The Hidden Environmental Cost of AI: Energy, Water and Africa

Every AI answer has a physical cost that never appears on your screen. Behind the clean chat window sits a data centre drawing megawatts of electricity and drinking water to stay cool. We talk about AI as if it lives in "the cloud," but the cloud is a building full of hot machines somewhere real - and increasingly, that somewhere is Africa. As the continent races to host AI infrastructure, we need to be honest about what it costs our grids, our water and our climate, and we need to make siting decisions with eyes open.

The numbers are not small anymore

The International Energy Agency has tracked a steep climb. Global data centre electricity consumption was estimated at roughly 415 TWh in 2024 - about 1.5% of global electricity - and is projected to reach around 945 TWh by 2030, under 3% of the world total in the IEA base case. Data centre electricity use surged again in 2025, with AI-focused facilities growing fastest of all, far outpacing overall electricity demand growth. The IEA estimated AI systems accounted for around 15% of data centre electricity demand in 2024, with AI power demand potentially reaching over 20 gigawatts by the end of 2025.

Then there is water. Data centres use water to cool their servers, and the volumes are large: research puts global data centre water consumption at roughly 560 billion litres a year, potentially rising toward 1,200 billion litres by 2030. The AI share alone could account for hundreds of billions of litres annually. A United Nations warning in 2026 cautioned that AI could eventually use as much water as more than a billion people. And the carbon follows the electricity: the IEA estimated data centre power generation produced on the order of 180 million tonnes of CO2 in 2024.

Training a frontier model is an industrial event. Running it for millions of users, every day, is a permanent industrial load. Both draw power and water from a real place, on a real grid.

Training versus inference

It helps to split AI's footprint in two. Training is the one-time, enormous effort of building a model - weeks of thousands of specialised chips running flat out, consuming a large burst of energy. Inference is every individual use of the finished model afterward. Each query is tiny, but multiply it by hundreds of millions of daily users and inference becomes the dominant, continuous cost over a model's life. For businesses this matters: the "cheap" AI feature you embed everywhere adds up to real, ongoing energy demand, not a one-off.

This split also explains why the footprint is growing so fast even after the headline-grabbing training runs are done. A model is trained once, but it can be queried billions of times. As AI gets embedded into search, office software, phones and customer service, the number of inferences per day climbs relentlessly, and each one draws a little power and, indirectly, a little water. The industry is chasing efficiency gains, but demand is rising faster than efficiency is improving - which is exactly why the IEA keeps revising its projections upward rather than down.

Why this lands differently in Africa

This is where the story stops being a global abstraction and becomes an African policy question.

Strained grids come first

Africa hosts only around 160 data centres - roughly 5.5% of the global total - and construction is accelerating in South Africa, Nigeria, Kenya, Uganda and Algeria. But many African grids are already stretched, with aging infrastructure and millions of people still lacking reliable power. Analysts now name energy supply as the single most critical constraint on African digital infrastructure, because grid expansion cannot keep pace with demand. That raises a hard question of priorities: when a data centre and a neighbourhood compete for the same scarce megawatts, who wins? If AI facilities pull power from grids that cannot yet keep homes and clinics lit, the technology built to advance us could deepen an energy divide.

Water and siting are not uniform

The water picture is genuinely mixed, and the nuance matters. Research on African data centres finds most surveyed countries actually show lower AI-related water intensity than the global average, largely because they generate electricity from sources with low water intensity. But there are sharp exceptions: countries like Ethiopia and the Republic of the Congo, whose grids rely heavily on hydroelectric power, can show substantially higher water consumption because of reservoir evaporation. The lesson is that siting is everything - the environmental cost of the exact same server depends heavily on where it sits and how that region makes electricity.

The asymmetry problem

Studies mapping where data centres get built against where water stress is worst find the same regions overlapping - and the communities near these sites are often not the ones using the AI running inside them. A facility can consume local water and power to serve users on another continent. For African governments courting data centre investment, this asymmetry has to be on the table, not buried in the excitement of a ribbon-cutting.

What responsible adoption looks like

The answer is not to reject AI or the jobs and infrastructure it can bring. It is to be deliberate.

• Site with renewables. Several African countries - among them Cote d'Ivoire, Gabon and Senegal - are pairing data centre ambition with renewable energy investment. That is the model: build the clean generation alongside the compute, not on the back of a grid that is already short.
• Demand transparency from AI vendors. Ask providers about their energy sources, efficiency and water use. Corporate buyers pushing these questions move the market.
• Right-size your own AI use. Not every task needs the largest model. Smaller, efficient models for routine work cut cost and footprint at once. Matching the tool to the job is good engineering and good stewardship.
• Insist on water-aware cooling and honest local impact assessments before approving facilities in water-stressed areas.

The bottom line

AI's intelligence is virtual, but its appetite is physical - and Africa is about to host a lot of that appetite. The continent has a rare chance to build its AI infrastructure the right way: powered by the renewable resources it has in abundance, sited where it does not steal power and water from communities that need them, and governed by leaders who ask the hard questions before the concrete is poured. This is part of the same discipline of not rushing AI decisions under pressure - see why you should not get pressured into buying AI before your business is ready. Get the siting and the power mix right, and AI becomes an engine of African development rather than another drain on it.