How AI Data Centres Draw Power, and Why It Matters for Electricity Pricing

AI is adding a new class of large, concentrated electricity demand. In its latest detailed global data centre outlook, the International Energy Agency estimates consumption at about 415 TWh in 2024 and projects around 945 TWh by 2030 in its Base Case. That is slightly more than double, with AI the most important driver of growth. For utilities and data centre operators, however, annual consumption is only the first question. The hourly load profile determines network exposure, wholesale risk and the value of flexibility.

There is no single, universal energy figure for an AI query. Consumption varies with the model, hardware, response length, batching, utilisation and cooling system. Comparisons with a web search can be useful illustrations, but they are too sensitive to assumptions to serve as a reliable basis for tariff analysis. A facility model should instead begin with measurable engineering inputs.

From chip ratings to the grid connection

A modern AI data centre combines accelerators, servers, networking, storage, cooling, power conversion and other building services. NVIDIA specifies a maximum thermal design power of up to 700 watts for an H100 SXM accelerator and 400 watts for an A100 SXM accelerator. Actual draw depends on workload and configuration, so nameplate values should not be treated as continuous metered demand. A bottom-up facility model moves from chip to megawatt scale by multiplying accelerator count by utilisation-adjusted power draw, adding the CPU, memory, networking and storage load for each server, then aggregating servers by rack and checking the result against rack power density and the facility's IT capacity.

The bridge between IT equipment and the grid meter is power usage effectiveness, or PUE. A PUE of 1.35 means that every 1.00 megawatt used by IT equipment requires 0.35 megawatts for cooling, power losses and supporting infrastructure. A 20 MW IT load would therefore imply up to 27 MW at the grid connection. The same PUE assumption is used throughout the model below: facility overhead is calculated as 0.35 times IT demand, not introduced as a separate parameter.

An illustrative daily dispatch profile

The chart separates IT demand into relatively time sensitive inference and potentially flexible batch computing. Training is shown moving into lower demand hours, illustrating how workload scheduling can reduce exposure to higher-cost periods and create a more credible demand response offer. Facility overhead is derived from the PUE assumption above, so the three stacked components reconcile exactly to the metered total.

Load factor changes the conversation

A data centre with high utilisation can be attractive to a retailer because it buys a large and comparatively stable volume. That does not automatically make it cheap to serve. A customer that contributes to regional peaks, requires new network capacity or cannot curtail during system stress may still impose substantial costs. The useful question is not whether the load is “good” or “bad”, but which costs are created and which risks the customer can reduce.

Flexible computing has value, but not all computing is flexible

Some training, fine tuning and batch jobs can be delayed or moved between regions. Live inference, latency commitments, data residency rules and accelerator availability constrain that flexibility. Any demand response offer should therefore specify the dependable megawatts, notice period, duration, recovery profile and number of events the operator can support.

A credible energy proposal does not promise that every workload can move. It identifies the portion that can move, the conditions under which it will move and the value created for both parties.

Turning the profile into a shadow cost model

A shadow cost model applies alternative pricing structures to the same half hourly or hourly load. It can test energy rates, network charges, capacity charges, loss factors, renewable products and demand response payments. The method is more defensible than applying a single discount to annual consumption because every saving can be traced to a change in price, load or risk.

The negotiating levers that can be quantified

Load scheduling. Shift only the computing jobs that can move without breaching service, data or hardware constraints. Value them against the actual hourly or settlement price exposure.

Capacity management. Reduce the billing peak where the tariff rewards it. Battery storage, cooling controls and workload orchestration may help, but the demand charge rules determine whether the reduction persists on the bill.

Demand response. Offer a firm, testable product rather than a broad statement about flexibility. The network or retailer needs to know what can be delivered during a real event.

Contract structure. High annual volume and a long operating life may support a tailored supply arrangement. The value depends on credit quality, volume certainty, connection costs and which party accepts wholesale and renewable certificate risk.

Conclusion

The dispatch profile is the foundation of a serious electricity procurement strategy for an AI data centre. It converts engineering design and computing schedules into tariff exposure, network impact and a measurable flexibility offer.

The strongest negotiating position is not that a data centre is large or technologically important. It is that the operator can demonstrate, with interval data and operational evidence, how its load behaves and what it can reliably change. That is the basis for a credible value exchange with a retailer, network business or system operator.