AI data centres are the physical infrastructure that keep digital intelligence services running. Servers and accelerators perform calculations, but those machines also need electricity, cooling, storage, networking and reliable operations. To understand how the system works, follow the complete flow from the user’s request through the computing facility and back to a verified result.
This chapter uses a fictional North Quay Compute Hall to calculate power usage effectiveness (PUE), daily electricity consumption and the difference between IT load and total facility demand. The worked figures show why an improved PUE can coexist with rising electricity use. An independent second facility provides a changed-context test. It fills topic 082 of How Super Intelligence Works.
Super Intelligence is eduKateSG’s editorial umbrella for examining modern AI mechanisms, not a claim that today’s infrastructure houses artificial superintelligence. Throughout this guide, invented examples remain labelled fictional, while current industry facts and forecasts are distinguished from physical observations.
Contents and reading routes
Understand the facility · North Quay packet · Calculate PUE · See the efficiency paradox · Independent practice · Sources
Chapters 1–9: 1. A data centre is a physical system for digital work · 2. Follow one request from user to machine room · 3. Power and energy are different quantities · 4. The complete fictional North Quay Compute Hall packet · 5. Power usage effectiveness has a precise denominator · 6. Calculate the fictional daily electricity use · 7. A lower PUE can coexist with higher total demand · 8. The non-IT share in the new configuration · 9. Cooling is a thermal-management system
Chapters 10–18: 10. Air cooling and liquid cooling are engineering choices · 11. Cooling loads cannot be guessed from PUE alone · 12. Electrical distribution and conversion affect availability · 13. Redundancy is not the same as independence · 14. Networks connect the facility to the user’s work · 15. Storage and backup answer different questions · 16. Capacity planning must consider peak and sustained load · 17. The grid connection sets another boundary · 18. Read the IEA forecast as a projection
Chapters 19–27: 19. Location changes energy and cooling trade-offs · 20. Water use is not inferred directly from electricity · 21. Carbon emissions require a separate calculation · 22. Service reliability is measured through outcomes · 23. Independent transfer: the smaller Gallery Compute Room · 24. Full transfer answers and interpretation · 25. A compact site evaluation checklist · 26. Frequently asked questions · 27. Sources and connected education routes
1. A data centre is a physical system for digital work
A data centre is a facility that houses computing, storage and network equipment together with the power, cooling, monitoring and security needed to operate them. AI services run on digital processors, but the processors are not suspended outside the physical world. They depend on electricity, temperature control, network connectivity, physical maintenance and the availability of replacement parts.
Calling a building an ‘AI factory’ can be a useful metaphor when it produces computational services, yet the actual outputs depend on software, data, model design and users. A building does not become intelligent merely by containing accelerators. The right question is what work the facility can reliably support, under which environmental and operational constraints.
Different sites have different designs. Some are shared colocation facilities, some belong to enterprises, and others serve large cloud providers. The mechanisms below are general, not an assertion that every facility uses the same power architecture, cooling technology, staffing level or national regulation.
2. Follow one request from user to machine room
When a learner asks an AI assistant a question, their device sends a request through networks to an application. The application may select a computing service, load or access model weights, perform inference and return a result through the network. Storage, identity management and logging may run on separate systems. Any one of those stages may affect response time.
The data centre supports part of that route by supplying functioning servers, accelerators, memory, storage, switches and infrastructure. A highly capable processor is no use when a network connection is unavailable or the application is not authorised to access the required data.
A system-level explanation must include the external service boundary. Electrical power to racks, cooling of equipment, storage durability, network capacity, application health and outcome verification are related but distinct aspects of a successful user request.
3. Power and energy are different quantities
Power is the rate at which energy is used; an electrical load can be described in kilowatts. Energy consumed over a period is measured in kilowatt-hours. A load of 200 kW maintained for twenty-four hours uses 200×24 = 4,800 kWh. A larger load sustained for a shorter period might use less total energy.
A facility’s maximum rated capacity does not prove it actually runs at that level all the time. Actual use varies by workload, climate, cooling demand, equipment utilisation and operating conditions. To calculate energy from a rated load, we must explicitly assume it remains constant.
This distinction matters when reading claims about AI electricity. A photograph of a high-capacity facility does not establish measured daily consumption, and a regional electricity-demand forecast is not a meter reading of one specific building.
4. The complete fictional North Quay Compute Hall packet
Our main case is an invented educational facility called North Quay Compute Hall. Assume it has a constant electrical power draw of 200 kW from its IT equipment and a total facility draw of 280 kW over a single twenty-four-hour illustrative day. IT includes the equipment counted by this simplified example; the remaining difference accounts for facility support loads.
The figures are deliberately fictional and not claims about any real data centre. We choose them because they permit transparent calculations of power usage effectiveness and energy. We assume all measurements use compatible electrical boundaries and the same time period.
The student’s task is to calculate the facility’s PUE, the day’s IT energy, the day’s total facility energy and the implied non-IT energy difference. Then a second operating configuration changes both loads and tests why a better PUE does not necessarily mean less total electricity.
5. Power usage effectiveness has a precise denominator
Power usage effectiveness, or PUE, expresses total facility power divided by IT equipment power under compatible measurement boundaries. In the fictional base case, PUE = 280/200 = 1.4. The number is dimensionless because both numerator and denominator use the same power unit.
The Uptime Institute describes PUE as an established metric for facility overhead efficiency. It does not measure how useful the computational work is, how accurate an AI model is, or whether the facility’s electricity comes from low-carbon sources.
Comparing PUE across sites requires care because climate, load, measurement scope and operating conditions differ. A lower number can suggest less facility overhead relative to IT load, but it is not a complete measure of sustainability or computing productivity.
6. Calculate the fictional daily electricity use
Under a constant-load assumption, IT energy is 200 kW×24 hours = 4,800 kWh. Total facility energy is 280×24 = 6,720 kWh. The difference is 1,920 kWh for non-IT facility loads within the simplified measurement boundary.
Dividing 6,720 by 4,800 again gives 1.4, as expected when both loads stay constant over the same period. This provides an independent consistency check using energy rather than instantaneous power.
One must not conclude that every one of the 1,920 kWh belongs to cooling. It is the aggregate difference under the stated definitions and could include other electrical support needs. A proper submetering system is needed to attribute the components.
7. A lower PUE can coexist with higher total demand
Now assume the facility’s IT power grows to 240 kW and total facility power grows to 312 kW. The new PUE is 312/240 = 1.3, better than the original 1.4 on this metric. However the new total power draw is 312 kW rather than 280.
Across an equally constant twenty-four-hour day, the new total energy is 312×24 = 7,488 kWh. That is 768 kWh more than the original 6,720, despite the lower PUE. The IT work and the overall facility demand changed together.
This is one of the most useful distinctions in energy reporting. Improved efficiency per unit of IT load and reduced absolute electricity consumption are separate outcomes. A credible article must say which measure changed and which broader claim remains unsupported.
8. The non-IT share in the new configuration
The revised non-IT power is 312−240 = 72 kW, versus 80 kW originally. That is a smaller overhead power draw despite a higher total facility load. Over twenty-four hours, non-IT energy falls to 72×24 = 1,728 kWh.
The system now has a larger IT workload, lower overhead and higher total consumption. It would be inaccurate to call the change either a universal energy saving or an unqualified regression. Each statement requires the metric and baseline.
The next decision depends on the user’s goal. An infrastructure operator might value lower overhead at higher useful throughput, while a regional grid planner may need the absolute additional 32 kW of facility demand. Both questions are legitimate.
9. Cooling is a thermal-management system
Electrical energy used by computing equipment eventually produces heat that must be managed. Facilities transfer heat away from equipment using combinations of airflow, liquid loops, heat exchangers and other methods, depending on their design. The objective is to maintain suitable operating conditions and reliability for the equipment.
Cooling is not just an accessory attached to a server. Air pathways, temperature sensors, coolant flow, weather and room layout influence how effectively heat is removed. A poorly managed thermal environment can reduce service reliability even when enough electrical power is available.
Do not assume one cooling method is universally superior. Site climate, available water, equipment density, maintenance capability and other trade-offs matter. Comparisons should use measured conditions rather than a photograph or a marketing term.
10. Air cooling and liquid cooling are engineering choices
Air-cooled designs transfer heat using moving air and cooling equipment. Liquid-cooled designs can move heat through coolant systems close to high-density components. Actual facilities may mix approaches. Each requires planning for capacity, fault detection, maintenance and safe operation.
A liquid-cooled rack still sits within a larger facility needing power distribution, networking, pumping, heat rejection and controls. Liquid cooling does not make heat disappear; it changes how the heat is transported. The external environment still receives that energy unless an appropriate recovery arrangement uses part of it.
For students, the key first-principles question is where the heat goes. Following the flow from electronic components to an external sink is more explanatory than memorising brand names for cooling equipment.
11. Cooling loads cannot be guessed from PUE alone
The North Quay example identifies non-IT electrical energy but does not subdivide it into pumps, chillers, lighting, electrical conversion losses and other components. An answer that says ‘cooling used exactly 1,920 kWh’ would invent an attribution not contained in the packet.
To estimate the cooling contribution, the operator would need appropriate submeters, definitions and possibly thermodynamic measurements. It could then compare the contribution across similar operating periods while controlling for weather and IT demand.
This is the same evidence discipline as the site’s Grounding chapter. A calculation can be correct while its proposed interpretation is wrong. The numerical boundary of a PUE calculation must not be expanded beyond the data supplied.
12. Electrical distribution and conversion affect availability
Power reaches computing equipment through electrical distribution infrastructure, protective devices and, in many facilities, arrangements designed to manage interruptions. Each stage has design limits and maintenance requirements. Depending on the architecture, an uninterruptible power supply can bridge short disruptions while other measures respond.
No generic promise of zero downtime follows from the phrase ‘redundant power’. Components can share hidden dependencies, and maintenance can create new failure states. A resilient design identifies critical paths, potential common failures and recovery procedures.
Readers can distinguish available electricity from usable delivered computing service. A server with uninterrupted power may still be unreachable because of a network failure; the same user-visible outcome can result from entirely different mechanisms.
13. Redundancy is not the same as independence
Two power feeds may look redundant but depend on the same upstream substation. Two network paths may share one trench. Two cooling pumps may require the same controller. Duplicate equipment reduces some risks while leaving common-mode failures.
A useful engineering assessment maps each supposedly independent component back to its real dependencies. The phrase ‘N+1’ describes an approach to spare capacity for a defined component set; it is not a universal guarantee that every other dependency can fail harmlessly.
Learning to detect shared dependencies transfers directly into multi-agent systems and school project management. Adding a second copy does not automatically create an independent source of reliability.
14. Networks connect the facility to the user’s work
Model inference may happen on a powerful accelerator, but the answer still travels through application servers, switches and external networks. Congestion, failed links or routing changes can affect latency and availability even when the GPU completes the numerical operation normally.
A data-centre design therefore balances compute, storage and network capacity. An application that requires large model weights or datasets to move between machines may depend heavily on internal network performance. Data movement and coordination can dominate an otherwise fast calculation.
The previous GPUs and Accelerators chapter uses a fictional timing laboratory to illustrate why fast computation alone is insufficient. The same principle scales from one accelerator to a facility.
15. Storage and backup answer different questions
Primary storage supports active data access. Backups aim to preserve recoverable copies under specified failure conditions. Replication can support availability, but an accidentally deleted record may be copied quickly to every replica. A backup that cannot be restored provides little practical reassurance.
Facilities and software operators need retention, access control, recovery testing and clear responsibilities. The relevant datasets may include model parameters, application databases, logs and training records. Their sizes and sensitivity differ.
Security and privacy duties do not disappear at the hardware boundary. A physical facility may store encrypted data whose use still depends on appropriate access, purpose and responsible organisational policy.
16. Capacity planning must consider peak and sustained load
A system may have enough average capacity while failing during simultaneous peaks. A school learning platform could see a burst of requests near an examination deadline, while normal traffic remains modest. Planning should examine peak demand, failure scenarios and how much of the load can be queued or shifted.
Compute, power, network and cooling limits may become bottlenecks at different times. The operator should identify which constraint reaches its limit first and what user-facing degradation follows. A service queue, request rejection or slower response can be preferable to an uncontrolled physical overload.
Do not call a facility’s nameplate power rating its measured consumption. Capacity, utilisation and actual delivered service require distinct observations.
17. The grid connection sets another boundary
A data centre needs an electrical grid or another reliable supply arrangement. Local transmission capacity, interconnection processes, operating rules and regional demand can constrain how quickly large new loads are served. Facility engineering therefore intersects with energy-system planning.
The International Energy Agency’s Energy and AI report, published in April 2025, studies how digital infrastructure may change electricity demand and how grids and generation may respond. Its forecasts are scenario-based, not claims that every proposed data centre will be built on schedule.
An accurate article keeps the local question separate from the global trend. A worldwide projection cannot tell an individual operator which grid connection has been approved, and an approved connection does not prove how intensively the servers will be used.
18. Read the IEA forecast as a projection
The IEA’s April 2025 base case projected global data-centre electricity consumption reaching around 945 TWh by 2030, more than twice the level of 2024. This is a modelled scenario estimate, not a meter reading for 2030 or proof that the actual future outcome must match it.
AI is an important driver in the analysis, but data centres also serve other digital work. It would be inaccurate to label all projected data-centre electricity as AI-only consumption. Readers should distinguish electricity used by AI-related accelerated servers from other equipment and infrastructure.
A responsible forecast discussion includes the publication date, scope and uncertainty. Different technology trajectories, policies, adoption rates and infrastructure constraints can shift future demand.
19. Location changes energy and cooling trade-offs
Climate, grid capacity, network connectivity, water availability and planning rules vary by location. A site that performs well in a cool environment may need a different thermal strategy in a tropical one. The useful comparison keeps the task and operating conditions visible rather than judging a whole region by one facility metric.
Geographic diversification can improve some kinds of resilience while introducing longer network routes and new operational dependencies. A multi-site application also needs to decide how data and users are routed under applicable requirements.
The relationship between computing and place is part of why the earlier Chips, Memory and the Geography of Computing article matters. Digital capability has a physical geography.
20. Water use is not inferred directly from electricity
Some thermal systems consume water and others use different methods or combinations. Water-use figures depend on climate, technology, operational choices and measurement boundaries. A lower PUE does not by itself establish lower water consumption.
Where water is relevant, readers should look for a properly defined water metric, metered quantities and the relationship to cooling load and season. Electricity and water are different resources with different local scarcity and environmental concerns.
A responsible operator considers the effects on the surrounding community rather than treating every design decision as a contest for the smallest energy ratio.
21. Carbon emissions require a separate calculation
Two facilities can consume the same kilowatt-hours while drawing electricity from different generation mixes. Emissions accounting requires information about the electricity source, timing and methodology, and potentially other sources within the chosen boundary. PUE is not a direct carbon-emissions metric.
An operator may procure energy certificates or enter contracts, but the accounting and physical grid effects need careful interpretation. This chapter does not infer a site’s emissions from a general statement about national energy shares.
Students can carry the same distinction from science: energy is a physical quantity, while attributed emissions require a model with additional inputs. A correct first calculation cannot supply missing second-stage data.
22. Service reliability is measured through outcomes
Equipment redundancy, backups and monitoring are design features. Their success should be tested through observed recovery behaviour and service availability, not assumed from the architecture diagram alone. A system may retain power while an application returns the wrong answer or becomes unreachable.
Define meaningful failure and recovery states for users. For an educational assistant, the observable result may be whether the request completes correctly within an acceptable time and whether work is preserved during an interruption.
The promise of an AI service cannot exceed the evidence from its components and full workflow. Reliability belongs to the complete system.
23. Independent transfer: the smaller Gallery Compute Room
Now use a different fictional site. In its first configuration, the Gallery Compute Room draws 100 kW for IT and 150 kW for the total facility. In a new configuration, IT rises to 160 kW and total facility power to 208 kW. Both loads are assumed constant for twenty-four hours.
Calculate both PUE values, the total daily energy for both configurations, and the absolute change in energy. Decide whether the lower PUE necessarily means the facility consumed less electricity.
Then explain which information is missing if an operator claims the new site has reduced greenhouse-gas emissions or water use. Solve from the printed numbers and do not invent outside measurements.
24. Full transfer answers and interpretation
The first PUE is 150/100 = 1.5. The revised PUE is 208/160 = 1.3. The efficiency ratio improves under the stated measurement boundaries. Yet total daily energy grows from 150×24 = 3,600 kWh to 208×24 = 4,992 kWh, an increase of 1,392 kWh.
The original non-IT load is 50 kW; the revised non-IT load is 48 kW. In both cases absolute non-IT power remains a separate quantity from total site demand. It is therefore possible for the overhead ratio and overhead power to improve while the total site demand increases.
No carbon or water conclusion follows from these numbers alone. The student would need generation-mix or emissions data with a defined accounting method and, separately, suitable water measurements and boundaries. This is an exercise in carrying evidence limits across contexts.
25. A compact site evaluation checklist
Start with a description of the intended computing workload, its demand pattern and reliability needs. List capacity and measured utilisation for compute, memory, storage, networks, power, cooling and other physical infrastructure. State where measurement boundaries begin and end.
Verify total facility and IT energy over matching periods, use PUE with its precise denominator and preserve actual kilowatt-hours alongside the ratio. Assess failures, common dependencies, backups, cooling, and appropriate privacy or security obligations. Identify external grid and community constraints.
A useful recommendation can be conditional: this design meets the stated case under specific assumptions and needs testing before procurement or deployment. It should never claim real certification solely because the fictional calculations balance.
26. Frequently asked questions
Is PUE a measure of AI model efficiency? No. PUE describes a facility energy ratio; it does not measure output quality, useful computation per watt, latency or the intelligence of a model.
Does better PUE always reduce electricity consumption? No. North Quay’s revised configuration has a lower PUE and a higher total daily energy use because IT load increased. Both outcomes must be reported together.
Does a backup generator guarantee a service remains available? Not by itself. Power distribution, fuel or maintenance arrangements, networking and application health may still fail. System reliability requires wider evidence.
Applied analysis: measurement boundaries and operational evidence
A lower ratio requires compatible measurement periods
If IT power is measured at midday while total facility power is averaged over a week, their quotient is not a defensible instantaneous PUE. Define whether the metric uses a simultaneous sample or matching energy integration over the same time interval. Record meter boundaries and any unavailable readings.
North Quay’s calculations work because the toy packet stipulates constant loads for an identical twenty-four-hour period. Real operation varies. An audit should verify sensor time alignment and avoid combining peak IT power with average facility power.
Peak demand is different from daily energy
A site could use the same total kilowatt-hours as another yet have a much higher short-duration power peak. Electrical infrastructure may be constrained by its maximum simultaneous load as well as cumulative energy use. These are different questions for the operator and grid planner.
For example, two hypothetical loads can each consume 2,400 kWh: a steady 100 kW over twenty-four hours, or 200 kW over twelve hours and zero over the next twelve. Their daily energy matches, but their peak power does not.
A service outage needs a system diagnosis
Suppose the fictional compute hall has normal rack power but users receive no answers. The cause could be a network fault, unavailable application process or improperly routed request. Turning up computing power would not fix an unrelated failure.
Trace observable states in order: is the facility supplied, are network services reachable, is the application accepting requests, is a model instance available, and did the correct result return? This makes the physical-infrastructure explanation relevant to the user’s actual task.
Thermal headroom and operating conditions
A cooling arrangement that succeeds on a mild day may be stressed during an unusually hot period, a higher IT load or a maintenance interval. Name the design conditions and tested limits. A single PUE snapshot cannot establish cooling resilience across every season.
The right evidence includes relevant temperatures, component capacity, monitored operating history and recovery procedures. This is not a call to operate equipment outside its safe limits; it is a method for identifying which engineering claims require domain-specific verification.
Efficiency per useful outcome is another metric
PUE says little about whether the IT equipment performs valuable, correct work. A building running an inefficient model that repeatedly produces wrong answers may have a good facility ratio and poor outcomes for users.
A useful comparison can supplement facility metrics with successfully completed, checked user tasks per unit of energy, provided the tasks and counting rules are comparable. That metric would still require careful treatment of request difficulty and quality; it should not be invented from facility consumption alone.
A truthful decision record
An accountable facility recommendation should list its proposed workload, measured and rated electrical inputs, relevant cooling and network constraints, evidence for reliability, and external limitations such as grid connection status. Describe forecasts as scenarios with publication dates.
If capacity is adequate only under a simplified assumption, say so. If a utility approval or site inspection has not occurred, do not imply it has. Distinguishing a calculated scenario from verified operational readiness is the principle that transfers from engineering class to investment and infrastructure decisions.
27. Sources and connected education routes
Read the IEA’s Energy and AI report and its data-centre electricity demand chapter for 2025 baseline analysis and scenario forecasts. Uptime Institute’s discussion of PUE comparisons explains why facility size, conditions and measurement matter. These sources support general mechanisms, not the invented North Quay calculations.
Continue through eduKate’s Gold Standard of Data Centres, From Data Centres to Intelligence Factories, and GPUs and Accelerators. Return to How Super Intelligence Works, the How X Works hub and The Gold Standard of Education for the full mechanism-first curriculum.
North Quay and Gallery Compute Room are fictional teaching facilities. Their power draws, calculations and operating scenarios are invented and not measured on any real site. No physical inspection, utility consumption certification or product benchmark is claimed. Sources reviewed: 22 October 2026.