Direct to Chip Cooling: What a Facility Has to Deliver Before You Sign
Direct to chip cooling puts a cold plate on the processor, runs coolant through it, and carries the heat out of the rack in a liquid loop rather than through the air. It is the only GPU liquid cooling method that serves current rack-scale systems, and it is the reason most colocation facilities cannot take them. A 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 saying it supports liquid cooling and a specific hall having a 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. Consider this your independent guide to data center liquid cooling at rack scale, written from the tenant’s side by an advisory practice that has taken liquid-cooled requirements to operators across North America and internationally, is paid by the operator you choose, and does not sell cooling equipment.
Summary: Direct to chip liquid cooling is required, not optional, above roughly 70 kilowatts per rack, and every rack-scale platform from Blackwell onward ships with it. What decides whether a facility can serve a deployment is not whether it lists liquid cooling as a capability but five specific things: the date the loop is commissioned as distinct from when the building opens, the water temperature delivered at the rack inlet, the flow rate per cabinet, who supplies the coolant distribution unit at each level, and whether the floor carries the weight. In a five megawatt search in 2026, almost every operator approached had space and almost none could answer those five. Data Center Knowledge published that finding across the first ten operators; the full search ran to more than twenty. This page is what it means for a tenant.
If you are placing a liquid-cooled requirement, tell us what you are deploying. The platform, the rack count, the date, and where. That is enough for a first read on which facilities can actually serve it and which are describing a portfolio. It comes to the principal directly and you will have an answer within 24 hours.
What is direct to chip cooling, and how does it work?
A cold plate is a metal block with internal channels, mounted directly on the processor package in place of a conventional heatsink. Coolant flows through the channels, absorbs heat by conduction, and leaves the server through flexible tubing to a manifold in the rack.
From the manifold the warm coolant 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 then carries the heat to a chiller, a dry cooler or a cooling tower, depending on the building and the climate.
That is the whole system, and three things about it matter to a tenant.
1. It is two loops, not one. The equipment loop runs a treated coolant through the servers. The facility loop runs water through the building. The coolant distribution unit sits between them and keeps them separate.
Who owns each loop, and who owns the unit between them, is a commercial question that has to be answered before hardware ships.
2. It removes most of the heat, not all of it. Cold plates cover the processors and sometimes the memory. Power supplies, network cards and other components still shed heat into the air. A direct to chip deployment typically moves 70 to 80 percent of rack heat into the liquid loop; the remainder still needs air handling. A hall with no air cooling at all is not the answer either.
3. It is warm water, usually. Modern cold plates work with inlet water at 30 to 45 degrees Celsius. That matters more than it sounds, because water at that temperature can often be cooled by outside air rather than by a chiller, which changes the operating cost of the building and the list of facilities that qualify.
How we help, and why the question is different here
With conventional colocation the question is price and terms. With a liquid cooling data center the first question is whether the facility can physically do it in the hall you would occupy, and most of what is published about high density cooling does not answer that.
1. We verify at the hall level, not the provider level.
An operator with liquid cooling somewhere in its portfolio is not an operator with a loop reaching the hall you would occupy. We ask five specific questions of every facility on a shortlist, and the answers are what decide whether it stays on the list. Those five are below.
2. We know what the answers actually look like.
Data Center Knowledge used the practice as its source in August 2026 on this exact gap: a five megawatt search at 140 to 160 kilowatts per rack across more than ten North American operators, where nearly every operator had space and almost none could serve the density with direct liquid cooling. That search is the reason we know which questions expose a portfolio answer.
3. We put the facilities that can do it in competition.
The pool that can actually serve rack-scale density is small. When three of them know they are competing on the same specification, the terms move, including on the ones that matter for liquid cooling: loop commissioning date, water temperature commitment, CDU responsibility.
4. It costs you nothing.
The operator you choose pays the placement fee as a standard part of its channel program. Going direct does not save it; it leaves it with the operator.
Metro Colo Advisory is an independent data center advisory practice. We do not own facilities and we do not sell cooling equipment. If the honest answer is that no facility on your timeline can serve your density, we say so, because that is a real answer and it is the one most searches at this tier arrive at.
Either way, tell us what you are deploying. Platform, rack count, date and market. That is enough to start.
What are the five things a facility has to answer?
These are the questions that separate a facility that can serve a direct to chip deployment from one that can describe liquid cooling. Every one of them has a number or a date as the answer, and a facility that answers with a capability statement instead has not answered.
| Question | What a real answer looks like | What a portfolio answer looks like |
|---|---|---|
| When is the loop commissioned in this hall? | A month. Distinct 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 45C, with what the plant can hold in summer. | "We can accommodate your requirements." |
| What flow rate per cabinet? | Liters per minute per rack, for the density you specified. | "Our engineering team will work with you." |
| Who supplies the CDU, and at what level? | "You bring in-rack. We supply row-level. 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 or kilograms per square meter. | "The building is designed for high density." |
The first question is the one that catches people. A building can be ready for service in February and the liquid loop commissioned in June, because the loop is often permitted and installed as separate works after the shell is handed over. A tenant who plans against the building date and discovers the loop date at contract stage has lost a quarter. It is the single largest cause of timeline failure in liquid-cooled placements, and the one operators are least likely to volunteer.
If you have a facility in front of you that has not answered these, send us what they have said. We can usually tell within a day whether it is a real answer or a portfolio one.
Why rack-scale GPU platforms require it
Air cooling has a physical ceiling around 30 to 35 kilowatts per rack. Rear door heat exchangers extend that to roughly 50 to 70. Every rack-scale GPU platform now shipping sits well above both.
| Platform | Rack System | Power per Rack | Cooling | Status |
|---|---|---|---|---|
| Blackwell | GB200 NVL72 | 120 to 140 kW | Direct to chip, CDU in-rack or in-row | Shipping |
| Blackwell Ultra | GB300 NVL72 | 140 to 160 kW | Direct to chip, CDU in-rack or in-row | Shipping |
| Vera Rubin | VR200 NVL72 | 190 to 230 kW | Direct to chip | Volume shipping second half 2026 |
| Rubin Ultra | Kyber NVL576 | ~600 kW | Direct to chip, Kyber rack | 2027 |
Two things follow. There is no air-cooled or rear-door configuration for any of these; the cold plates are part of the system as shipped. And the envelope moves every twelve months, so a facility engineered for 140 kilowatt racks today is not a facility that serves 230 kilowatt racks next year without more plant, more flow and more floor. A lease signed now typically runs five to ten years, which means the second question about a facility is what its path to the next generation looks like.
GPU liquid cooling at smaller scale, single servers with two to eight accelerators, is a different case, and it is where most of the published material on liquid cooling GPU deployments is actually about. Those often run air-cooled or with a rear door, and they are a bare metal or retail colocation conversation rather than a rack-scale one. The AI and GPU colocation page covers where each tier fits.
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 does three things: keeps the two circuits separate so a fault in one does not contaminate the other, controls the temperature and flow of coolant to the racks, and provides the pumping to move it. A CDU data center deployment has them at one or more of three levels.
In-rack
A unit inside the cabinet itself. Rack-scale systems from Blackwell onward frequently ship with in-rack CDUs, depending on the system integrator, in which case the tenant brings them and the facility supplies water at the right temperature and flow to the rack's connection points. Where the integrator's design uses row-level distribution instead, the responsibility shifts.
In-row
A larger unit serving a row of racks, typically sitting at the end of the row. Sized in the hundreds of kilowatts. Usually supplied by the facility, though not always, and this is where the responsibility question most often goes unanswered.
Facility-level
A plant room unit serving a hall or a floor, sized in megawatts. Always the facility's.
In a CDU data center deployment, the question that has to be settled before hardware ships is which of these the tenant provides and which the facility provides, because it changes capital cost, lead time, and who is responsible when something fails.
A tenant arriving with in-rack CDUs and a facility expecting to supply them is a duplicated cost. A tenant expecting row-level distribution and a facility that has not built it is a delay. It sounds obvious. It is asked far less often than it should be.
What is cold plate cooling, and what does it mean for the tenant?
Cold plate cooling is the component-level view of direct to chip. The plate itself is a machined block, usually copper, with internal microchannels that put coolant within millimeters of the die. Manufacturers such as CoolIT, Asetek, Boyd and Vertiv make them; rack-scale platforms ship with them fitted. A tenant deploying rack-scale hardware does not choose or install cold plates. They arrive on the system.
What the tenant does have to understand is what the plates need from the building. Three things.
Inlet temperature.
Cold plate liquid cooling is designed around a supply temperature, and the range has been moving warmer. Older designs wanted 20 to 25 degrees Celsius. Current rack-scale platforms accept 30 to 45. The higher the tolerated inlet temperature, the more facilities qualify and the cheaper the building is to run, because warm water can often be cooled by outside air.
Flow rate.
Plates need a specified volume of coolant per minute to carry the heat away at the design temperature rise. For a fully populated rack-scale cabinet that is on the order of 130 to 150 liters per minute. A facility loop that cannot deliver that per rack, sustained across the deployment rather than to a single demonstration cabinet, cannot serve the density regardless of what the plant is rated at.
Coolant chemistry.
The equipment loop runs treated water or a glycol mix with corrosion inhibitors. It has to be compatible with the plates, the tubing and the CDU. This is usually specified by the hardware vendor and it is one more reason the equipment loop and the facility loop stay separate.
Direct to chip, rear door, or immersion?
Three liquid cooling approaches, and they are not interchangeable.
| Method | How it works | Practical ceiling | What the facility supplies | Serves rack-scale? |
|---|---|---|---|---|
| Rear door heat exchanger | 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 processors, coolant loop into the rack via CDU. | 200 kW+ per rack | Water loop to the white space, CDU at row or facility level | Yes. Required. |
| Immersion | Hardware submerged in dielectric fluid in a tank. | Highest of any method | Tanks, fluid handling, different floor layout | Not in current rack-scale form factors |
- Rear door heat exchangers are the least invasive way to add liquid cooling to an existing hall, which is why they appear in most retrofit projects. They work well up to the mid tens of kilowatts. They do not serve rack-scale systems and a facility offering them as its liquid cooling answer for a 140 kilowatt requirement has misread the requirement.
- Immersion cooling reaches the highest densities of anything and is used in specialized deployments, but it requires tanks rather than racks, different hardware handling, and a different floor layout. Current rack-scale GPU platforms do not ship in immersion configurations. For a tenant deploying those systems, immersion is not on the menu.
- Direct to chip is the answer for rack-scale hardware because it is the only one the hardware is built for. The comparison worth making is not between the three methods but between facilities that have a direct to chip loop reaching the white space today and facilities that do not.
What does a search at rack-scale density actually find?
The clearest way to show what these questions are for is what happened when they were asked. The figures below are from a 2026 search for five megawatts of liquid-cooled capacity at 140 to 160 kilowatts per rack, commissioning inside a fixed quarter, across more than twenty operators in North America and internationally. Operators are not named.
| What the operator said first | What the five questions established | Outcome |
|---|---|---|
| Liquid-cooled capacity available at the required density in a primary market | Building ready for service in one quarter. Liquid loop commissioned in the following quarter, after handover, permits and works. Capacity capped at roughly a third of the requirement with no approved growth path. | Ruled out on date and on ceiling |
| Multiple sites, one at ten megawatts, three tenants competing | No building could be named. No power position, floor load or loop date could be given. Sites were existing buildings requiring density modifications, engineered after a tenant commits. | Build-to-suit, not capacity. Ruled out on date. |
| Six sites across a region, one to two and a half megawatts each | None reached five individually. A further site offered later had a growth path to five and then ten, ready for service two quarters past the date. | Ruled out on date. Held for expansion. |
| Purpose-built for density from the start, liquid cooling as the design premise | All five answered with figures. Capacity available two to three quarters past the date. | Ruled out on date only. Strongest technical fit in the search. |
| Declined at first contact, seven of ten in one channel | No availability at the density in current inventory or forecast. | Ruled out |
Three things to take from it.
- The loop date is the failure mode. The first row is the one that catches experienced buyers. Building live, loop live, and capacity available to a tenant were three different dates, months apart, and only the first was offered.
- The facilities that answered with numbers were the real ones. The fourth row is the only operator that gave all five answers as figures, and it was the only one whose only problem was timing. That correlation is not an accident.
- Almost nobody has this today. The honest state of the market at this tier is that capacity capable of serving rack-scale density inside a near-term quarter is close to nonexistent, and requirements that can flex on date or on density place far more easily than requirements that cannot. That is what the search found and it is what Data Center Knowledge reported.
What does a liquid cooling retrofit actually involve?
Most facilities that will serve direct to chip in 2027 are being converted rather than built. It is worth knowing what that means for timeline, and the same applies to an immersion cooling data center, which requires even more change to the floor.
A data center liquid cooling retrofit into an operating hall is a construction project, not a provisioning task. It typically involves a new or extended chilled water plant, pipework from the plant to the hall, a distribution manifold in the white space, row-level CDUs, and structural work if the floor was not rated for rack-scale weight. Each of those has permits, procurement lead times and works that cannot start until the hall is available.
Realistic durations run nine to fifteen months from decision to commissioned loop for a hall that already has adequate power and floor. Longer where power or structure also has to change. The plant itself is frequently the long pole, since chiller and dry cooler lead times have stretched with demand.
Which is why “we are adding liquid cooling” is not the same as a date. A facility that can name the month the loop is commissioned in the specific hall has a plan and a schedule. One that describes the intent is describing a project.
Why does warm water change which facilities qualify?
One consequence of direct to chip is often missed by tenants and it affects cost.
Because cold plates work with warm water, a GPU liquid cooling facility can frequently reject heat to outside air through dry coolers for much of the year rather than running chillers. That is free cooling, and in a cool climate it can cover most of the year. Facilities in Quebec publish free cooling for roughly 80 percent of annual hours on this basis, and cool-climate sites elsewhere in North America make similar claims.
Two things follow for a tenant. A deployment specified to accept warm water at the rack inlet, 40 to 45 degrees rather than 25, qualifies at more facilities and costs less to operate, because the building is not running compressors to serve it. And the operator’s power usage effectiveness on that hall is meaningfully better, which shows up in the power line of the invoice if the lease passes cooling cost through.
A requirement that specifies the coldest water the hardware can accept rather than the warmest it tolerates is narrowing its own facility pool and raising its own bill. The specification should say the warmest.
What this page cannot tell you
It cannot tell you which specific facility has a commissioned loop in the hall you would occupy this quarter, because that is a live search and it changes monthly. It cannot tell you what an operator will commit to on water temperature or flow rate in a contract, because that is negotiated. It cannot tell you whether your hardware vendor’s cold plate specification matches a given facility’s plant, because that requires both documents. And it cannot tell you whether a facility that says it can serve you actually can, because that is what the five questions are for.
Those are the placement. This page is the framework it uses.
How the practice works on liquid-cooled requirements
Placement.
You give us the requirement: platform, rack count, density, date, markets. We take it to operators who can plausibly serve it, ask the five questions of each, and bring you the ones that answer with numbers. Then we run the process to a signed contract, including the terms that only matter for liquid cooling: loop commissioning date as a delivery obligation, water temperature and flow as service levels, CDU responsibility stated. The operator you choose pays us. You pay nothing.
Consulting, for developers and operators.
The other side of the same question is what a building can serve. What density a site supports once the floor, the loop and the electrical path are accounted for, what a specification forecloses once drawings lock, and which tenants a facility built to a given spec would attract. That is the consulting side, paid by the client at a fixed fee, and it serves the people building capacity rather than the people leasing it.
Frequently Asked Questions
What is direct to chip cooling?
A liquid cooling method where a cold plate is mounted directly on the processor, coolant flows through it to absorb heat, and the warm coolant leaves the rack through a loop to a coolant distribution unit that transfers the heat to the facility water system. It removes 70 to 80 percent of rack heat into liquid, with the remainder handled by air. It is required for every current rack-scale GPU platform and is the only high density cooling method that serves densities above roughly 70 kilowatts per rack.
What is the difference between direct to chip cooling and immersion cooling?
Direct to chip keeps hardware in a standard rack and cools the processors through cold plates connected to a coolant loop. Immersion submerges the entire server in a tank of dielectric fluid. Immersion reaches higher densities and is used in specialized deployments, but it requires different hardware handling and a different floor layout, and current rack-scale GPU platforms do not ship in immersion form. For tenants deploying those systems, direct to chip is the only option.
What is a coolant distribution unit?
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 exist at three levels: in-rack, which frequently ships with rack-scale platforms and the tenant brings; in-row, serving a row of racks and usually supplied by the facility; and facility-level plant units serving a hall. Which party supplies which level has to be agreed before hardware ships.
What is cold plate cooling?
Cold plate cooling is the component-level mechanism inside direct to chip. A machined metal block with internal microchannels mounts on the processor in place of a heatsink, and coolant flowing through the channels removes heat by conduction. Cold plates arrive fitted on rack-scale GPU systems; the tenant does not select or install them. What the tenant has to establish is that the facility can supply the inlet temperature and flow rate the plates are designed for.
What water temperature does direct to chip cooling need?
Current rack-scale platforms accept inlet water at 30 to 45 degrees Celsius, with the range moving warmer each generation. Older designs wanted 20 to 25. The warmer the tolerated inlet, the more facilities qualify and the cheaper the building runs, because warm water can often be cooled by outside air without chillers. A requirement should specify the warmest water the hardware accepts rather than the coldest.
What flow rate does a liquid-cooled GPU rack need?
For a fully populated rack-scale cabinet at 140 to 160 kilowatts, on the order of 130 to 150 liters per minute. The figure scales with density and with the temperature rise the plates are designed for. What matters is that the facility can sustain that flow to every rack in the deployment simultaneously, not to a single demonstration cabinet.
How much does a rack-scale GPU system weigh?
A fully populated GB200 or GB300 NVL72 cabinet runs approximately 1,360 kilograms, about 3,000 pounds. Standard data hall floors are often specified at 250 pounds per square foot distributed load; rack-scale deployments generally want 300 and above, and the concentrated load at the rack footprint matters as much as the distributed figure. Floor loading should be confirmed in writing before hardware is ordered.
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, longer where those also have to change. The work involves plant, pipework, in-hall distribution, CDUs and often structural reinforcement, each with permits and lead times. It is a construction project rather than a provisioning task, and the loop is frequently commissioned months after the building itself is ready for service.
Can a rear door heat exchanger cool a GPU rack?
For single-server GPU deployments at up to roughly 50 to 70 kilowatts per rack, yes. For rack-scale systems from Blackwell onward, no. Those platforms ship with cold plates fitted, frequently with in-rack CDUs depending on the integrator, and require a direct to chip loop from the facility. A facility offering rear door heat exchangers as its liquid cooling answer for a rack-scale requirement has not understood the requirement.
What should a lease say about liquid cooling?
Five things beyond a conventional agreement: the date the loop is commissioned in the specific hall, as a delivery obligation distinct from the building date; the water temperature at the rack inlet, as a service level; the flow rate per cabinet; which party supplies the CDU at each level; and the floor load, in writing. The data center lease guide covers how those sit alongside the terms every lease carries.
Related reading
- The wider picture: the AI and GPU colocation guide for which markets hold rack-scale capacity and what the platforms require, and the wholesale colocation guide for the commercial side of multi-megawatt leases.
- The contract: the data center lease guide, including the five terms a liquid-cooled lease has to carry.
- For developers and operators: the consulting practice, powered shell and build to suit, for what a building can be made to serve.
- Markets: Northern Virginia, Silicon Valley, Chicago, Dallas, Atlanta. For how placement works, see how it works and why companies use an independent advisor.
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 which are describing a portfolio, and you will have it within 24 hours. If no facility on your timeline can serve your density, we will tell you that, because it is a real answer and it is the one most searches at this tier arrive at.
North American and international coverage. Paid by the operator you choose. We do not sell cooling equipment.
