SA must learn from China’s digital white elephants and build a sustainable AI tech ecosystem

Recent reports show that China’s aggressive AI data centre construction is facing up to 80% underutilisation in some regions. This cautionary tale of overbuilding raises questions about how Africa might avoid similar pitfalls while still embracing AI’s transformative potential.

Hardy Pemhiwa, the president and group CEO of Cassava Technologies, reckons the AI market opportunity for Africa is R95-billion, with the bulk of that accounted for in South Africa.

“The primary drivers pushing this demand are the need to preserve local culture and language for LLMs [large language models], regulatory compliance, intellectual property protection and data sovereignty,” he told Daily Maverick in an exclusive interview.

“We are planning to build large language models for major African languages, which will help democratise the use of AI applications across Africa.”

A sound academic foundation

Before the onset of the age of private capital, all research for major project builds like GPS, the internet and other major human advances was paid for by the government. Now private industry is moving faster than universities.

At the University of Cape Town (UCT), Professor Jonathan Shock, associate professor in the department of applied maths and director of UCT’s AI initiative, is leading efforts to establish the university as an AI research hub focused on pressing African challenges.

“There’s amazing research going on in areas like health and wellbeing, climate of the environment, poverty and inequality, along with AI safety,” explains Shock. “We want to accelerate this research hugely by bringing in the right people and proper capacitation.”

However, this acceleration faces infrastructure challenges. Many UCT graduate students rely on temporary resources such as Google’s free TPU access program — a situation that highlights the urgent need for sustainable, local computing power.

“As we build, we’re not going to be able to keep using this sort of free access, and having our own compute [processing power, computation or computing resources] is going to be absolutely vital,” he stresses. “Compute sovereignty is going to become a real issue.”

Enough compute to go around

UCT’s approach centres on “knowledge translation”, focusing on how research outputs can become practical applications, including clinical use, spin-off companies and supporting startups.

The priority is ensuring researchers and entrepreneurs are “not short of compute” while pursuing solutions to uniquely African challenges.

“We don’t know what’s going to happen in terms of the US computes,” adds Shock, referring to potential export restrictions.

“It’s very clear that we do need to have our own compute and that may be compute which is on premises which is shared with the rest of the South African and even further African AI communities.”

Enterprise perspective

While infrastructure dominates headlines, Sashen Naidu, the global VP for customer experience services at NTT, believes that successful AI implementation requires more than raw computing power.

“The challenge for enterprises isn’t solely about accessing AI models or infrastructure but also about being able to sustain the benefits,” explains Naidu. This requires “a strategy to manage and operate and sustain a bot” similar to managing human employees, including training, quality management and performance metrics.

The economic benefits are already materialising. “Agentic AI [smaller, task-driven AI that can perform specific tasks like schedule a payment] is coming in around $2 to $3 [per transaction],” he points out, compared to traditional costs of “$10 to $12 per transaction”.

This represents “a significant decrease in costs but not just decreasing costs — we found that it’s increasing metrics like customer satisfaction”.

The impact on processing times is equally dramatic. Tasks that once took “hours ... now we’re talking about seconds and minutes”, says Naidu, highlighting how AI is transforming business operations beyond cost savings.

Infrastructure as a service 

Robin Fisher, the head of Europe, Middle East and Africa growth markets at Salesforce, says that AI customers are “moving from a subscription model to consumption model” with AI adoption.

He explains this evolution as a natural progression in cloud computing. “So now what we’re seeing is exactly that reason we’re now going to consumption-based models, which a lot of the hyperscalers have already been doing.”

The Agentforce — Salesforce’s agentic AI solution — pricing model is based on an idea that customers will pay for successful interactions or “resolutions” achieved through the platform.

Fisher says the aim is for customers to derive more value than the cost per resolution, turning high adoption rates into a positive indicator of Agentforce’s effectiveness and value proposition.

This “pay as you go” approach, also applicable to Data Cloud, is designed to cater to high-value interactions that ultimately benefit a business’s bottom line.

Building AI for Africa 

Cassava Technologies offers perhaps the most optimistic view, announcing that “Cassava is investing significantly in Nvidia supercomputing for its AI factory in South Africa.

“Africa’s AI future is owned by Africans who will use their datasets, models, languages and voices to create solutions that leave no African behind,” predicts Pemhiwa.

To meet these needs, Cassava is deploying “Nvidia HGX H200 supercomputers” with cutting-edge capabilities. “These are one of the first GPUs [graphics processing units] with HBM3e [high bandwidth memory designed for demanding applications like AI training, high-performance computing and supercomputing, where massive amounts of data need to be processed quickly and efficiently], which makes them faster, and they have more significant memory to fuel the acceleration of generative AI and LLMs while advancing scientific computing for HPC [high-performance computing] workloads,” explains Pemhiwa.

Cassava’s approach includes accessibility measures for smaller players through “tiered pricing models enabling customers to either reserve GPUs or use them on demand” and collaboration with educational institutions to “foster local AI development and innovation”.

What this means for you 

Africa benefits from the cost-effectiveness of new AI hardware. Efficient, open-source models like DeepSeek shift the emphasis from building advanced large language models to effectively using them.

Instead of massive infrastructure investments, Africa can focus on optimised infrastructure and practical applications. South Africa’s strategy, particularly in local languages and cultures, is promising. Pemhiwa’s plan to develop LLMs for African languages aims to democratise AI, creating unique value. Continued alignment between infrastructure investment and demand will be crucial to avoid pitfalls seen in other regions. DM