Africa’s GPU race has begun. Power, price and sovereignty will decide who wins

Africa’s AI debate is moving out of the conference hall and into the machinery.

On 12 August, Smart Africa closed an expression-of-interest process seeking private-sector partners capable of supplying GPU-enabled computing across its member states. The document is not a contract award, and it does not commit Smart Africa or any government to buy capacity. But it is still consequential: it turns the familiar promise of “AI for Africa” into a practical test of infrastructure, economics and control.

The proposal imagines a federated marketplace in which public institutions, researchers, startups and companies could access GPU capacity on demand—even when that capacity is hosted in another participating country. Providers may include hyperscale cloud companies, data-centre operators and specialist GPU-as-a-service businesses, working alone or in consortia.

That sounds technical. It is also political.

The companies and countries that shape this layer will influence what African researchers can build, how much startups pay to train and run models, where sensitive workloads are processed, and whether local operators participate in the value created around AI.

The first contest is affordability

Compute scarcity is not solved merely by installing more accelerators. It is solved when the right users can access them at a price and on terms that match their work.

Smart Africa asks prospective providers to support transparent consumption models such as pay-as-you-go, reserved capacity or subscriptions. The proof-of-value phase is expected to test availability, performance, provisioning time, cost, security, interoperability, energy efficiency and user satisfaction.

Those measurements matter because a nominally African compute market could still reproduce the access problems of the global cloud economy. Capacity can exist while remaining too expensive for university labs, early-stage founders and public-interest teams. A successful system needs more than a catalogue of GPUs. It needs prices that researchers and startups can actually sustain, predictable access during periods of high demand, and a route for smaller users that cannot sign enterprise-scale contracts.

The unanswered question is who absorbs the early risk. Smart Africa says the proof-of-value is expected to rely mainly on provider contributions and negotiated commercial terms. It offers no guaranteed upfront capital expenditure, minimum purchase commitment or automatic government contract. That protects the institution from promising money it has not committed, but it also means providers will decide how much capacity they are prepared to expose before demand is proven.

The second contest is electricity

There is no independent AI-compute strategy without an energy strategy.

Smart Africa asks providers to show that their facilities have adequate power, cooling, security and connectivity. It also asks for current or targeted power-usage effectiveness, information about renewable or green energy sourcing, and alignment with regional grid conditions.

The International Energy Agency reported in April that electricity demand from data centres increased by 17% in 2025, with AI-focused facilities growing faster. Its wider analysis warns that the impact of a data centre can be highly concentrated in the local grid even when the sector remains a relatively small share of global electricity use.

For African markets, the meaningful question is therefore not simply how many GPUs can be announced. It is where they will run, what power they will use, how reliably they will operate, what cooling conditions they require and whether surrounding communities or businesses carry any hidden infrastructure cost.

A compute project that depends on fragile electricity, expensive backup generation or weak connectivity may be African in location while remaining commercially inaccessible in practice.

The third contest is sovereignty without isolation

Smart Africa’s proposal attempts a difficult balance. It calls for cross-border capacity sharing and open interoperability while also requiring compliance with national data-protection, localisation and sovereignty rules.

Those goals can pull in different directions. A pooled market becomes more useful when workloads can move between providers and jurisdictions. Governments and regulated organisations, however, may need certain data or processing to remain within a defined territory. The hard work sits in identity, workload isolation, auditability, billing, portability and enforceable contracts—not in the marketing language of a “sovereign cloud.”

The African Union’s Continental AI Strategy already identifies reliable electricity, broadband, data centres, cloud infrastructure and computing power as foundations for meaningful participation in AI. It also argues for a people-centred and development-oriented approach. The new compute initiative should therefore be judged by more than capacity installed. It should be judged by who gains useful access and who retains leverage.

The case for global scale

Large cloud and infrastructure companies can bring mature platforms, procurement experience, security controls, high-performance networks and the ability to deploy capital across several markets.

Google’s July Cloud Summit in Johannesburg illustrates the scale of that approach. The company announced an Eastern Cape connectivity hub linked to new international routes, an applied AI lab in Accra and further programmes for founders and skills. These investments can expand the technical surface available to African companies and institutions.

Global providers may also make it easier for developers to connect compute with models, storage, security and deployment tools they already use.

The case for local leverage

Scale alone does not guarantee local ownership, affordable pricing or meaningful technology transfer.

African data-centre operators, telecommunications companies, cloud providers, universities and specialist infrastructure firms understand local licensing, connectivity, power constraints and procurement environments. Their participation could keep more operating knowledge, employment and commercial value within the continent.

The strongest outcome may not be a symbolic choice between global and local providers. It may be a system that makes global scale contestable: open interfaces, portable workloads, transparent prices, enforceable data rules and room for African operators to compete as more than resellers.

What changes next

The expression-of-interest deadline has passed, but the result is not yet the continental GPU market described in the document. The next evidence should be concrete.

Which providers were shortlisted? Which countries and facilities will participate? What capacity will be available for training, fine-tuning and inference? What will one usable unit of compute cost? How will scarce capacity be allocated between governments, researchers, startups and larger companies? Which power sources and grid agreements will support the facilities? What information will be published about uptime, utilisation, security incidents and user outcomes?

Until those answers arrive, the initiative should be treated as a serious procurement experiment—not as completed infrastructure.

That distinction is not pessimism. It is the difference between an announcement and an operating system. Africa’s AI opportunity will be shaped by models and applications, but the more durable contest is underneath them: who owns the compute, who powers it, who can afford it and who gets to set the rules.