The Compute Divide: Why Africa Risks Being Locked Out of AI

Every serious conversation about African AI eventually hits the same wall, and it is not talent, and it is not ideas. It is compute. The graphics processing units, data centres and cheap electricity that turn an AI idea into a working product are overwhelmingly located outside the continent, and access to them is metered in hard currency that African startups do not have. My thesis is uncomfortable but necessary: unless Africa treats compute as critical infrastructure - the way it treats roads, ports and power - it risks being permanently locked into renting intelligence from abroad while everyone else owns it.

What the compute divide actually is

Modern AI runs on specialised chips, mostly Nvidia GPUs, packed into data centres that draw enormous amounts of electricity. Training and even running large models is expensive, and demand has so outstripped supply that GPUs became one of the most fought-over commodities on earth. For a founder in Nairobi, Lagos or Kigali, the practical experience of this shortage is simple: you pay foreign cloud providers in dollars, at premium rates, for capacity that sits on another continent, with latency and data-residency headaches attached.

As Cassava Technologies CEO Hardy Pemhiwa put it, at its very base, the cost of AI is going to be the cost of energy. That single sentence explains why the compute divide overlaps almost perfectly with the electricity divide. You cannot run an AI factory on an unstable grid.

The three shortages stacked on top of each other

Africa is not facing one gap. It is facing three at once.

• Chips. GPUs are scarce and expensive globally, and Africa sits at the back of the allocation queue behind US and Chinese hyperscalers.
• Data centres. The continent has a tiny share of the world's data-centre capacity relative to its population, and most of it is not AI-optimised.
• Power. AI facilities are extraordinarily power-hungry, and grid stability is now widely described as the primary bottleneck for AI expansion. Many African grids cannot absorb that load without an overhaul.

Stack these and you get a compounding disadvantage. A brilliant model trained on African languages is worth little if there is nowhere affordable to train and serve it at home. And the three gaps reinforce each other: without reliable power you cannot attract data-centre investment, without data centres you cannot justify importing scarce GPUs, and without local GPUs your best engineers keep renting foreign capacity that drains hard currency out of the economy. Each missing piece makes the next one harder to secure, which is exactly how a divide hardens into a permanent structural gap rather than a temporary lag.

The money is starting to move, but read it carefully

It is not all bleak, and I refuse to write a doom piece. Real capital is arriving.

• Nvidia and Cassava Technologies are partnering to deploy thousands of GPUs across Africa Data Centres facilities, with reports of up to 12,000 GPUs through an AI Factory build-out and hundreds of millions of dollars targeted for expansion. The explicit goal is to end the dependence on expensive foreign cloud credits.
• Microsoft and the UAE's G42 announced a roughly $1 billion package for Kenya, including a green data centre for an East Africa Azure region.
• MTN, Airtel and others are building AI-optimised, GPU-ready capacity in Nigeria, with facilities costing hundreds of millions of dollars coming online through 2026.

The World Economic Forum has argued that investment in green compute could unlock enormous economic value across Africa. So the thesis is not that nobody is building. It is that who owns the build matters enormously.

Renting compute is fine when you are experimenting. It becomes a trap when your entire economy's intelligence runs on infrastructure you neither own nor control.

The dependence trap

Here is the strategic risk. Much of this new capacity is being financed and controlled by foreign hyperscalers and chip vendors. That is genuinely useful in the short term - it puts GPUs on African soil and cuts latency. But if Africa's AI future is simply a set of foreign-owned data centres physically located here, we have moved the building without moving the ownership. Pricing, priorities and access still sit with the parent company, and geopolitics can change access overnight. We have seen how quickly technology can become a bargaining chip in trade disputes.

Sovereignty is not xenophobia about foreign investment. It is refusing to let the most important input to the next economy be a tap that someone else can turn off.

Paths to compute sovereignty

What would a serious response look like? Not one silver bullet, but a portfolio.

1. Treat compute as national infrastructure. The AU's proposed AI Fund and Pan-African Centre are the right instinct. Pool continental demand so Africa negotiates with Nvidia and the hyperscalers as a bloc, not as 54 separate small customers.
2. Solve power and compute together. Africa's renewable potential - solar, geothermal in the Rift Valley, hydro - is a genuine competitive advantage for green data centres. Kenya's geothermal base is a real asset here. Pair every serious data-centre plan with a dedicated clean-energy plan.
3. Build shared public compute for researchers and startups. Not every university and founder needs to rent frontier GPUs at dollar rates. National or regional AI compute clusters, subsidised and shared, are how smaller nations elsewhere have bootstrapped capacity.
4. Lean into models that do not require frontier training. You do not need a $100 million training run to deliver value. Fine-tuning and serving open-weight models on modest hardware is a realistic path, which I explore in the companion piece on open versus closed models.
5. Negotiate ownership and skills transfer into every deal. When a hyperscaler builds here, the terms should include local ownership stakes, local jobs, and training - not just a badge that says the servers are physically in-country.

What this means for African businesses today

If you are a founder, do not wait for continental infrastructure to be ready. Be deliberate about your compute strategy now. Understand exactly what you are paying foreign providers, in which currency, and what your exposure is if prices spike or access tightens. Design your products so they can run on cheaper, smaller models where possible. And resist the temptation to burn scarce capital on the most expensive infrastructure before you have proven demand - a discipline that connects directly to why I warn businesses against buying AI before your business is ready.

The compute divide is the single most physical, most concrete of all the AI challenges facing Africa. You cannot code your way around a chip shortage or a weak grid. But the divide is not destiny. The capital is arriving, the renewable-energy advantage is real, and the continental will to coordinate exists on paper. The task of this decade is to make sure that when the AI economy is fully built, Africa is an owner in the room, not a tenant paying rent in dollars.