Power On: Why AI runs on electricity

Every AI application depends on electricity. When a person asks a chatbot a question, or a bank runs a fraud check on a mobile payment, computers in a data centre process the request, and those computers need a constant supply of power. Before AI can answer a question, an energy system somewhere has to keep that data centre running.

This article, the first in the Powering AI series, explains what that dependency involves and why it has become a question for policymakers. The effect of a data centre on an African power system depends on how much electricity it uses, how it connects to the grid and the terms of that connection. Energy regulators, system planners, system operators and the ministries that approve large connections determine many of those terms, yet they are rarely part of discussions on AI policy.

What a data centre contains and why its power supply matters

A data centre is a building that houses the equipment used to store, process and transmit data. It contains servers, many of them fitted with specialised processors such as graphics processing units (GPUs) that carry out the calculations AI requires, networking equipment that moves data in and out, and cooling systems that stop the equipment from overheating. All of these run on electricity.

The services a data centre hosts cannot tolerate an interruption in power. If grid supply fails and backup batteries and generators do not take over within moments, those services stop. Operators therefore seek highly reliable supply, and guarantees of availability, backup arrangements and redundancy are central points in their negotiations with utilities and governments. However, some flexibility is possible. The International Energy Agency (IEA) notes that grid operators can offer data centres faster connections in return for reducing the power they draw from the grid when the system is under stress, with on-site batteries or generators keeping the facility running (IEA 2026). The next article in this series looks inside the building in more detail.

Why AI uses more electricity than ordinary computing

Sending an email or loading a web page requires relatively little computing. AI tasks involve far more calculations, carried out on specialised processors that draw much more power than the servers used for most conventional digital services. The picture is changing quickly. According to the IEA, the energy used per AI task has fallen at least tenfold each year in recent years, and a simple text query now uses less electricity than running a television for the same length of time. At the same time, newer uses such as video generation, reasoning and agentic tasks can use hundreds or thousands of times more energy per query than simple text (IEA 2026). The total electricity used by AI therefore depends on three trends at once: gains in efficiency, growth in the number of users, and changes in the kinds of tasks people use AI for.

Training and inference

This electricity demand comes from two main activities: training and inference. Training builds the model. For the largest models, it involves thousands of specialised processors working together, often continuously for weeks or months. Meta, for example, used a cluster of more than 16,000 GPUs to train its largest Llama 3 model (Grattafiori et al. 2024). Training concentrates a very large amount of electricity demand in one place for the duration of each training run. Because training does not need to respond to users in real time, it can be carried out far from where the model will eventually be used. Developers can therefore site it where electricity is available in bulk, reliable and relatively inexpensive (IEA 2025c).

Inference is the use of the model once it has been built, and it happens every time someone sends a request. Each request uses far less electricity than training, but requests arrive continuously and multiply as more people use AI, so inference creates an ongoing demand for electricity that grows with adoption. That demand is drawn from the grid wherever the data centres serving users are located, which may be in the user's own country or in another one. Most AI providers do not publish where consumer requests are processed. OpenAI, for example, lets eligible business customers store their data in ten regions, none of them in Africa (OpenAI 2025b). It offers in-region inference for its ChatGPT business products only in the United States and Europe (OpenAI 2025a).

This distinction matters for Africa. Because training is not sensitive to delay, it tends to be located where electricity is cheapest and most abundant, which currently means outside the continent. Inference has more reason to be located near users, although most AI services used in Africa today are probably processed in data centres abroad. That may change as AI use grows and as governments introduce data localisation rules and demand locally hosted services. If inference infrastructure is built in African countries, it will add new demand to national power systems. Where grids cannot supply it reliably, operators are likely to rely on their own generation or compete with existing users for limited capacity.

How much electricity data centres use

According to the IEA, data centres used about 415 terawatt-hours (TWh) of electricity in 2024, around 1.5% of global consumption (IEA 2025c). The agency's 2026 update estimates 485 TWh in 2025 and projects about 950 TWh by 2030, around 3% of global electricity demand (IEA 2026). That is slightly more than Japan's total electricity consumption today (IEA 2025a). These are central projections; the IEA also publishes alternative scenarios and revises its figures regularly.

National and global totals do not show how this demand is distributed. The IEA notes that an AI-focused data centre can draw as much electricity as an aluminium smelter, and that data centres are much more geographically concentrated than other large consumers (IEA 2025c). Total demand can therefore look manageable while one large facility places heavy demand on a particular part of the grid. Grid operators need to consider where data centres connect and whether the local network can carry the extra load. This is made harder by uncertainty over how much power a new facility will actually use. Data centres fill with servers gradually and often reserve more grid capacity at the start than they need (IEA 2026).

What a data centre asks of a grid

A large data centre needs a set amount of firm capacity, meaning power the network can guarantee to deliver at all times, at a specific site. The connection must also be designed so that the failure of a single line or transformer does not cut off supply. These requirements limit where data centres can be built. Access to transmission is usually one of the most important factors in choosing a site, alongside fibre connectivity, land and water.

Timing adds a further difficulty. According to the IEA, a data centre can be operational within two to three years, while the energy infrastructure needed to supply it usually takes longer (IEA 2025c). In Europe and the United States, developing a new extra-high-voltage transmission line takes around eight years on average (IEA 2025b). The delay comes from route planning, land acquisition, environmental approvals and financing. It also comes from delivery times for cables and large transformers, which almost doubled between 2021 and 2024 (IEA 2025d). Some new generation, such as solar power, has short lead times (IEA 2025c), but large hydropower, geothermal and other conventional plants take considerably longer to build.

As a result, new electricity demand can arrive before the grid is ready to serve it. When that happens, data centres may wait for a connection, build their own generation, or add load to a network that is already under strain. In the United States, slow grid connections have already led some developers to pursue on-site gas-fired generation (IEA 2026). Governments seeking to attract data centre investment therefore need to plan grid and generation projects ahead of expected demand, so that the power system is ready when new facilities are completed.

Africa's data centres and the terms on which they connect to the grid

More companies want to build data centres in Africa, but the continent starts from a small base. The African Energy Chamber, an industry body, counted 223 data centres in 38 African countries as of mid-2025. The largest numbers are in South Africa (56), Kenya (19) and Nigeria (17), which together account for about 41% of the total (Ayuk 2025).

Kenya and South Africa show how quickly this could change. The Chamber estimates that Kenya has about 40 megawatts of IT load capacity and expects total capacity to exceed 155 megawatts by 2029 (African Energy Chamber 2026). In South Africa, announced large-scale projects could add as much as 1,000 megawatts to national electricity demand if they are all built (Carnie 2026a).

These amounts are significant for both countries. Kenya's total installed generation capacity is about 3,000 megawatts (Kinghorn 2026), so 155 megawatts would equal around 5% of it. Kenyan media have also reported that in November 2025 the president acknowledged daily evening load-shedding to keep the grid stable (Kimani 2026).

In South Africa, 1,000 megawatts is about the amount of power removed from the grid during one stage of load-shedding. An energy analyst has warned that new data centre demand, combined with other new loads, could lead to constrained supply by 2029 unless generation capacity is added quickly (Carnie 2026a).

The pressure is greater at the local level, because data centres cluster in a few places. A proposed data centre in Durban, reported at 400 megawatts, would need the equivalent of about a quarter of the city's current electricity demand, although the city says such figures are not yet confirmed (Carnie 2026b). Each facility connects to a particular substation and power line, and these may be unable to carry the extra load even when the national system has power to spare.

Nigeria shows a different problem. Its 17 data centres need about 137 megawatts in total, but the Chamber reports that the national grid often supplies power for only a few hours a day, so operators rely heavily on diesel generators (Ayuk 2025). In Nigeria, the main difficulty is the reliability of supply from the grid. Diesel is expensive and polluting, and these costs are rarely mentioned when new investments are announced. A country can have the demand, the investors and the sites for data centres, and still find that a facility can only operate if it generates much of its own electricity.

Data centres also bring an opportunity. They use large amounts of electricity at a steady rate and usually sign long-term contracts to buy it. This makes them reliable customers, which can help new power plants and grid projects raise the money they need (Ayuk 2025). Whether other electricity users benefit depends on the terms under which each facility connects to the grid.

Those terms settle four questions:

• who pays to upgrade the network so that it can carry the new load;
• which electricity tariff the facility pays;
• whether the facility keeps its supply or is cut off when there is not enough power for everyone; and
• what the operator provides in return, such as building its own generation or reducing its demand at peak times.

Some of these questions are answered by published rules, such as national tariff schedules and grid codes. Other matters are usually settled in agreements between the operator, the utility and, in some cases, the government. How much of these agreements is made public varies from country to country. Because these agreements shape how a data centre affects the power system and its other users, later articles in this series examine them in more detail.

Kenya's experience shows how much these questions can matter. Microsoft and G42 announced a $1 billion data centre at Olkaria in 2024, with a first phase of 100 megawatts, and it missed its planned May 2026 opening (Butler 2026). The president has said publicly that the grid could not support the project at its intended scale. Reports also indicate that talks stalled after the government was unable to guarantee the annual capacity payments Microsoft had requested. The government says the project has not been withdrawn and is being restructured (Kinghorn 2026).

This suggests that national AI strategies are more likely to include enforceable commitments on power when the energy regulator, the system operator and the ministry that approves large connections are involved in preparing them. Many of the decisions that will determine where AI infrastructure is built in Africa, and on what terms, are being made now. They are being made in investment and energy negotiations that AI policy discussions seldom address.

References

African Energy Chamber (2026), How Data Centers Are Reshaping Africa's Power Market, press release, 27 January, Cape Town: African Energy Chamber, accessed 21 September 2026.

Ayuk, NJ (2025), Data Centers Could Be the Spark Africa's Power Sector Needs, 29 December, Johannesburg: African Energy Chamber, accessed 21 September 2026.

Butler, Georgia (2026), Microsoft and G42 Data Center in Kenya Stalled Due to Lack of Power Capacity, DatacenterDynamics, 7 May, accessed 21 September 2026.

Carnie, Tony (2026a), Major SA Data Centre Expansion Plans Emerge, Set to More than Double Current Power Demand, Daily Maverick, 9 February, accessed 21 September 2026.

Carnie, Tony (2026b), SA's Biggest Data Centre "Monster" Set to Consume 25% of Durban's Electricity, Daily Maverick, 4 March, accessed 21 September 2026.

Grattafiori, Aaron, Abhimanyu Dubey, Abhinav Jauhri, et al. (2024), The Llama 3 Herd of Models, arXiv:2407.21783, Menlo Park, CA: Meta AI, accessed 21 September 2026.

IEA (International Energy Agency) (2026), Key Questions on Energy and AI, Paris: IEA, accessed 21 September 2026.

IEA (International Energy Agency) (2025a), AI Is Set to Drive Surging Electricity Demand from Data Centres While Offering the Potential to Transform How the Energy Sector Works, press release, 10 April, Paris: IEA, accessed 21 September 2026.

IEA (International Energy Agency) (2025b), Building the Future Transmission Grid: Strategies to Navigate Supply Chain Challenges, Paris: IEA, accessed 21 September 2026.

IEA (International Energy Agency) (2025c), Energy and AI, Paris: IEA, accessed 21 September 2026.

IEA (International Energy Agency) (2025d), Rising Component Prices and Supply Chain Pressures Are Hindering the Development of Transmission Grid Infrastructure, press release, 25 February, Paris: IEA, accessed 21 September 2026.

Kimani, Brian (2026), Kenya's $1B Microsoft Data Centre Project in Limbo over Power Concerns, Citizen Digital, 23 May, accessed 21 September 2026.

Kinghorn, Jess (2026), Microsoft's $1 Billion Data Center Project in Kenya Stalls amid Row over Payment and Power, PC Gamer, via Yahoo Finance, 13 May, accessed 21 September 2026.

OpenAI (2025a), Introducing Data Residency in Europe, 5 February, updated 16 January 2026, San Francisco, CA: OpenAI, accessed 21 September 2026.

OpenAI (2025b), Expanding Data Residency Access to Business Customers Worldwide, 25 November, San Francisco, CA: OpenAI, accessed 21 September 2026.