The $700 Billion Compute Divide

American technology companies are committing close to $700 billion in capital spending this year, most of it aimed at AI infrastructure, and roughly double what the same firms spent the year before. That is not a product cycle. It is the physical layer of the next economy being poured, and where it gets poured is being decided now.

The number the comparison needs

Calling Africa's position a rounding error is easy and useless without the figure, so here it is. The continent has roughly 360 megawatts of active data centre capacity, with a few hundred more under construction and a larger pipeline announced. Global installed capacity is on the order of 122 gigawatts.

That puts Africa at well under one percent of global capacity, against roughly 19 percent of the world's population. Inside the continent the concentration is tighter again, with South Africa, Kenya and Nigeria holding about 41 percent of what exists.

Those two numbers next to each other are the whole story. A fifth of the world's people, a rounding error of the world's compute, and most AI workloads touching African users served from somewhere else.

The gap also tells you something about latency and sovereignty that the megawatt figures alone do not. Workloads served from outside the continent carry a round trip that is measurable to the user and a jurisdiction question that is measurable to the regulator. Both of those become commercial problems for anyone building on top of them, which is how physical capacity turns into a business constraint rather than a statistic.

The scarce input is not chips

The instinct is to read this as a capital gap, and to conclude that Africa cannot compete because it cannot match hyperscaler budgets. That misreads what the budget buys. Compute is infrastructure now, and infrastructure is geography. The scarce inputs are power, cooling, land near both, and fibre to reach them.

A data centre is a power customer first and a technology asset second. A hundred megawatts of committed, reliable, affordable electricity is the hard part, and it is hard in exactly the places where the population is. That is why the constraint on African capacity is generation and transmission rather than willingness to invest.

It also means the competition is not with hyperscalers. Nobody in Lagos is outbidding Microsoft on chips. The position available is the one underneath, which is the power, the site and the connectivity that any operator needs before a rack goes in.

South Africa, Kenya and Nigeria holding most of the continent's capacity is not a coincidence either. Those are the markets with the grid, the fibre and the regulatory clarity to make an interconnection agreement worth signing. Capacity follows power and rules, and it will keep following them regardless of which market has the largest population.

Why the timing matters

Capacity decisions compound. A region that gets a facility gets the latency, then the workloads, then the developers who build for local latency, then the next facility. A region that does not gets served from abroad indefinitely, and every year of that makes the local build harder to justify commercially.

The share of global capacity Africa holds in 2035 is being determined by power projects that reach financial close in the next few years, not by AI policy announcements.

There is a version of this that goes right, and it does not require anyone to out-build a hyperscaler. It requires enough reliable generation in two or three additional markets that a colocation operator can sign an anchor tenant. That is a power project problem with a technology customer attached, and it is financeable by people already in this market.

The window for that is defined by other people's decisions rather than by African readiness, which is the uncomfortable part. Capacity commitments being made in the next few years set where the workloads sit for the decade after.

The case against acting on this

The strongest objection is that this is an AI capex cycle that corrects. If hyperscaler spending stalls, demand for frontier-market capacity never materialises and anyone who built power for it has stranded an asset against a thesis rather than a contract.

That objection is serious, and the answer to it is the shape of the bet rather than its direction. Power and connectivity have demand whether or not the AI build reaches the continent. Speculative capacity built for workloads that have not been contracted does not. The first is infrastructure. The second is a wager wearing infrastructure's clothes.