Every AI answer is paid for in electricity, water and computing power. Africa is short on all three.
One email, one bottle of water
Asking a model like GPT-4 to write one 100-word email uses about 519 ml of water, a little more than a bottle, and 0.14 kWh of electricity, enough to light 14 LED bulbs for an hour. Once a week for a year, that becomes 27 litres of water and 7.5 kWh. Now multiply by a whole country.
Electricity: we cannot power what we do not have
Training one major AI model can use as much electricity as 1,000 homes consume in a year, and about half of it goes to cooling, not thinking. Meanwhile, roughly 600 million Africans still live without electricity. Every megawatt sent to a data centre is a megawatt not sent to a clinic, a school or a home.
Water: cooling machines while people wait for water
AI chips run so hot that data centres evaporate large volumes of water to cool them. In regions already facing drought and water stress, spending clean water to cool servers is a choice we cannot afford.
Compute: the price keeps rising
Building a frontier model has gone from roughly $1,000 (GPT-1, 2018) to millions (GPT-3) to close to $100 million (GPT-4). One high-end AI chip costs $25,000 to $40,000, and large models need 20,000 to 30,000 of them. Add top researchers earning $500,000 to $1 million a year. At that price, Africa becomes a customer and a data source, not a builder.
So what do we do?
We stop copying a model built for rich grids and full reservoirs. Africa should invest first in reliable, renewable power; in smaller, efficient models; in shared regional computing; and above all in people. Programming, data analysis, mathematics and statistics are the new Office suite. We must master them to use AI wisely, not just consume it.
— Akuété Giana MATHEY-APOSSAN, CEO & Founder, HOPE FOR AFRIKA · PMI Authorized Trainer · PfMP | PgMP | PMP | CPMAI | RMP
Figures on cost, electricity and water come from IST 692 Responsible AI course material (Syracuse University, Week 1). Email estimates come from a 2024 Washington Post / UC Riverside analysis for US data centres and vary by location. Electricity access figure: IEA, approximate.
Version française : Pourquoi l’Afrique ne peut pas se payer l’IA.
Leave a Reply