Part 2 of 3: Techno-Feudalism and the AI Divide
There is a number that tells you almost everything you need to know about where AI will and will not take root.
Sub-Saharan Africa's per capita electricity consumption — excluding South Africa — is roughly 180 kilowatt hours per year. Europe's is around 6,500. The United States sits at approximately 13,000. A single AI training run at a frontier lab can consume more electricity than some African nations generate in a month.
You cannot run a data centre on 180 kilowatt hours per capita. You cannot train a model. You cannot host inference. You cannot even keep the servers cool. Before the AI divide is a question of talent, regulation, investment, or ambition, it is a question of watts. And the watt gap is not closing.
Electrification and the shape it takes
Africa's overall electrification rate sits at roughly 47% in Sub-Saharan Africa. But that headline number hides a pattern that matters more than the average.
Electrification does not spread evenly. As GDP grows, power flows to capital cities, industrial corridors, port zones, and extraction sites. Rural areas do not catch up. They fall further behind. The urban-rural split is not a transitional phase on the way to universal access. In many countries, it is the stable equilibrium. Cities get power because cities generate revenue. Rural regions remain dark because the return on connecting them is low.
This is the shape electrification takes when it is driven by market logic rather than universal provision. And it has a direct consequence for AI: the infrastructure that could support compute capacity will cluster in the same few nodes that already have power — the capital cities of the continent's largest economies, and almost nowhere else.
Kenya has made real progress, pushing electrification from 32% to 75% in under a decade. That is a genuine achievement. But Kenya is also the exception that proves the pattern: its gains were driven by specific policy interventions and sustained investment that most nations on the continent have not been able to replicate. And even Kenya's gains are concentrated along the Nairobi-Mombasa corridor. Step outside those zones and the picture changes fast.
The continent holds 18% of the world's population and produces roughly 2% of global GDP. Its average per-capita electricity consumption is about a quarter of the global average. These are not temporary conditions. They are structural. And they define the boundary within which any conversation about African participation in the AI economy must take place.
The mine that powers itself
There is an image that has stayed with me since I started thinking about this series.
The Kibali gold mine sits in the northeast of the Democratic Republic of Congo, in the Haut-Uélé province. It is one of the largest gold mines in Africa. It is owned by a joint venture between Barrick Gold of Canada and AngloGold Ashanti of South Africa, with a 10% stake held by the Congolese state mining company SOKIMO. Barrick operates it.
Kibali generates its own power. Three hydroelectric stations produce 44 megawatts. A 16-megawatt solar farm was recently commissioned. Battery storage backs up the system during the dry season. At peak, the mine runs on 85% renewable energy. For six months of the year, it operates on 100% renewables. It is, by any measure, a sophisticated energy system — microgrid management, prescriptive maintenance, battery-backed frequency control. Cutting-edge engineering.
None of this was built to electrify the region. It was built to extract gold.
When Barrick started construction roughly fourteen years ago, the surrounding area was described as one of the DRC's most underdeveloped regions. The community around the mine has grown from about 30,000 to over 500,000 people, and Barrick has built 300 kilometres of road, schools, and clinics. One of the hydro plants does supply some electricity to nearby communities. These are real contributions.
But the fundamental dynamic is worth stating plainly: a foreign mining company built a world-class energy system in one of the poorest regions on earth, primarily to power an extraction operation whose output — gold — leaves the country. The region did not get a power grid. It got a mine that happens to have one.
This is not unique to Kibali. It is the pattern. Across the continent, the most reliable infrastructure often belongs to extractive industries. The roads that work are the ones connecting mines to ports. The power that is stable is the power that feeds industrial operations. The connectivity that exists serves export logistics.
AI infrastructure will follow the same grooves unless something changes. Data centres need stable power. They will go where stable power exists. Stable power exists where economic returns justify the investment. In much of Africa, that means extraction zones, capital cities, and almost nowhere else.
Cloud rent on a continent that does not own the cloud
In Part 1, I described Varoufakis's argument that cloud platforms have become the new land — assets that everyone must rent access to in order to participate in the economy. That argument sharpens considerably when you look at it from the African continent.
Africa holds under 1% of global data centre capacity. The hyperscale providers — AWS, Azure, Google Cloud — have a combined physical presence in exactly two African countries: South Africa and, more recently, limited deployments in a handful of others. For the majority of the continent, using cloud services means routing data to servers in Europe or the Middle East. Every API call, every model inference, every byte of storage crosses an ocean, accrues latency, and generates revenue for a company headquartered in Seattle or Redmond.
This is cloud rent in its purest form. African businesses, developers, governments, and institutions pay to access infrastructure they have no ownership stake in, no governance role over, and no ability to influence the pricing or terms of. The data flows out. The money flows out. The capability remains elsewhere.
And unlike physical rent, cloud rent scales. A business in Lagos or Luanda paying for AWS compute is not just renting a server. It is renting the entire stack: the hardware, the operating system, the networking, the orchestration layer, the AI APIs, the model weights, the training data that those models learned from. At every layer, value was created somewhere else and is being accessed under terms set by someone else. The tenant has no leverage. The landlord has no obligation to invest locally.
The mobile phone revolution is often cited as evidence that Africa can leapfrog infrastructure limitations. And it is true that mobile adoption was extraordinary — precisely because a phone requires minimal infrastructure. A single tower covers a wide area. Handsets are cheap. The barrier to entry was low enough that market forces alone could drive adoption.
AI is the opposite of that. It is infrastructure-heavy, capital-intensive, power-hungry, and dependent on specialised supply chains. It cannot leapfrog. It needs the grid, the fibre, the cooling, the hardware, the talent pipeline. Without those, adoption means consumption, not production. It means renting, not owning.
The resource paradox
Many of the minerals critical to AI hardware — cobalt, lithium, tantalum, coltan — are extracted from African soil. The DRC alone produces roughly 70% of the world's cobalt, a key component in the batteries that power everything from electric vehicles to data centre backup systems. Those minerals are mined in Africa, shipped to refineries in China or Southeast Asia, manufactured into components in East Asia, assembled into hardware in factories around the world, and installed in data centres that African companies then pay to access.
The value chain starts in African ground and ends in cloud rent paid by African tenants. At no point does the continent capture the high-value portion of that chain. It exports raw materials and imports finished services. This is not a new pattern. It is the oldest pattern. Colonialism worked this way. The commodity has changed — from rubber and copper to cobalt and data — but the direction of extraction has not.
Varoufakis's framework gives this a useful name. The cloud landlords extract rent from their platforms. But those platforms are themselves built on a physical supply chain that extracts resources from the very populations who will be charged the highest relative cost to access the finished product. It is feudalism layered on top of colonialism, intermediated by technology.
What is not working
The standard responses to this problem tend to fall into a few categories, and most of them are inadequate.
"Invest in STEM education" is popular and partially correct, but it puts the burden on the tenant to become useful to the landlord. A well-trained engineer in Nairobi or Accra can get an excellent job — at a company headquartered in San Francisco, contributing to infrastructure owned in Virginia, serving customers in Europe. The talent develops. The infrastructure does not follow.
"Attract foreign investment" assumes that the interests of foreign capital align with local development. Sometimes they do. Often they do not. Kibali is a useful illustration: the investment is real, the infrastructure is genuine, and the primary beneficiary is a Canadian mining company. The region gained roads and clinics. It did not gain sovereignty over its own energy system.
"Build local startups" is appealing and important, but it runs into the infrastructure wall almost immediately. A startup in Lagos can build an excellent application. It will deploy it on AWS. It will use an API from OpenAI or Anthropic or Google. It will store data on servers in Cape Town or Dublin. Every layer of the stack generates rent for someone else. The startup may succeed. The dependency deepens regardless.
These efforts are not pointless. But without addressing the structural question — who owns the infrastructure — they operate within the constraints of the existing system rather than changing them.
The question underneath
The question I keep coming back to is not "how does Africa catch up in AI." That framing already concedes too much. It assumes a race with fixed rules, run on someone else's track.
The better question is: what does it look like to build AI infrastructure that serves the people who live here, rather than extracting value from them?
That is a different kind of engineering problem. It involves energy sovereignty, data governance, compute ownership, and the political will to treat digital infrastructure as a public good rather than a market opportunity. It is hard. It is expensive. And it will not happen by accident.
In Part 3, I will talk about what it looks like to start. Not at the policy level. At the level of a keyboard and a language that the machines do not speak yet.
This is Part 2 of a three-part series on techno-feudalism and the AI divide. ← Part 1: Cloud Rent and the New Landlords | Part 3: The Lion Learns to Write →