Data centres, power and networks are becoming a public argument in Africa. Much of that argument is imported, and it does not describe what is happening here. This is a plain record of the numbers: what the infrastructure costs, what it does not cost, and who it serves.
In the United States the fight over data centres is a fight about abundance. Those communities already have the power, the water and the connectivity, and are deciding how much more to give away. Africa is not in that position. Same subject, different argument.
Africa is not deciding whether to add capacity to a system that already has plenty. It is deciding whether to have the system.
Capacity today against demand in 2030 · McKinsey, November 2025
Four things change when the capacity is local.
Storing data in Europe sends every request ten thousand kilometres and back. Nigeria's exchange point more than doubled domestic traffic in eleven months once local hosting arrived.
Nigeria's offshore hosting bill runs at about $850 million a year, with roughly 70 percent of government agencies keeping data outside the country.
Only 5 percent of Africa's AI researchers can reliably get the compute they need, and only 1 percent have GPUs in their own building.
Thirty-six of 54 African countries now have data protection law. None of it reaches far while the servers holding citizens' records sit in another country's courts.
Some are true and some are not, and the honest answer is not always the one that suits us.
True or false?
Half right. Around 600 million Africans have no electricity, more than 80 percent of the global gap. Data centres did not cause that and refusing them will not close it. What matters is whether a facility adds generation or only draws from it. Kenya gets over 90 percent of its power from renewables and passed 1,000 MW of installed geothermal in March 2026, which is the kind of grid that can carry this load without taking from anyone. Its entire data centre load is about 40 MW today, against a grid measured in thousands.
The standard we hold to. Arrive with contracted, additional generation, and pay as an anchor customer that makes new supply financeable. The real risk runs the other way: private captive power that bypasses the utility and leaves the grid no better off.
A fair concern. Evaporative cooling uses roughly 25 million litres per megawatt a year and loses 80 percent of it to the air. In August 2026 civil society groups called for an inquiry into a proposed 174 MW campus near Cape Town on those grounds. The objection is reasonable. But evaporative cooling is a choice. Closed-loop and air-cooled designs use almost no water, at the cost of a little more electricity.
The standard we hold to. Closed loop wherever water is stressed, with the figure published yearly and open to audit. An operator who will not publish it should not get the site.
Largely true. A colocation site supports three to six direct jobs per megawatt. A 100 MW build is about 500 construction jobs and 50 permanent roles. Anyone promising a factory’s worth of employment is selling something. The case sits downstream, in what can only run on capacity people can reach, and in a technical trade Africa currently imports.
The standard we hold to. Publish real headcount, construction and permanent, counted separately. Hire and certify locally instead of flying crews in.
Not the real constraint. Data centres have one of the smallest footprints in heavy infrastructure. McKinsey expects two thirds of new African capacity at 1 to 50 MW, not the 500 MW campuses of Virginia and Texas. Those fit inside industrial zones on a few hectares. What binds here is proximity to firm power and fibre.
The standard we hold to. Build inside existing industrial and energy zones. Publish the site plan before construction, not after, and hold a session where people can read it.
True so far. The largest facilities here were financed to serve global cloud platforms, and much of the value flows straight back out. That is an objection about ownership, not about whether the infrastructure should exist. And 36 of 54 African countries now have data protection law, which counts for little while the servers sit under another country’s jurisdiction.
The standard we hold to. Local ownership and local tax residence. A published share of capacity for universities, public research and local startups at prices they can pay. No exclusivity that outlasts the hardware.
Easily designed out. A data centre reads 40 to 59 decibels at the boundary. A library is 30, a conversation 60, and hearing damage starts at 85. The noise comes from cooling and from backup generators, and generators idle almost always, running about 30 minutes a month for testing. In Northern Virginia, the densest cluster anywhere, they account for under 4 percent of regional nitrogen oxides. Setbacks, acoustic walls and tree lines work, because sound falls off with distance.
The standard we hold to. A sound study before construction and after commissioning, both published. A boundary limit in the permit, and a named person a neighbour can call when it is breached.
Where a figure is American we say so, and we say what carries across.
Kenya runs on more than 90 percent renewables and passed 1,000 MW of installed geothermal in March 2026. Its entire data centre load is about 40 MW today and is forecast near 155 MW by 2029, against a national grid measured in thousands of megawatts. The gap that matters is the 600 million people with no supply at all, and no data centre caused it.
Kenya data centre IT load · African Energy Chamber, 2026
Contracted, additional generation, and an anchor customer that makes new supply financeable. The real risk runs the other way: private captive power that bypasses the utility and leaves the grid no better off.
Evaporative cooling uses roughly 25 million litres per megawatt a year. Closed-loop and air-cooled designs use almost none, for a little more electricity. Between 25 and 30 percent of new data centres used advanced closed-loop cooling in 2026, against under 5 percent a few years earlier.
Annual water use, United States · LBNL 2024 and industry figures
Water per megawatt per year, by cooling method
Closed loop wherever water is stressed, with the figure published yearly and open to audit. An operator who will not publish it should not get the site.
Data centres occupy one of the smallest footprints in heavy infrastructure. McKinsey expects two thirds of new African capacity at 1 to 50 MW, not the 500 MW campuses of Virginia and Texas, and those fit inside industrial zones on a few hectares.
Share of total land, United States · USDA and industry figures
Build inside existing industrial and energy zones. Publish the site plan before construction, not after, and hold a session where people can read it.
A data centre reads 40 to 59 decibels at the boundary. Backup generators idle almost always, running about 30 minutes a month for testing, and in Northern Virginia they account for under 4 percent of regional nitrogen oxides. Setbacks, acoustic walls and tree lines work, because sound falls off with distance.
Measured at the boundary · Virginia JLARC report 598
A sound study before construction and after commissioning, both published. A boundary limit in the permit, and a named person a neighbour can call when it is breached.
A colocation site supports three to six direct jobs per megawatt. A 100 MW build is roughly 500 construction jobs and 50 permanent roles. Anyone promising a factory’s worth of employment is selling something. The employment case sits downstream, in what can only run on capacity people can reach.
Per 100 MW build · industry figures
Publish real headcount, construction and permanent, counted separately. Hire and certify locally instead of flying crews in.
The water and land comparisons are American, because the United States has counted more carefully. Decibels, land arithmetic and cooling physics carry across. The catchment a site draws from does not, and has to be answered site by site.
What we ask of every operator building here, Navon included. Each one is a number or a document an outsider can check.
Contracted and additional, not taken from the queue in front of households. Publish the source and the term.
Publish water use effectiveness annually, in litres, open to audit. Accept the small efficiency penalty.
Power use effectiveness on the same schedule, so a site can be compared against its own past and its neighbours.
Construction and permanent roles counted separately. Hire and certify locally, and fund the training.
A published share of every site kept for universities, public research and local startups, in local currency where the law allows.
Publish the plan, the power draw and the water design before construction starts, and take questions on it.
Current as of September 2026. Where sources disagree we show the range. If a figure is wrong, out of date, or reads more favourably than the research supports, tell us and we will correct it.
Capacity today and to 2030, investment ranges, facility size mix.
Share of global capacity, committed investment, grid and localisation risks.
Traffic leaving the continent, latency before and after local exchange points.
Researcher access to GPUs, on premise and on cloud.
Annual offshore hosting spend, share of government agencies hosting abroad.
Kenya data centre load today and to 2029.
Evaporative cooling water intensity, the Cape Town objection.
Continental policy position, phased implementation from 2025 to 2030.
The four US fact sheets behind the comparison figures.
US data centre water consumption, the 17.4 billion gallon figure.
Effect of US data centre load on retail electricity rates, 2015 to 2024.
Boundary noise levels, generator run hours, generator share of regional emissions.
Regulators, utilities, universities, community groups: if you have a question about a project near you, we would rather answer it than read about it later. If you have better data, send it.