The 5-step pipeline.
How energy turns into a deployed, financed, offtake-matched cluster, and what SLYD does at every step.
How SLYD works →Platform
SLYD Cloud
New and recovered GPU systems.
NVIDIA and AMD systems through documented manufacturer and qualified channel supply, with financing and deployment coordinated on the same platform.
Explore GPU hardware →By accelerator
Infrastructure and services
Compute, hardware, and power in one book.
Browse available GPU capacity by accelerator, configuration, region, and price, or bring supply to qualified demand.
Open marketplace →Compute
Bring supply
For buyers
Start with an indicative structure.
Tell us deal size, structure, and offtake. Any range is preliminary and subject to underwriting, diligence, and documentation.
Open Configure →For lenders
GPU market trends and deployment playbooks.
Infrastructure best practices, hardware comparisons, and industry analysis from the SLYD team.
Read the blog →Company
Cooling infrastructure for AI systems
Cooling method is a property of the server and the site, not of the accelerator generation. This page compares the four architectures on what each actually asks of the building, and on how to read the efficiency claims that surround them.
What is AI data center cooling?
AI data center cooling removes heat from GPU servers, racks, and supporting equipment using facility- and system-specific air or liquid architectures. The correct design depends on server approval, rack density, heat capture target, water temperatures, flow, facility loops, redundancy, water quality, serviceability, climate, and deployment constraints.
What this page is, and what it is not
Published examples are planning illustrations, not site engineering. Cooling design and water systems require qualified engineering, manufacturer compatibility review, and project-specific commissioning.
This page exists to help a buyer ask the right questions early and to make manufacturer claims readable. It does not contain a design, and no figure on it should be carried into one.
Cooling is configuration specific, not generational
It is commonly said that a given accelerator generation requires liquid cooling. Manufacturer specifications do not support that as a general rule. AMD publishes Instinct accelerator cooling as Passive OAM for MI300X, MI325X, and MI350X and as Passive and Active for MI355X. HPE publishes the same 8-way GPU server at 5U for direct liquid cooling and 6U for air. NVIDIA describes GB300 NVL72 as a fully liquid-cooled rack-scale architecture, which is a statement about that platform rather than about Blackwell generally.
The useful version of the question is narrower and answerable: what cooling method is this exact server approved for, and can this exact building supply it. Everything else on this page follows from those two answers.
Four ways to move the heat
Compared on what each asks of the facility and what it costs operationally, rather than ranked. The right choice is the one the servers are approved for and the building can support.
| Architecture | How it works | Facility requirement | Operational impact |
|---|---|---|---|
| Air with containment | Conditioned air through the rack, with hot or cold aisle containment preventing recirculation | Air handling capacity, containment, and adequate airflow at the rack face | Familiar operations, standard service access, well understood failure modes |
| Rear-door heat exchanger | A liquid-cooled coil at the back of the rack removes heat before it enters the room | Facility water to each rack, plus the existing air path through the servers | Servers stay air-cooled internally, so server service access is unchanged |
| Direct-to-chip liquid | Cold plates on the hottest components carry heat into a liquid loop | Facility water loop, CDUs, heat rejection, water quality management, leak detection | New maintenance discipline for the loop; residual heat still leaves as air |
| Immersion | Equipment is submerged in a dielectric fluid | Purpose-built tanks, fluid handling, heat rejection, and facility design around them | Substantially changed service practice, component compatibility, and warranty terms |
Equipment must be approved by its manufacturer for the cooling method used. That approval is specific to a model and configuration, and using equipment outside it can affect warranty and support.
What a water loop actually commits the site to
Liquid cooling is often treated as an equipment choice. It is a facility programme, and the items below have their own lead times and their own owners.
Supply temperature and flow
Servers are approved for a defined water temperature range and flow rate. The facility has to hold those under worst-case conditions, not just typical ones, and the design has to state what worst case means for that site.
Coolant distribution and redundancy
CDU sizing, placement, and redundancy topology, plus what happens to the load when one unit is taken out for service.
Heat rejection
Where the heat ultimately goes, which depends on climate, water availability, and site constraints, and which is frequently the item with the longest lead time in the loop.
Water quality and treatment
Chemistry, filtration, and monitoring to protect cold plates and CDUs. Neglected water quality is a common cause of expensive failures well after handover.
Leak detection and response
Detection at the rack and the loop, a defined response procedure, and a tested shutdown path. This has to be commissioned, not just installed.
Maintenance model and ownership
Who maintains the loop, what spares exist, and how escalation works. Liquid cooling adds a system that needs an owner, and that owner should be named before commissioning.
How to make an efficiency claim usable
Cooling marketing is full of percentages. Most of them are true for the configuration they were measured on and meaningless anywhere else. These questions separate the two.
- Which exact system and configuration was measured
- What rack density was in place
- What facility water temperatures were supplied
- What climate and heat rejection design applied
- What baseline the improvement is measured against
- Whether the figure is "up to" or typical
- Who measured it and when
- Whether the same conditions can exist at this site
SLYD publishes no PUE target, efficiency percentage, energy saving figure, rack density, water temperature, noise level, or control accuracy value. Where a manufacturer publishes one for a named configuration, this site reproduces it with that configuration attached, as on the GIGABYTE and Supermicro pages. Those remain manufacturer claims about manufacturer equipment.
Choosing from requirements rather than rankings
What is the server approved for
If the chosen system only ships liquid cooled, the decision is made and the question becomes whether the site can support it. If both are available, the choice is genuinely open.
What can the building supply today
Existing air capacity, existing water, and what could be added within the project timeline. This frequently settles the question on its own.
Is lower density acceptable
Spreading the same equipment over more racks reduces per-rack load and can keep an air-cooled room viable, at the cost of floor space and cabling. Often the pragmatic answer for a first deployment.
Who will operate it
A liquid loop needs an owner with the skills and the spares. If that owner does not exist and cannot be hired or contracted, that belongs in the decision rather than in the risk register.
To scope cooling
- Target systems and their approved cooling methods
- Number of racks and target density
- Existing cooling capacity and type
- Whether facility water is available, and at what temperature
- Site climate and heat rejection options
- Redundancy and concurrent maintainability requirements
- Existing monitoring and building management systems
- Who will own loop maintenance
- Target date
- Known code or environmental constraints
SLYD helps define the cooling requirement and evaluates sourcing paths for the equipment. SLYD does not perform mechanical engineering, water system design, installation, or commissioning, and publishes no monitoring service, maintenance program, or response commitment. That work is performed by qualified professionals identified for the specific project.
Cooling questions
Do GPU servers require liquid cooling?
Not as a rule tied to the accelerator generation. Cooling method is a property of the specific server and the site. Manufacturers publish the same accelerator platforms in both air-cooled and liquid-cooled systems, and AMD publishes Instinct accelerator cooling as Passive OAM for MI300X, MI325X, and MI350X and as Passive and Active for MI355X. What matters is what the exact server is approved for and what the building can supply.
What are the main cooling architectures?
Air cooling with hot or cold aisle containment, rear-door heat exchangers that remove heat at the back of the rack, direct-to-chip liquid cooling using cold plates on the hottest components, and immersion cooling where equipment sits in a dielectric fluid. They differ in the density they support, what the facility must provide, and how serviceable the equipment is.
Does direct-to-chip liquid cooling remove all the heat?
No. Cold plates cover the hottest components, and manufacturers publish heat capture figures for specific named configurations rather than for the technology in general. The remaining share leaves the server as air and still has to be handled by the room. A facility plan that budgets no air-side load because the system is liquid cooled will be wrong.
What does a liquid-cooled deployment require from the facility?
A facility water loop with a defined supply temperature and flow rate, coolant distribution units with a redundancy design, heat rejection, water quality management and treatment, leak detection and response, and a maintenance model with named owners. Those are facility commitments with their own lead times and should be confirmed before the servers are selected.
How should manufacturer efficiency and PUE claims be read?
As results for a specific configuration under specific conditions, which is what they are. A heat capture, PUE, or energy saving figure depends on the exact system, rack density, water temperatures, climate, heat rejection design, and the baseline being compared against. Without all of those stated it is not a planning input, and it does not transfer to a different system or a different site.
When is immersion cooling the right answer?
Rarely, and only when the constraints genuinely point there. Immersion changes serviceability, warranty terms, component compatibility, facility design, and operational practice substantially. Equipment must be approved for immersion by its manufacturer, and that approval is specific rather than general. It should be evaluated against direct-to-chip liquid cooling rather than adopted because it sounds more advanced.
Can an existing air-cooled room take high-density GPU racks?
Sometimes, at reduced density per rack. Spreading the same equipment across more racks lowers per-rack load and can bring a deployment within what an existing room can handle, at the cost of floor space and cabling. Whether that works is a question about the actual room, its airflow, and its containment, answered by a site survey rather than by a rule.
What information is needed to scope cooling?
The target systems and their approved cooling methods, the number of racks and target density, existing cooling capacity and type, whether facility water is available and at what temperature, site climate and heat rejection options, redundancy and maintainability requirements, and the target date.
Scope a cooling requirement
Share the target systems, rack count, existing cooling capacity, and whether facility water exists. Those answers usually settle the architecture before anyone opens a catalog.
Page updated: August 18, 2026