D1.1 · What an AI data centre actually does
Before discussing batteries or grid capacity, we need to understand the business being powered. An AI data centre combines computing, data movement, electricity and heat removal to deliver a usable digital service. This lesson follows a job from its arrival to its result, explains what changes between training and inference, and then connects those activities to the facility’s energy requirements. No electrical-engineering background is assumed.
Read the concepts and system relationships in sequence before attempting the workshop. Key terms are explained in the context of their function. If a task is unclear, return to its section and explain the reasoning before calculating.
What should you be able to judge after this lesson?
- Explain the roles of computing, storage, networking, power and cooling in one complete service.
- Distinguish model training, fine-tuning and inference without assuming they share one load profile.
- Read an IT-capacity claim, facility-demand figure and energy-efficiency metric at the correct boundary.
- Explain why available power, continuity, cooling and commercial permission must be considered together.
Follow the computing service
Users and data
A task arrives; storage provides inputs, model files and a place to retain results.
Coordinated computing
CPUs, accelerators, memory and networks complete training or inference together.
Usable business result
Delivery requires the agreed timing, quality and availability, not merely powered equipment.
Then follow the supporting energy path
Permitted sources and distribution
Grid or on-site sources supply electricity through switching, transformation and distribution.
Equipment and conversion
IT computes; UPS, fans, pumps and other supporting systems also consume power.
Continuous heat removal
Air or liquid carries heat away; the heat-rejection chain must remain usable in each relevant state.
1. A data centre sells a dependable computing service, not simply floor space
Information Technology (IT) equipment includes the servers, storage and networks doing the digital work. A server performs calculations and runs software. Storage keeps the data and model files that the software needs. Networks move requests and results between users, storage and servers. A data centre houses these components and provides the electricity, environmental conditions, physical security and operations needed to keep them usable. A powered building with empty racks is therefore not the same as a functioning computing service.
For example, a customer may submit an image for defect detection. The request must reach a running model, the model must read its parameters and input, an accelerator must perform the calculation, and the result must return within the customer’s time limit. A server that has power but cannot reach the model file or network still fails this service. The relevant outcome is a completed useful task, not merely an energised socket.
This distinction explains why an energy project begins with a service requirement. The operator must say which service must remain available, what delay or interruption is acceptable, and which supporting systems it depends on. Those answers determine which loads are critical and which can be adjusted.
2. “AI” describes the computing work; it does not define a single building type
Artificial intelligence (AI) is a broad family of computational methods. Machine learning (ML) is one approach within it: a model learns patterns from examples rather than relying only on rules written individually by a programmer. A model contains numerical parameters that determine how it transforms an input into an output. In this course, an AI data centre means a facility hosting substantial AI computing; it is not a claim that the building itself operates autonomously.
An AI cluster is a group of computing nodes that cooperate on a workload. Nodes may contain central processing units (CPUs), graphics processing units (GPUs) or other accelerators. A CPU handles general tasks and coordination; an accelerator can perform many suitable operations in parallel. Memory holds the working data close to the processor, while storage retains larger datasets and files. Faster processors alone cannot deliver useful throughput if data movement becomes the bottleneck.
A facility can mix AI and conventional workloads. Some AI inference runs on CPUs or smaller edge systems, while large training jobs may use dense accelerator clusters. Therefore “AI ready” is not a complete electrical specification. You still need the actual hardware population, rack power, network/storage arrangement, utilisation and expansion plan. NVIDIA’s infrastructure overview describes compute, networking, storage, power and cooling as an integrated stack. [1]
3. Training changes a model; inference uses a trained model
During training, the system presents examples to a model, calculates an output, compares it with the training objective and adjusts model parameters. This repeats over many batches. In distributed training, multiple devices may exchange intermediate information so that they can update a shared model consistently. Computation, communication and saving progress can therefore occur in different phases rather than at a constant electrical demand.
Fine-tuning continues training from an existing model to adapt it to a narrower task or dataset. Inference is different: the deployed model processes new inputs to produce answers, predictions or classifications. A conversational application may generate its answer step by step, while an inspection model may classify an image. Both are inference, but their latency, throughput and hardware requirements can differ substantially.
It is tempting to describe all training as flexible and all inference as inflexible. That is too crude. A training job may face a delivery deadline, expensive restart or coordination constraints; batch inference may tolerate delay, while a real-time user request may not. Electrical flexibility must be authorised for the particular workload, including pause, checkpoint, restart and recovery consequences.
4. Follow three connected flows: information, electricity and heat
Information enters through networks, is read from storage and processed by servers, and leaves as a result. Electricity enters through the utility connection or an authorised on-site source, passes through switchgear, transformers and distribution equipment, and reaches server power supplies. The two flows are related, but they are not the same: a network request is not an electrical dispatch command.
Most electricity consumed inside computing equipment ultimately becomes heat within the facility. Air or liquid transfers heat away from chips and racks; mechanical systems then reject it to the environment or recover part of it for another use. Cooling itself needs energy for fans, pumps and, where applicable, compressors. Liquid cooling changes the transport path; it does not make heat disappear or automatically eliminate all mechanical energy use.
The useful operating limit is therefore whichever constraint becomes binding first. A hall may have spare electrical capacity but insufficient heat rejection at the required ambient condition. It may have spare cooling capacity but no approved grid capacity. An expansion decision must check both paths and their shared support systems, rather than treating a transformer rating as the whole facility capability.
5. Separate a chip rating, IT capacity and facility demand
A chip’s stated power is not the power of its entire server, and a server’s power is not the total rack or hall demand. Memory, networking, storage and power-conversion losses add to the equipment boundary. At facility level, cooling and other supporting loads also contribute. Ask what is included whenever a project describes itself as a “20 MW data centre”.
Installed capacity describes what has been fitted; contracted or approved capacity describes a right or limit; operating demand describes what is actually consumed at a time. They can all have different values. A hall built for future racks does not necessarily consume its full design power on day one, while a contracted connection limit may prevent all installed equipment from operating together.
Power is the rate of energy use: 1 MW maintained for 1 hour consumes 1 MWh. A 10 MW import permission is not a store of 10 MWh and does not by itself promise availability through an outage. Likewise a 10 MWh battery does not tell you the converter power it can deliver. Keep capacity rights, instantaneous power, energy and continuity as four separate questions.
6. PUE explains an energy boundary, not the whole business
Power Usage Effectiveness, or PUE, is total facility energy divided by IT equipment energy over the same boundary and period. If a fictional facility uses 78 GWh in a year and its IT equipment uses 60 GWh, annual PUE is 1.30. The remaining 18 GWh belongs to the non-IT part of that defined boundary. The calculation says nothing on its own about how much useful AI work was completed.
A lower PUE can reflect more efficient supporting systems, but comparisons need consistent climate, load level, measurement and operating conditions. An inefficiently used cluster can have a good PUE while producing little useful work. Electricity carbon intensity, water use, reliability and cost also require their own measures. DOE’s design guide treats IT, cooling, electrical systems and performance metrics as connected design subjects. [2]
An annual energy ratio cannot simply be multiplied by an IT peak to prove a facility peak. The cooling and electrical overhead at that instant may differ from its annual average relationship. For a preliminary calculation, use an explicitly stated coincident power assumption and label it as an assumption until time-aligned measurements or a validated design model support it.
7. AI changes density and coordination; it does not prescribe one load shape
When more computing power is concentrated in a rack, the local power distribution and heat-removal duty rise. If many accelerators perform related phases together, their electrical changes may also align. However, different job mixes, utilisation, scheduling and power controls can smooth or reinforce those changes. The label AI alone cannot tell a utility the peak, ramp or duration it should design for.
A utility meter reporting 15-minute averages can support an energy or billing analysis while concealing faster events. A rack measurement can reveal a local event without proving that it reaches the grid unchanged. DOE’s 2026 discussion of oscillation monitoring illustrates why measurement bandwidth and filtering matter. [3] The third lesson develops this point after the service and system boundaries are understood.
The Point of Common Coupling (PCC) is the defined connection boundary between the site and the external network. Begin a data request with a question such as “what load reaches the PCC when this cluster starts a job?” rather than requesting an unspecified AI profile. Ask for workload state, metering location, time resolution and simultaneous cooling behaviour. This makes the requested data useful for a specific decision.
8. The owner, facility operator and compute customer may want different things
A colocation provider may supply space, power and cooling while a tenant owns the servers and controls workloads. A hyperscale operator may control more of the full stack. A developer may still be securing a connection before any compute customer is contracted. Do not assume the person requesting an energy proposal has authority to pause computing or commit backup equipment to grid services.
A service-level agreement, or SLA, translates business expectations into obligations such as availability, response or recovery conditions. It does not automatically identify the electrical design that satisfies them. The IT team, facility team and energy supplier must jointly translate those obligations into load groups, interruption limits, thermal limits and restoration priorities.
This is why an apparently cheap energy-saving opportunity may be unacceptable. Pausing a job could save a small electricity cost yet delay a contracted delivery or strand expensive computing equipment. Compare energy benefits with the actual service consequence, and obtain the authorised owner’s decision rather than assuming all technical flexibility is commercially available.
9. Grid supply, UPS, BESS and generation solve different problems
Grid supply normally provides ongoing energy within a connection arrangement. An Uninterruptible Power Supply (UPS) maintains power to its connected loads across defined input disturbances using its conversion system and stored energy. A separate Battery Energy Storage System (BESS) may support peak management or longer energy duties, depending on how it is connected and controlled. A generator converts fuel to electricity but needs a validated starting, switching and load-acceptance sequence.
These functions can overlap in a designed system, but the equipment labels do not establish that they do. A campus battery connected upstream of a disconnected switch cannot support a downstream load merely because the battery has enough MWh. A generator’s MW rating does not prove it can accept the required step load or start the cooling system. Trace the electrical and control paths before adding their capacities.
A large persistent gap between facility demand and grid capacity is therefore not automatically a battery project. A battery shifts energy; it must obtain that energy and restore reserve. Possible responses include phased occupancy, a confirmed connection upgrade, authorised workload limits or an appropriately permitted firm source. The right answer depends on duration, operation and commercial constraints, not on choosing the most fashionable technology.
Workshop: form a decision from the given inputs
Attempt these tasks before opening the reference answers.
- Explain what training and inference contribute to the service, and name one supporting-system failure that can stop either without a loss of server power.
- Calculate present demand, present import headroom and the later-phase gap. Explain why “20 MW ready” does not answer these questions.
- Calculate annual PUE and explain which operating claim it cannot establish.
- Explain who must authorise workload flexibility and what evidence is needed before choosing an energy solution.
Answer 1 · Describe the service chain before naming equipment
Training produces or updates model parameters; inference uses a deployed model to answer a new request. Both also require usable data paths and acceptable thermal conditions. A network failure that prevents access to storage or delivery of a result can stop the service even while server power stays on. Loss of cooling can force throttling or shutdown despite electrical continuity.
The output from this step should be: A service map linking the customer outcome to computing, data, power and cooling.
Answer 2 · Keep the operating balance separate from the marketing label
Current IT demand is 6 + 2 = 8 MW. Total facility demand is 8 + 2.4 = 10.4 MW, leaving 12 − 10.4 = 1.6 MW of arithmetic import headroom. This does not yet establish a usable expansion allowance because equipment ratings, reserve and cooling remain unverified.
Later-phase demand is 12 + 3 = 15 MW, a 3 MW gap against the stated cap. “20 MW ready” could describe future fit-out or another boundary; ask for its definition and evidence. Do not add UPS or battery nameplates to the 12 MW import right or assume that they create continuous energy.
The output from this step should be: A current/later phase balance showing the measurement boundary and unresolved cooling conditions.
Answer 3 · The annual ratio does not settle the peak
Annual PUE is 78/60 = 1.30, and the non-IT annual energy is 18 GWh. These values describe the separate annual dataset. Multiplying 12 MW by 1.30 gives a hypothetical 15.6 MW, not a demonstrated later-phase peak. The provided coincident later-phase balance is 15 MW; real peak and transient requirements need aligned data or a supported model.
The output from this step should be: A metric note stating the period, boundary and limit of the PUE interpretation.
Answer 4 · Obtain service authority before promising flexibility
The tenant controls jobs, so the building owner cannot promise to pause them without the relevant contractual authority. Ask the tenant which services tolerate delay and how restart affects delivery; ask the facility team for single-lines, cooling boundaries and failure duties; ask the connection party for capacity conditions and a credible upgrade schedule.
With these answers, compare phased operation, approved job limits, connection reinforcement and permitted on-site supply. The current evidence justifies recognising a later-phase constraint and requesting data. It does not justify selecting battery size, guaranteeing availability or claiming investment returns.
The output from this step should be: An evidence request owned jointly by the tenant, facility operator and electricity connection party.
The key decision from this lesson
An AI data centre is a service system in which computing and data movement depend on continuous power and heat removal. First define the service, operating boundary and authorised flexibility; only then choose an energy architecture.
Sources and scope
Reviewed 25 October 2026. Primary sources provide background; numbered references in the text identify the relevant documents below. Cases, calculations and action suggestions are original teaching derivations. They do not replace site specifications, network permissions or contract review.
[1] NVIDIA · What Is AI Infrastructure? · infrastructure background ↗
[2] U.S. DOE FEMP · Best Practices Guide for Energy-Efficient Data Center Design · 2024 ↗
[3] U.S. DOE · Monitoring Oscillations from Large Data Centers · 28 May 2026 ↗
Further reading · DOE SEAB · Powering AI and Data Center Infrastructure · July 2024 ↗