In brief
Africa’s role in the global AI economy is increasingly defined by the intersection of energy infrastructure and compute demand. As AI drives exponential growth in data‑centre power consumption, electricity availability has become the primary constraint on expansion, particularly in South Africa. Despite projected growth, the continent remains under‑represented in global capacity, highlighting a widening scale gap. Developers are responding by integrating renewable energy, storage and innovative power arrangements to ensure project viability. At the same time, AI pricing models are shifting toward usage‑based structures, bringing the true cost of compute—driven largely by energy—into sharper focus. For businesses and investors, the evolving relationship between megawatts (MW) and token economics will be central to the future of AI deployment in Africa.
Key takeaways
- Power is the constraint: Power availability is now the longest-lead bottleneck in data centre deployment, ahead of capital or chips.
- The scale gap: Africa holds just 0.6% of global data centre capacity. Projecting to triple to ~1.2 GW by 2030 merely tracks global expansion without closing the gap.
- Upstream engineering: Developers are integrating utility-scale solar, virtual wheeling, and battery storage from the outset to manage transmission risks and ensure bankability.
- Realised costs: With software providers shifting to consumption-tied pricing, the core metric is moving from headline token prices to the realised cost per completed task.
In more detail
The megawatt is the unit of AI
Conversations about energy infrastructure typically begin on the supply side, with generation, transmission, storage and the path toward a renewable-led portfolio. However, the most consequential new source of demand now sits at the other end of the wire. Artificial intelligence (AI) has turned the data centre into one of the fastest-growing loads on any modern grid, and the unit in which that demand is now measured is the megawatt.
For most of the digital era, the sector was a remarkably light consumer of power: through the 2010s, efficiency gains, the shift to hyperscale facilities and steadily improving power usage effectiveness allowed digital services to expand rapidly while the related electricity demand stayed broadly flat. AI has broken that coupling, because the largest historical efficiency levers have already been pulled and accelerated GPU computing draws power on a vastly different scale. Globally, the digital sector could consume more than 1,000 terawatt-hours of electricity in 2026, close to the annual consumption of Japan, according to the International Energy Agency (IEA). Demand from data centre facilities alone rose 17% in 2025, while electricity use by AI-specific data centres climbed roughly 50% in the same year. A genuine paradox sits underneath these numbers: the power drawn per AI task is actually falling rapidly as hardware and models become more efficient, yet aggregate demand keeps rising because more people use AI and the heaviest new uses, such as autonomous agents in new ‘agentic’ AI ecosystems, are the ones that consume the most compute. On current trajectories, the IEA expects data centre consumption to double by 2030, with AI-driven demand tripling.
The capital behind this is both vast and concrete. Combined capital expenditure by the five largest technology companies passed USD 400 billion in 2025 and is set to rise by around three-quarters again in 2026. The flagship United States programme, Stargate, carries a headline commitment of USD 500 billion and 10 gigawatts of capacity, with close to 7 GW of sites already announced and a single flagship campus rated at 1.2 GW. One gigawatt can power tens of thousands of AI accelerators (i.e., specialized hardware designed to speed up artificial intelligence and machine learning tasks). Power, ahead of chips or buildings, has become the longest-lead constraint, with power purchase agreements (PPAs) for large facilities in established markets now carrying waits of three to five years. The physical world has become the bottleneck for the digital one, and operators are increasingly funding their own generation and storage rather than waiting in the interconnection queue.
Africa’s place in the build-out
Against this backdrop, Africa’s position can be described as marginal in scale, yet consequential in stakes. The continent holds an estimated 0.6% of global data centre capacity while home to roughly 18% of the world’s people, on figures from the Africa Data Centres Association and the World Economic Forum. The United States alone hosts close to 45% of global data centre capacity. Installed capacity on the African continent, spread across some 220 to 230 facilities in 38 countries and concentrated in South Africa, Nigeria, Kenya and Egypt, is forecast to roughly triple to about 1.2 GW of IT load by 2030. The sobering qualifier is that this growth tracks global expansion rather than closing the gap. Putting that simply, the entire continent’s projected 2030 capacity is about the size of just one large American AI campus.
The contrast with the Gulf is instructive, because it shows what targeted policy and cheap, reliable power can achieve. State-backed programmes and integrating power, land and sovereign demand from the outset, have produced multi-gigawatt, AI-native platforms. This includes the UAE-US AI Campus, a 10-square-mile site in Abu Dhabi unveiled in May 2025, which is billed as the largest outside the United States at 5 GW. The Middle East and Africa AI data-centre market is projected to grow from roughly USD 2.5 billion to USD 8.2 billion between 2026 and 2031, a compound annual growth rate close to 27%. Much of that reflects deliberate sovereign strategy rather than opportunistic co-location. Africa’s challenge is to convert its own advantages, above all its solar resource, into a comparable, bankable signal to investors.
South Africa: engineering around the grid
South Africa is the continent’s springboard, and its trajectory makes the energy stakes clear to understand. As at Q1 2026, data centre IT load sits near 190 MW for the largest operator alone, and recently announced expansion plans could push national IT load past 1,200 MW, equivalent to roughly one stage of load shedding. New AI-focused facilities draw 100 MW or more each, against 10 to 25 MW for a conventional data centre, and rack densities for AI workloads run four to six times higher than traditional deployments, frequently exceeding 30 kilowatts per rack. The response has been to engineer around grid scarcity through utility scale solar feeding campuses via the Eskom virtual wheeling framework, long-term renewable purchase agreements, and on site generation and storage. These arrangements also meet the carbon targets demanded by global cloud tenants, which makes clean power a commercial requirement. Power availability, ahead of connectivity, is now the principal constraint on data centre build-out, and operators are building facilities around where firm, affordable electricity, and increasingly water for cooling, can be secured.
The bankability of these arrangements turns on several factors. Creditworthy offtake from hyperscalers or co location operators has supported longer-tenor PPAs, but lenders remain focused on transmission risk, particularly curtailment and grid access constraints. Wheeling frameworks, while increasingly standardised, still depend on municipal and Eskom cooperation, creating fragmentation across jurisdictions. As a result, developers are favouring hybrid structures that combine utility-scale solar with on-site generation, battery storage, and in some cases backup gas or diesel capacity to ensure reliability.
Across the region, similar dynamics are emerging but with wider dispersion in regulatory frameworks. Markets such as Kenya and Egypt offer clearer pathways for independent power production, while others remain constrained by vertically integrated utilities. The common theme is that successful data centre build-outs are being designed around integrated energy solutions from the outset, rather than treating power as a downstream input. Power structuring has become a core workstream in project development, not an ancillary consideration.
From subsidised compute to the true cost of a token
The downstream story is quickly turning from physics to economics, and this is where the next two years diverge sharply from the last two. Through 2023 the cost conversation was dominated by AI model training, a one-time expense that only the largest laboratories could bear. This centre of gravity has, however, since shifted to AI inference, i.e., the cost of actually running models in production, which now accounts for the large majority of AI compute spend. The unit cost of inference has fallen sharply, with the price of a million tokens across major providers dropping from roughly USD 10 to USD 2.50 in a single year. Nonetheless, the bill has risen for most adopters, because agentic workflows consume tokens in multiples of what earlier applications did. The cost of a unit of intelligence is falling while the cost of deploying intelligence at scale climbs.
What many don’t yet realise is the pricing of AI inference most enterprises see today rests on a subsidised foundation. Frontier labs and LLM providers, backed by venture capital and hyperscaler balance sheets, have been pricing inference below cost to capture market share, creating what analysts describe as a ‘false floor’ in the market. That floor is now lifting. By 2026, an estimated 85% of software providers had moved to usage-based or hybrid pricing tied directly to consumption, and credible analysis points to API price normalisation of 30 to 50% over the following 18 months as capital discipline returns to the sector. The era of effectively flat-rate, all-you-can-consume AI is closing, and a pay-as-you-go pricing model is taking its place. Token economics, until recently a niche concern of infrastructure teams, is becoming a board-level budget line for non-frontier enterprises, and the meaningful unit of measurement is shifting from the headline price per token to the realised cost per completed task.
The grid sets the cost of intelligence
A token is, at root, a small quantity of computation, and computation amounts to electricity and hardware, amortised over time. As subsidies recede and providers price toward genuine unit economics, the cost of power becomes visible in the cost of AI inference. Where electricity is expensive or unreliable, inference becomes expensive. Where it is cheap, firm and clean, the economics of running models locally improve markedly. The value chain here runs in a straight line: the grid sets the cost of the megawatt, the megawatt sets a large part of the cost of the token, and the token sets the cost of the AI solution that an African business can actually afford to buy.
It is important to note that two forces already pull inference onshore for African countries like South Africa, which doesn’t have a general data localisation requirement baked into its data protection laws. The first is policy direction rather than a general mandate: the National Data and Cloud Policy, adopted in 2024 and still being implemented, signals a clear sovereignty orientation and requires that government data bearing on national security be stored within the Republic, even as Protection of Personal Information Act (POPIA) continues to permit cross-border processing where adequate protection travels with the data. The second is commercial and technical: latency, the economics of serving local users, and the preferences of regulated sectors such as financial services point the same way. Today, many African enterprises still run workloads in offshore or South African cloud regions because local accelerator capacity is scarce and expensive. As the World Economic Forum and the Africa Data Centres Association frame it, the binding constraint is the alignment of Graphics Processing Unit (GPU) economics, power costs, and uncertain demand, rather than the supply of buildings, compounded by capital mispriced against the continent’s actual default record. Cheap, firm renewable power is one of the few levers capable of moving several of those variables at once and falling bandwidth costs as undersea cable capacity expands reinforce the shift.
Compute as a new class of energy load in Africa
For the developers, investors and advisers, the practical implication is that AI compute is emerging as a new class of anchor load with an unusual profile. This industry is large, may be willing to sign long-dated power purchase agreements, and it will likely pay a premium for reliability. However, it is also volatile, producing rapid and deep swings in demand that strain conventional on-site gas turbines. That volatility is precisely why on-site battery storage is moving from a refinement to a core design element, and why a well-configured AI campus, with storage and flexible load, can become an asset to the grid rather than only a burden on it. Renewables are already projected to meet nearly half of the growth in global data centre electricity demand through 2030.
At a continental level, the opportunity sits in pairing Africa’s natural resource advantage with scalable project frameworks. Cross border power trading, regional transmission build-out, and harmonisation of regulatory regimes will be critical to unlocking larger, multi-gigawatt platforms. At the same time, investors are increasingly focused on grid resilience, recognising that data centre developments both depend on and can support system stability when designed with flexible load and storage capability.
Looking ahead
The temptation is to treat AI as a demand problem; however the more useful framing is that Africa’s energy endowment, its solar resource and its room to build storage-led capacity, is a compute advantage waiting to be priced in. As the subsidised era for AI inference ends and the true cost of a token surfaces, the question for the continent is whether it builds the firm, affordable power that turns an energy story into a compute story, or whether it tracks global growth while the gap stays exactly where it is. The megawatt and the token will very quickly become the same conversation, and over time, the energy discussion will direct the cost of AI in Africa.