already planning, using the Chatsworth Products (CPI) ZetaFrame™ cabinet system, and that knowledge went on to shape how the new room was laid out. Raised early, it's a planning input. Discovered late, it's a problem.
What changed once it went live?
Processor temperatures fell from roughly 72°C to 66°C and held flat under load, eliminating the thermal throttling that had been quietly taxing performance — up to 30% lower processor temperatures and a 20% improvement in compute performance.
Asked what changed operationally, though, Sullivan's first answer was that not much did, which is rather the point. Working with Dell, the team took BIOS-level control of the server fans and wound them down, cutting power on a per-machine basis. The temperature curve disappeared. And the heat itself stopped being waste, redirected instead to warm nearby buildings in support of the university's sustainability goals.
Sullivan's project was never really about buying cooling. It was about refusing to accept that modernising meant starting over — and the estate you already own is usually the cheapest capacity you'll ever add, provided the cooling can meet it where it is.
Chris Sullivan spoke with Alison Deane of ZutaCore on DCD's Keeping It Cool broadcast. Watch the full session →
Frequently asked questions
Can liquid cooling be retrofitted into an existing data center? Yes. Oregon State University retrofitted ZutaCore HyperCool two-phase direct-to-chip cooling into an existing data center, keeping its Dell, HP and Supermicro servers and connecting to the campus chilled water system already in place.
Do you have to replace servers to deploy direct-to-chip liquid cooling? No. Direct-to-chip cold plates mount onto existing processors. Oregon State replaced zero servers, which was decisive because its hardware had been purchased with research grant funding that could not be spent again.
How long does a liquid cooling retrofit take per server? At Oregon State, in-house technicians completed each machine in 15 to 20 minutes. The full data center went from first conversation to deployed in approximately one year.
Does a liquid cooling retrofit require downtime? It did not at Oregon State. The retrofit was carried out without taking the data center offline, while a $220 million ocean observatory programme continued streaming minute-level data.
Is two-phase liquid cooling better than single-phase for a retrofit? Two-phase operates at significantly lower flow rates than single-phase water cooling and uses a dielectric fluid rather than water, keeping water away from the electronics. Oregon State chose two-phase after seeing water on the floor at data centers running single-phase systems.
Does liquid cooling reduce data center noise? Yes. Removing heat at the processor allows server fans to be turned down. Oregon State ran a full rack inside an occupied laboratory for close to a year at full performance without noise complaints.
Want the numbers on one page? Download the Oregon State University proof-point summary.