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

Africa’s AI infrastructure deserves serious attention. For those of us building businesses, the test is practical: whether a developer, researcher or small company gets reliable computing power at a price their work supports.

I want to see African founders build useful products, serve paying customers and retain control over their businesses. Every infrastructure announcement should bring us closer to those outcomes.

Cassava Technologies offers one example of this ambition. In its 18 March 2026 update, the company described deploying an NVIDIA-powered AI factory in South Africa, with expansion planned for Nigeria, Kenya, Egypt and Morocco. Its offering includes access to GPUs and AI services. GPUs are processors used to accelerate the calculations behind many AI systems. [Cassava’s deployment update](https://www.cassava.ai/2026/03/18/cassava-scales-african-ai-infrastructure-with-nvidia-powered-ai-factories-to-accelerate-sovereign-data-capabilities/)

The distinction between deployment and planned expansion matters. An announced location gives us something to follow. An operational service gives a founder something to build with.

We should measure progress through available capacity, customer access and performance under real demand. My view is straightforward. Power, price and sovereignty will determine how much value these investments create for African businesses.

Electricity comes first because every promise about AI depends on a functioning physical system.

The International Energy Agency’s 2026 analysis estimates global data-centre electricity consumption reached 485 terawatt-hours in 2025. Its central projection puts consumption at roughly 950 terawatt-hours by 2030. Those figures cover data centres overall. Electricity consumption from AI-focused facilities is projected to triple over the same period. [IEA’s 2026 energy and AI analysis](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary)

For any proposed African facility, I want to understand the electricity supply, the connection timetable and the cost of maintaining service during an interruption. Cooling, maintenance and backup systems belong in the same conversation.

An investor should also explain how the project contributes to the surrounding electricity system. New generation, grid upgrades and a fair allocation of costs deserve attention from the beginning.

The IEA makes an essential distinction here: the effect of data centres on electricity prices depends on local supply conditions and policy choices. Additional demand puts pressure on a constrained system. Where spare supply exists, it offers an opportunity to spread infrastructure costs across greater consumption. [IEA’s assessment of electricity affordability](https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary)

I would therefore judge a project by its actual power arrangements. A credible service commitment needs functioning infrastructure behind it.

Price is the next test, and I mean the full cost of delivering a useful result.

As a founder, I would assess an AI service through the cost of completing the work my customer needs. An advertised GPU-hour rate is only one part of the calculation. Storage, data transfers, support, minimum commitments and idle capacity belong in the budget.

Consider a hypothetical comparison using one GPU on each service. One charges $2 per GPU-hour and completes a workload in four hours. Another charges $3 per GPU-hour and finishes the same workload, at equivalent quality, in two hours. The processing costs are $8 and $6 respectively, before other charges.

The higher hourly rate produces the lower bill.

Hardware differences help explain why hourly prices alone tell us so little. Memory bandwidth affects how quickly data reaches the processor. For work spread across multiple GPUs, the connections between them also affect performance. NVIDIA’s technical guidance identifies memory and interconnect bandwidth as factors worth measuring for the workload being run. Neither establishes a fixed speed advantage without testing. [NVIDIA’s guidance on memory and interconnect performance](https://developer.nvidia.com/blog/nvidia-nvbandwidth-your-essential-tool-for-measuring-gpu-interconnect-and-memory-performance/)

This is why I want providers to publish clear pricing and offer realistic trials. Founders need to test their own workloads, understand the charges and establish what happens after introductory credits expire.

For a business earning in local currency, a dollar-denominated bill also introduces exchange-rate exposure. Local billing would make planning easier, although the supplier’s own imported equipment and financing costs still need to be paid.

We should be precise about affordability. Hosting equipment in Africa does not establish a lower total price. Providers need to demonstrate the advantage through comparable workloads and transparent invoices.

They also face established global competitors.

AWS operates its Africa region in Cape Town. Microsoft lists Azure’s South Africa North region in Johannesburg, where Google Cloud also operates a region. These are existing cloud locations, although a regional presence alone does not establish availability of a particular GPU or AI service. Buyers need to verify the exact service and location required. [AWS Cape Town](https://aws.amazon.com/local/africa/cape-town/), [Microsoft’s region directory](https://learn.microsoft.com/en-us/azure/reliability/regions-list), [Google Cloud Johannesburg](https://cloud.google.com/blog/products/infrastructure/heita-south-africa-new-cloud-region)

Startup incentives add another competitive pressure. AWS Activate offers credits to eligible startups, while Google’s startup programme also provides credits subject to eligibility and programme conditions. [AWS Activate](https://aws.amazon.com/startups/credits/), [Google Cloud startup benefits](https://cloud.google.com/startup/benefits)

My assessment is these incentives make global providers a serious option for a founder managing limited cash. African providers need a persuasive commercial offer covering performance, support, access and long-term cost.

Founders, in turn, should model the bill after the credits run out. A subsidised prototype needs a credible route to an affordable production service.

Then comes sovereignty, a word which deserves more demanding scrutiny.

For me, meaningful control includes knowing where information is processed, who administers the systems, which suppliers are involved and what happens when a customer wants to leave.

A business should understand how to export its data, move its application and continue operating if a provider changes its terms. It should also understand the licences governing the models it uses and whether its information is retained or reused.

These are purchasing requirements I would put into a supplier assessment. They turn an ambitious claim about sovereignty into specific commitments a customer understands.

Local infrastructure offers an opportunity to strengthen this control. The value depends on contracts, technical design and the skills of the people operating the service. I would apply the same scrutiny to an African provider and an overseas supplier.

Regional cooperation deserves equal attention. The African Union’s Continental Artificial Intelligence Strategy, endorsed in July 2024, calls for coordinated national approaches and stronger regional and international cooperation. [African Union’s Continental AI Strategy](https://au.int/en/documents/20240809/continental-artificial-intelligence-strategy)

I would like to see this translated into shared research access, practical procurement standards and opportunities for smaller businesses to buy computing services across participating markets.

A university should have a clear route to apply for research capacity. A startup should understand the eligibility rules for support. Public funding should come with published access criteria and measurable outcomes.

The products built on this infrastructure deserve as much attention as the infrastructure itself.

My interests sit across technology, media, education and community resilience. From this perspective, I see work worth testing: transcription for African media, language tools assessed by local speakers, learning support aligned with a specific curriculum, and business software which reduces repetitive administration.

Each needs evidence. A language service needs testing across accents and everyday speech. An education product needs subject review. A business tool needs to demonstrate time saved, accuracy and a price customers accept.

Local hosting alone does not deliver those outcomes. Teams still need appropriate data, permission to use it, clear evaluation and people who understand the users.

I would also resist making large model training the default ambition for every startup. Founders should start with the task and compare the available approaches. Training a model requires a specific technical and commercial case. The right investment meets the required quality and supports the business.

For infrastructure providers, this means helping customers reach production. Documentation, technical support, dependable billing and engineers who resolve problems all deserve investment alongside the hardware.

For governments, it means evaluating proposals through public value. I would look for additional electricity capacity where needed, skilled jobs with progression, research access and viable opportunities for domestic suppliers.

For founders, it means asking for evidence before committing. Test the service, compare the complete bill and understand the exit terms. Establish whether your customer receives a better result.

I am optimistic about Africa’s AI infrastructure because these investments offer businesses more choices. My confidence will grow with evidence of reliable services, sustainable prices and products people continue using after the launch.

I will judge success by the African businesses able to build, operate and grow on this infrastructure, with control over their work and enough margin to keep going.