Direct to Chip Cooling
Cold Plates, CDUs and What a Data Center Has to Deliver Before You Sign
Every rack-scale GPU system now ships with direct to chip liquid cooling, and most data centers cannot serve it. Here is how it works, what a facility has to prove, and how to find one that can.
Price My Liquid-Cooled RequirementA liquid cooled data center in the marketing sense is common; one with a direct to chip loop reaching a specific hall on a specific date is not. A facility that "supports liquid cooling" and a hall with a commissioned loop to the white space you would occupy are two different answers to the same question, and the gap between them is where deployments go wrong. This guide to data center liquid cooling is written from the tenant's side, by an advisory practice that places liquid-cooled requirements with operators across North America, is paid by the operator you choose, and does not sell cooling equipment.
How Direct to Chip Cooling Works
A cold plate is a metal block with internal channels, mounted on the processor package in place of a heatsink. Coolant flows through the channels, absorbs heat, and leaves the server through flexible tubing to a manifold in the rack. From there it goes to a coolant distribution unit, which transfers the heat to the facility water system and returns cooled liquid to the rack. The facility water carries the heat to a dry cooler, chiller or cooling tower.
It is two loops, not one
The equipment loop runs treated coolant through the servers; the facility loop runs building water. Who owns each, and the unit between them, is a commercial question to settle before hardware ships.
It removes most of the heat, not all
Cold plates cover processors and sometimes memory. Power supplies and network cards still shed heat into the air, so a hall with no air handling is not the answer either.
It runs on warm water
Current cold plates accept 30 to 45°C inlet water, which can often be cooled by outside air instead of chillers. That changes both the operating cost and which facilities qualify.
The Five Questions a Facility Has to Answer
Each has a number or a date as its answer. A facility that answers with a capability statement has not answered.
| Question | A real answer | A portfolio answer |
|---|---|---|
| When is the loop commissioned in this hall? | A month, separate from the building's ready-for-service date | “The facility supports liquid cooling.” |
| What water temperature at the rack inlet? | A figure, typically 30 to 45°C, and what the plant holds in summer | “We can accommodate your requirements.” |
| What flow rate per cabinet? | Liters per minute per rack at your density, sustained across every rack | “Our engineering team will work with you.” |
| Who supplies the CDU, and at what level? | “You bring in-rack. We supply in-row. The loop is ours from the manifold out.” | “That is part of the design conversation.” |
| What is the floor load? | A figure in pounds per square foot, distributed and at the rack footprint | “The building is designed for high density.” |
Send the platform, rack count, date and markets before you contact an operator. You'll hear back within 24 hours with a first read on which facilities can serve it.
Why Rack-Scale GPU Systems Require It
Air cooling tops out around 30 to 35 kW per rack, and rear door heat exchangers stretch that to roughly 50 to 70. Every rack-scale GPU platform now shipping sits well above both, with cold plates fitted at the factory, so there is no air-cooled version to fall back on.
From NVIDIA and system vendor specifications and industry cooling limits, October 2026. Rubin Ultra racks planned for 2027 are expected near 600 kW. Free to cite with a link to this page.
The envelope moves every year, so a hall engineered for 140 kW racks today will not serve 230 kW racks without more plant, more flow and a stronger floor. A lease signed now typically runs five to ten years, which makes a facility's path to the next generation the second question to ask. Prices and specs for the current systems are on GB200 and GB300 NVL72; smaller GPU servers with two to eight accelerators often run air-cooled or with a rear door, and are covered on AI and GPU colocation.
What Is a Coolant Distribution Unit, and Who Supplies It?
The coolant distribution unit, or CDU, sits between the facility water loop and the equipment loop. It keeps the two circuits separate, controls coolant temperature and flow to the racks, and does the pumping. In a CDU data center deployment, units can sit at three levels, and which party supplies which is the question to settle before hardware ships.
In-rack CDU
Inside the cabinet. Rack-scale systems often ship with one, depending on the system integrator, in which case the tenant brings it and the facility supplies water at the right temperature and flow.
In-row CDU
Serves a row of racks from the end of the row, sized in hundreds of kilowatts. Usually the facility's, though not always, and the level where responsibility most often goes unanswered.
Facility-level CDU
A plant room unit serving a hall or floor, sized in megawatts. Always the facility's.
A tenant arriving with in-rack CDUs at a facility that planned to supply them pays twice. A tenant expecting in-row distribution at a facility that never built it waits. Both are avoidable with one question in writing.
What Is Cold Plate Cooling?
Cold plate cooling is the component-level view of direct to chip: a machined block, usually copper, with microchannels that bring coolant within millimeters of the die. Rack-scale systems ship with the plates fitted, so the tenant never chooses or installs them. What the tenant has to confirm is what the plates need from the building.
Inlet temperature
Current platforms accept 30 to 45°C, up from 20 to 25°C in older designs. The warmer the accepted inlet, the more facilities qualify and the cheaper the building runs.
Flow rate
On the order of 130 to 150 liters per minute for a fully populated rack-scale cabinet, sustained to every rack at once, not just a demonstration cabinet.
Coolant chemistry
Treated water or a glycol mix with corrosion inhibitors, specified by the hardware vendor, which is one more reason the two loops stay separate.
Direct to Chip vs Rear Door Heat Exchanger vs Immersion Cooling
| Method | How it works | Practical ceiling | What the facility supplies | Serves rack-scale GPUs? |
|---|---|---|---|---|
| Rear door heat exchanger | A liquid-cooled radiator replaces the rack's back door; exhaust air passes through it | 50 to 70 kW per rack | Facility water to the door | No |
| Direct to chip | Cold plates on the processors, coolant loop into the rack through a CDU | 200 kW+ per rack | Water loop to the white space, CDU at row or facility level | Yes, and required |
| Immersion | Hardware submerged in dielectric fluid in a tank | Highest of any method | Tanks, fluid handling and a different floor layout | Not in current rack-scale form |
Rear door heat exchangers
The least invasive way to add liquid cooling to an existing hall, and the usual retrofit answer up to the mid tens of kilowatts. A facility offering one for a 140 kW requirement has misread the requirement.
Immersion cooling
Reaches the highest densities, but needs tanks instead of racks and a different floor. Current rack-scale GPU systems do not ship in immersion form, so an immersion cooling data center is not an option for them.
Direct to chip
The answer for rack-scale hardware because it is the only method the hardware is built for. The comparison that matters is between facilities with a loop reaching the white space today and those without.
What a Liquid-Cooled Search Usually Finds
When the five questions are asked across a market, the answers fall into a few recognizable patterns. Data Center Knowledge reported on this gap in August 2026, drawing on our findings.
The building is ready; the loop is not
Space is offered at the right density, but the liquid loop is commissioned a quarter or more after the building opens. The most common way a timeline fails.
Build to suit, described as capacity
An existing building that would be modified for density after a tenant commits. Useful later, but not capacity on a near-term date.
Several sites, none big enough
A portfolio of smaller halls that together reach the target but individually cannot, with growth paths that land later.
Purpose-built, answering in numbers
Facilities designed for liquid cooling from the start answer all five questions with figures. When something rules them out, it is usually the date, not the engineering.
What a Liquid Cooling Retrofit Involves
Most facilities that will serve direct to chip in 2027 are being converted rather than built. A data center liquid cooling retrofit is a construction project, not a provisioning task.
That is for a hall that already has adequate power and floor loading; longer where those also have to change. It is why "we are adding liquid cooling" is not the same as a date. A facility that can name the month its loop is commissioned in your hall has a schedule. One that describes the intent is describing a project.
Send the platform, rack count, density, date and markets. You'll hear back within 24 hours, and the operator you choose pays us.
How We Help on Liquid-Cooled Requirements
You send the requirement
Platform, rack count, density, date and markets. You'll hear back within 24 hours, at no cost.
We verify at the hall level
We ask the five questions of every facility that might serve it and bring you the ones that answer with numbers, not the ones that describe a portfolio.
We negotiate the liquid-cooling terms
The loop date as a delivery obligation, water temperature and flow as service levels, and CDU responsibility in writing, alongside the rate and term. The operator you choose pays us.
Developers and operators asking the other side of the question, what density a site or building can serve and what a specification forecloses, can hire us through data center consulting at a fixed fee.
Direct to Chip Cooling: Frequently Asked Questions
What is direct to chip cooling?
A liquid cooling method where a cold plate is mounted on each processor, coolant flows through it to absorb heat, and the warm coolant leaves the rack through a loop to a coolant distribution unit that passes the heat to the facility water system. It moves 70 to 80 percent of rack heat into liquid and is required for every current rack-scale GPU platform.
What is cold plate cooling?
The component-level mechanism inside direct to chip: a machined metal block with internal microchannels mounted on the processor in place of a heatsink, with coolant flowing through it. Cold plates arrive fitted on rack-scale GPU systems; the tenant only has to confirm the facility can supply the inlet temperature and flow they need.
What is a coolant distribution unit (CDU)?
A CDU sits between the facility water loop and the equipment coolant loop. It keeps the two circuits separate, controls coolant temperature and flow to the racks, and provides pumping. CDUs come at three levels: in-rack, in-row and facility-level.
Who supplies the CDU in a colocation deployment?
It depends on the level. In-rack CDUs often ship with rack-scale systems and the tenant brings them; in-row CDUs are usually, but not always, supplied by the facility; facility-level units are always the operator's. Agree which party supplies which in writing before hardware ships.
What water temperature does direct to chip cooling need?
Current rack-scale platforms accept inlet water at 30 to 45 degrees Celsius, up from 20 to 25 in older designs. Specify the warmest water your hardware accepts, since warm water can often be cooled by outside air, which widens the list of qualifying facilities and lowers operating cost.
What flow rate does a liquid-cooled GPU rack need?
For a fully populated rack-scale cabinet at 140 to 160 kW, on the order of 130 to 150 liters per minute. The facility has to sustain that flow to every rack in the deployment at once.
Direct to chip vs immersion cooling: which is better?
For rack-scale GPU systems, direct to chip, because it is the only method those systems are built for. Immersion reaches higher densities but needs tanks, special hardware handling and a different floor layout, and current rack-scale GPU platforms do not ship in immersion form.
Can a rear door heat exchanger cool a GPU rack?
For GPU servers up to roughly 50 to 70 kW per rack, yes. For rack-scale systems such as the GB200 and GB300 NVL72, no: they ship with cold plates fitted and need a direct to chip loop from the facility.
Which GPU systems require direct to chip cooling?
Every rack-scale NVIDIA platform from Blackwell onward: the GB200 NVL72 at about 120 to 130 kW per rack, the GB300 NVL72 at about 140 to 160 kW, and the Vera Rubin NVL72 at about 190 to 230 kW.
How much does a rack-scale GPU system weigh?
A fully populated GB200 or GB300 NVL72 cabinet weighs roughly 1,360 kilograms, about 3,000 pounds. Many data halls are rated around 250 pounds per square foot; rack-scale deployments generally want 300 and above, and the load at the rack footprint matters as much as the average. Confirm it in writing before ordering hardware.
How long does a liquid cooling retrofit take?
Nine to fifteen months from decision to commissioned loop for a hall that already has adequate power and floor loading, and longer where those also have to change. It is a construction project, and the loop is often commissioned months after the building is ready for service.
What should a lease say about liquid cooling?
Five things beyond a standard agreement: the loop commissioning date in your hall as a delivery obligation, the water temperature at the rack inlet and the flow rate per cabinet as service levels, which party supplies the CDU at each level, and the floor load in writing.
Why does warm water change which facilities qualify?
Warm-water cooling can often reject heat to outside air through dry coolers instead of running chillers, which in cool climates covers most of the year. Facilities in Quebec, for example, publish free cooling for roughly 80 percent of annual hours. That lowers the operating cost and widens the list of facilities that can serve the deployment.
How many data centers can host direct to chip cooling today?
Far fewer than claim liquid cooling capability. Capacity with a commissioned loop at rack-scale density inside a near-term quarter is scarce, and requirements that can flex on date or density place much more easily than those that cannot.
What does it cost to use Metro Colo Advisory?
Nothing for placement: the operator you choose pays us through its channel program. Developers and operators who want to know what a building can serve can hire us for fixed-fee consulting.
Tell Us What You Are Deploying
The platform, the rack count, the date and the markets. That is enough for a first read on which facilities can serve it, and you'll hear back within 24 hours. If no facility on your timeline can serve your density, we will tell you, because that is a real answer and often the one at this tier.
Once you have contacted an operator directly, it usually will not work through an advisor on your deal.
The hardware: GB200 and GB300 NVL72, B200 and B300 rental, GPU cluster rental and the GPU rental hub.
Hosting it: AI and GPU colocation, wholesale colocation, the data center lease guide and data center cost.
Building it: data center consulting, powered shell and build to suit.
Markets: Northern Virginia, Dallas, Phoenix, Chicago, Atlanta and Ohio, and how a data center broker works.