What Does a Data Center “Facility Standard” Actually Require?

The phrase “facility standard” does not, by itself, identify a rule that every data center must follow.

The useful distinction is what a document covers, whether it offers guidance or has been specified for a project, and what must still be checked locally.

Data center fundamentals series

Does “facility” tell you which rules apply?

No. IBM’s “What Is a Data Center?” describes a data center as a physical room, building, or facility that houses IT infrastructure for applications and services and stores and manages their data. That explains what the facility is; it does not identify a design requirement or a law. To interpret a claimed “facility standard,” start with the equipment and work the document actually addresses.

IBM — What Is a Data Center? | IBM

What can a document’s name lead you to assume?

These common readings confuse a document’s title or publisher with its scope:

“A government guide must be mandatory for every data center”

What the document addresses

The U.S. Department of Energy’s July 2024 revision of “Best Practices Guide for Energy-Efficient Data Center Design” offers suggestions for energy-efficient design across IT systems and environmental conditions, air management, cooling, electrical systems, and heat recovery

What it does not establish

Whether a particular site has a legal obligation

“A standard must cover the whole building”

What the document addresses

The Fiber Optic Association (FOA) describes its standards as guidelines for designing, installing, and testing fiber-optic cable plants

What it does not establish

Whether other data-center systems meet their requirements

“Following one document proves the facility complies”

What the document addresses

DOE’s design guidance and FOA-1’s method for testing loss in an installed fiber-optic cable plant concern different work

What it does not establish

Whether requirements in the other field have been met

Even terms *within* a guide need care. In the ASHRAE environmental ranges discussed by DOE, the recommended range is an operating target intended to support energy efficiency and high reliability. The allowable range describes boundaries tested by equipment manufacturers for functionality, not a boundary that guarantees reliability. Neither label, on its own, establishes a legal limit for a particular site.

The Fiber Optic Association — FOA Standards

U.S. Department of Energy — Best Practices Guide for Energy-Efficient Data Center Design

What changes when a project specifies a standard?

An industry guideline and a requirement named in project documents are different things. FOA gives an example of citing FOA-1 in a statement of work, request, or contract to specify loss testing of an installed fiber-optic cable plant. The question then becomes which standard was cited for which task—not whether the word “standard” appears somewhere in the paperwork.

**Hypothetical example:** Suppose a data-center cabling contract names FOA-1 as its test method. That reference alone does not supply the project’s acceptable loss result. FOA says test specifications, referenced standards, and acceptable results based on a design-stage loss-budget analysis should be set out in project paperwork, with the required test methods agreed in advance.

Installation scope is a separate question from that test method. FOA describes its installation standard as accounting for differences such as premises versus outside-plant and underground versus aerial work; it excludes submarine cables and allows relevant sections to be adapted for a project’s scope of work.

Reading the cited task and the project’s acceptance terms is therefore more useful than treating a standard’s name as a complete specification.

The Fiber Optic Association — The FOA Reference For Fiber Optics - Fiber Optic Network Design

How do you determine whether it is a legal obligation?

These documents cannot settle legal applicability for a particular data center. In the fiber-cabling context, FOA notes that the route and installation location are affected by local building codes and laws, and that a project may require permits, inspections, or locally required licenses. That is an illustration of why location matters, not a legal checklist for every data-center system.

The defensible next step is to identify the site’s jurisdiction, the work and route at issue, the applicable local requirements, and the documents specified for that project. Until those are known, keep three statements separate: “the guide recommends it,” “the project requires it,” and “the law requires it.”

Sources

• U.S. Department of Energy — Best Practices Guide for Energy-Efficient Data Center Design
• IBM — What Is a Data Center? | IBM
• The Fiber Optic Association — The FOA Reference For Fiber Optics - Fiber Optic Network Design
• The Fiber Optic Association — FOA Standards

What Is an AI Accelerator? A GPU Is Not a Model or a Cluster

An AI accelerator is hardware that helps execute AI computations; a GPU is one example.

A trained model, the work of training or inference, and the server or cluster running that work are different things. A spam-filter example and a misconception map make those boundaries easier to recognize.

Data center fundamentals series

What does an AI accelerator do?

An AI accelerator is computing hardware used to speed up AI operations. A GPU is one example: it can perform many operations concurrently and is used for the matrix and vector computations common in AI tasks, including training and running models. That describes a computing role, not a particular performance result or a requirement that every AI system use a GPU.

GPUs are not the only examples. TPUs are designed to accelerate tensor computations, while NPUs are designed for neural-network computations. These are hardware options, not a list of parts that every system must contain.

IBM — What Is an AI Data Center? | IBM

IBM — What is AI Inference? | IBM

IBM — What is AI Infrastructure? | IBM

If a GPU runs a model, which part has learned?

Consider a hypothetical email spam filter. Fitting its model to example emails is training. In a supervised-learning approach, the model’s predictions are evaluated and its parameters are adjusted to reduce error. Using the trained model to classify a new email is inference: an ordinary inference pass produces an output without itself updating those parameters.

The learned parameter values are not the GPU. They belong to the model; a GPU is hardware that can execute the computations involved. Nor are training and inference permanent labels for different machines. GPUs can be used for either kind of work, and a model already serving predictions can later be fine-tuned. That does not mean each prediction is a training step.

What changes when the description moves from GPU to cluster?

These terms mark different boundaries. The map helps separate what does the computing from what is learned, what work is being done, and where it runs.

“The GPU is the trained model.”

Useful distinction

The GPU is hardware; the model uses learned parameter values during inference.

“Training and inference are types of accelerator.”

Useful distinction

They name work: training adjusts model parameters, while inference uses a trained model on new input. The same kind of hardware can be used for both.

“A GPU-equipped server is a cluster.”

Useful distinction

A GPU is a compute component. A server is a larger system within infrastructure that also supports data movement and storage; a cluster comprises servers. If work exceeds one GPU’s capacity, it can be divided across multiple processors, but not every workload needs that arrangement.

“A cluster is an AI data center.”

Useful distinction

AI infrastructure includes hardware and software, with compute, network and storage resources. An AI data center is the facility housing such infrastructure and supplying the power and cooling it needs.

“Inference always runs on a data-center GPU.”

Useful distinction

Inference can also run on an end user’s device, subject to that device’s compute capacity; that does not imply the device has a GPU.

When you encounter the phrase “AI accelerator,” ask which hardware performs the computation, which model and task are involved, and where the task runs. The accelerator’s name alone cannot tell you whether the work is training or inference, or whether it needs one server, a cluster or neither.

Sources

• IBM — What Is an AI Data Center? | IBM
• IBM — What is AI Infrastructure? | IBM
• IBM — What is AI Inference? | IBM

Related reading

• What Is a GPU Cluster? A Worked Example of Servers, Clusters, and AI Data Centers

What Does Simple Payback Mean for a Data Center Efficiency Upgrade?

Simple payback estimates how long assumed annual net savings would take to recover an upgrade’s additional upfront cost.

A hypothetical calculation shows what goes into that estimate, followed by the distinctions it cannot settle: changing savings, whole-life cost, and other operating outcomes.

Data center fundamentals series

What costs does the payback calculation compare?

Simple payback is the additional upfront cost of an efficiency improvement divided by its positive annual net savings. If those savings stay constant, the result is the number of years it would take to recover the additional cost. The costs and savings must be expressed in the same currency, with savings stated per year.

This is a calculation based on assumptions, not a forecast supplied by the U.S. Department of Energy’s Best Practices Guide for Energy-Efficient Data Center Design.

First define the baseline: both options should deliver the same IT work and meet the required reliability conditions. Otherwise, a change in service could be mistaken for an efficiency saving. The upfront figure is not everything previously spent on the data center; it is what the proposed improvement costs *in addition to* the baseline option.

Its installation scope matters. For example, the DOE guide notes that adding an air-side economizer to an existing facility can be difficult because of the large duct installations involved. That does not mean every upgrade needs ductwork.

The denominator is money saved per year, not electricity saved in kilowatt-hours. Estimate the difference in annual costs under the same conditions, then subtract any additional recurring costs.

IBM notes that liquid cooling requires specialized infrastructure and maintenance; a proposal involving it would need to place applicable costs in the upfront or recurring category. Avoid counting a benefit twice: if reduced cooling or UPS loads are already reflected in the total electricity-bill difference, do not add them again as separate savings.

U.S. Department of Energy — Best Practices Guide for Energy-Efficient Data Center Design

IBM — What Is a Green Data Center? | IBM

How could a hypothetical upgrade produce a five-year result?

This example is entirely hypothetical. None of its amounts are measurements or projected results for a particular facility or technology. Assume both options provide the same IT work and meet the required reliability conditions:

• Baseline annual electricity bill: KRW 8 million.
• Upgrade annual electricity bill: KRW 5 million, for assumed electricity savings of KRW 3 million per year.
• Additional annual maintenance for the upgrade: KRW 1 million, leaving net savings of KRW 2 million per year.
• Additional upfront cost of the upgrade: KRW 10 million.

If net savings remain KRW 2 million every year, simple payback is **KRW 10 million ÷ KRW 2 million per year = 5 years**. The answer depends on both the stated cost boundary and the assumption that annual net savings remain constant.

Which conclusions would the five-year figure not support?

The calculation answers a narrow recovery-time question. This misconception map separates that answer from claims it cannot establish:

• **“Five years is a guaranteed recovery date.”** Actual operating loads, future loads and part-load conditions can affect savings. The DOE guide calls for attention to those conditions when selecting equipment. If net savings change from year to year, examine their annual cumulative total to see when, if ever, it reaches the additional upfront cost; dividing by one fixed annual amount no longer describes that path.
• **“Zero or negative net savings still give a payback time.”** With a positive additional upfront cost, annual net savings of zero or less cannot recover it through those savings in a positive, finite time. Division by zero does not mean zero years, and a negative quotient does not mean the cost has been recovered.
• **“A lower PUE reveals the payback period.”** The DOE guide defines PUE as total annual facility energy divided by annual IT-equipment energy. That energy ratio contains neither the additional upfront cost nor annual net monetary savings.
• **“Payback settles the investment decision.”** It says when assumed savings recover the additional cost, not what costs and savings look like afterward or across a common service period. The DOE guide also distinguishes initial from life-cycle costs and calls for considering total cost of ownership. Comparing whether to keep or replace equipment over such a period is a different question.
• **“Payback also establishes reliability, water or carbon outcomes.”** Reliability is a separate design requirement, and the DOE guide discusses water and carbon metrics alongside energy metrics. None automatically becomes an annual monetary saving; a proposal would have to identify an actual cost change before including it in that denominator.

What makes the estimate useful for a real proposal?

Keep the hypothetical five years out of any claim about actual performance. For a real proposal, state the baseline, the additional upfront costs and exactly which changes make up annual net savings. Then use relevant energy metering to test the energy assumptions as operating data become available.

The DOE guide says sufficient metering is needed for ongoing energy management and describes trending and retaining measured values to obtain annual energy totals.

Simple payback can therefore frame a useful question: *How long would these specified savings take to cover this specified extra cost?* Checking the assumptions makes that answer more informative; comparing costs over the full service period and assessing required reliability remain separate parts of the decision.

Sources

• IBM — What Is a Green Data Center? | IBM
• U.S. Department of Energy — Best Practices Guide for Energy-Efficient Data Center Design

Related reading

• Keep or Replace Data Center Equipment? Set the Total-Cost Comparison Boundary First

How to Evaluate a Data Center CDU Against IT and Facility Requirements

A CDU’s name does not establish that it can connect a particular server to a particular facility cooling loop.

This hypothetical cold-plate proposal shows which conditions to compare on each side, which labels cannot substitute for specifications, and when to select, exclude, or defer a decision.

Data center deep-dive series

What are the two interfaces a CDU must serve?

Evaluate the IT-side supply and the facility-side connection separately. The U.S. Department of Energy’s *Best Practices Guide for Energy-Efficient Data Center Design* explains that, in many liquid-cooling approaches, a cooling distribution unit (CDU) interfaces with the facility cooling loop while supplying liquid at conditions appropriate for the IT equipment. That description establishes a role, not compatibility with a specific installation.

Consider an explicitly hypothetical proposal: servers use cold plates that transfer heat from chips into flowing liquid, and a candidate CDU is offered to connect them to a facility loop. Cold plates are one form of direct liquid cooling; this example does not describe every liquid-cooling arrangement. Neither the server requirements, the facility conditions, nor the candidate CDU’s operating limits have been supplied.

U.S. Department of Energy — Best Practices Guide for Energy-Efficient Data Center Design

Which conditions belong in the comparison?

Use the table to request evidence for the same loads and operating conditions on both sides. Every status is **evidence missing** in this hypothetical proposal; it is not a finding that a real product fails.

IT-side liquid

IT equipment or facility evidence needed

Equipment-required fluid type and chemistry

Candidate CDU evidence needed

IT-side supply fluid and supported chemistry

Hypothetical proposal status

Evidence missing; “supports liquid cooling” does not establish compatibility

IT-side operation

IT equipment or facility evidence needed

Equipment-required supply temperature, pressure, flow, and heat load

Candidate CDU evidence needed

IT-side operating range at the required flow and load

Hypothetical proposal status

Evidence missing; no common setpoint can be assumed

Changing loads

IT equipment or facility evidence needed

Initial and future heat loads, including part- and low-load operation

Candidate CDU evidence needed

Operating range under those conditions

Hypothetical proposal status

Evidence missing; a maximum-load figure alone is insufficient

Facility connection

IT equipment or facility evidence needed

Actual loop supply and return temperatures, available flow and pressure, and conditions as load changes

Candidate CDU evidence needed

Required facility-side conditions and heat-transfer capability

Hypothetical proposal status

Evidence missing; an IT-side supply claim does not establish facility compatibility

Boundary between sides

IT equipment or facility evidence needed

Equipment and facility connection and fluid requirements

Candidate CDU evidence needed

Schematic and specifications showing whether the fluids are the same and how the circuits are arranged or separated

Hypothetical proposal status

Evidence missing; the name “CDU” does not establish a heat-exchanger or circuit arrangement

Remaining air cooling

IT equipment or facility evidence needed

Heat captured by liquid and heat left to room air

Candidate CDU evidence needed

Liquid-side heat load the CDU is proposed to handle

Hypothetical proposal status

Evidence missing; CDU capacity does not establish that room air cooling can be removed

The IT-side liquid might be treated water, a glycol-based solution, or a dielectric fluid, depending on the technology; that list does not mean the hypothetical cold plates accept all three. Some liquid-cooled systems also leave part of the IT heat load to air cooling. Request the equipment’s heat split before treating the CDU as a replacement for room cooling.

Can a temperature label fill a missing specification?

No: first identify what the temperature describes. The DOE guide’s example of a medium-temperature chilled-water loop supplied at **55°F or higher** is design guidance, not the hypothetical facility’s measured temperature or a universal CDU requirement. The facility row needs that facility’s own operating conditions.

The guide also describes ASHRAE W17, W27, W32, W40, and W45 as classes whose numbers indicate upper limits, in degrees Celsius, for **server-side liquid supply temperature**. A W label alone does not establish fluid or pressure compatibility, or the facility-side chilled-water conditions.

By contrast, the A1–A4 recommended and allowable temperature envelopes discussed in the guide concern **air-cooled equipment inlet air**. The recommended envelope guides efficient operation while maintaining reliability; the allowable envelope describes boundaries tested for equipment functionality, not reliability. Neither is a CDU liquid-supply specification.

Would the same comparison apply to immersion or prove an efficiency gain?

The questions about interfaces still matter, but this cold-plate table cannot be copied over as an immersion-system specification. The DOE guide distinguishes single-phase immersion, in which a CDU pumps dielectric fluid around electronics, from two-phase immersion, in which vapor transfers heat to an exchanger, condenses, and returns in a passive cycle.

“Immersion cooling” alone therefore does not identify the proposed IT-side circuit or CDU arrangement.

Heat rejection is another separate question. The guide explains that warmer cooling-water conditions can facilitate a water-side economizer and that higher W classes may permit dry-cooler use. Whether either option works depends on the particular facility loop and outdoor conditions. It is not an automatic efficiency benefit of the candidate CDU.

When can the proposal be selected, excluded, or held?

For each table row, record **documented match**, **documented mismatch**, or **evidence missing**. If the equipment requirements, facility-loop data, or CDU operating ranges for either side are absent, hold the compatibility decision. A documented mismatch is grounds to exclude the proposal unless a design change is separately evaluated; missing evidence is not itself proof of a mismatch.

Even matching documents do not complete installation approval. Plan subsequent operating checks around the temperatures, flows, and other values used in the comparison, including the accuracy and calibration status of the instruments. Obtaining the missing documents and comparing them under the same operating conditions turns a product description into a testable interface decision.

Sources

• U.S. Department of Energy — Best Practices Guide for Energy-Efficient Data Center Design

Related reading

• Immersion Cooling: What a Hypothetical Rack Example Helps You Separate
• Air Cooling vs. Liquid Cooling: Define the System Boundary Before Comparing Efficiency

Where Should a Shared Document Live? A Data Governance Worked Example

Choosing where a document is stored does not decide who may use it, how long it should remain, or who manages those decisions.

Follow one hypothetical project draft from its storage choice through access, retention, and responsibility.

Data center fundamentals series

Does choosing a storage location settle the policy?

No. Suppose a project team needs to share a meeting draft. A shared file store on the organization’s network and an off-site cloud file service are both possible ways to make files available for collaboration. For this hypothetical example, the team chooses the cloud service—not as a recommendation, but to see what decisions remain.

A data center may restrict entry to its building and equipment areas. Those physical controls do not decide who may open this particular draft or when it should stop being kept.

The Fiber Optic Association — The FOA Reference For Fiber Optics - Data Centers -

IBM — What Is Cloud Storage? | IBM

IBM — What Is Data Storage? | IBM

Who may read or change the draft?

Assume team members need to edit the draft, while an external reviewer only needs to read it. The team must distinguish viewing from editing and name someone to approve the reviewer’s access. Whether those permissions can be configured separately depends on the chosen service and must be checked there.

Access also needs a review point. When the external review ends, the person responsible can decide whether the reviewer still needs access, then check the actual permission and how to change it. The team should not assume that the service removes access automatically.

Why keep the draft, and what happens to copies?

The team first needs to decide whether this is a working draft needed only for review or a record that must remain afterward. That purpose informs its retention condition; a general description of data storage cannot supply a legal retention period or deletion date for this hypothetical document.

If the organization uses backups or immutable snapshots, it must consider those copies separately. A backup provides a copy for recovery after data loss. An immutable snapshot is a point-in-time copy protected from change or deletion for a set or indefinite period.

Neither type of copy can be assumed to exist for this draft, and deleting the original cannot be assumed to delete its copies. The team must check its actual storage and retention settings.

Does the cloud provider take over these decisions?

Not under the general responsibility split IBM describes: the provider manages and secures the underlying cloud infrastructure, while the customer is responsible for securing its data and applications within it. In this example, the team still needs an internal owner for sharing approvals, permissions, and retention decisions.

The exact division of duties must be checked against the service and contract rather than inferred from where the file is stored.

What would the team record for this document?

The hypothetical team can turn the discussion into a short decision record. The middle column shows a proposed decision for this example, not a setting or rule that applies to every organization.

Storage

Hypothetical team record

Put the draft in the chosen cloud file service

Check before using it

Confirm the service and its available settings

Access

Hypothetical team record

Team members edit; the external reviewer reads; a named team owner approves access

Check before using it

Confirm that these permissions can be set and who currently has access

Access review

Hypothetical team record

Reconsider the reviewer’s access when the review ends

Check before using it

Confirm how to inspect and change that access

Retention and copies

Hypothetical team record

Decide whether the draft is needed after review; address any backups or snapshots separately

Check before using it

Check applicable obligations and the service’s retention and deletion behavior

Responsibility

Hypothetical team record

Assign an internal owner for sharing, access, and retention decisions

Check before using it

Check the service and contract for the actual provider–customer split

The useful result is not a universal retention period or a preferred storage location. It is a set of decisions the team can assign and verify: where the file lives, who can use it, why it remains, what copies may remain, and who is accountable for each answer.

Sources

• The Fiber Optic Association — The FOA Reference For Fiber Optics - Data Centers -
• IBM — What Is Data Storage? | IBM
• IBM — What Is Cloud Storage? | IBM

Related reading

• Physical Security vs. Data Security in a Data Center