Mustafa Ehsan, head of research firm Convly, has warned that African nations and startups should avoid the costly pursuit of building frontier AI models, urging a shift toward high-value application and economic integration.
Ehsan, who tracks the price and performance of 33 different AI models, argues that the current global competition to build massive large language models (LLMs) is an economic arms race that most African entities cannot win or sustain.
Frontier models, such as those developed by OpenAI and Google, require billions of dollars in compute power, specialized hardware, and massive electricity reserves. For most African economies, pursuing this path risks wasting scarce capital on infrastructure that is rapidly depreciated by faster-moving global players.
The Convly research indicates that the gap between the most expensive frontier models and highly efficient open-source alternatives is closing. This suggests that the value in the AI stack is shifting from the base model to the implementation layer.
Ehsan suggests that Africa’s competitive advantage lies not in creating a general-purpose model from scratch, but in building vertical AI solutions that solve specific regional challenges in agriculture, finance, and healthcare.
The Economic Barrier to Frontier Model Development
The cost of training a frontier model involves a combination of high-end GPUs, such as the Nvidia H100, and immense datasets. These requirements create a “compute divide” that makes it difficult for smaller economies to compete on a level playing field.
Industry experts argue that the “national champion” model—where a government funds a single, massive AI project to ensure sovereignty—often fails to deliver practical economic returns compared to supporting a diverse ecosystem of startups.
Instead, the recommended strategy is to leverage existing frontier models via APIs or open-source frameworks like Hugging Face, and then apply a process called fine-tuning. This allows developers to adapt a powerful existing model to a specific language, dialect, or industry without the cost of initial training.
Another viable path is Retrieval-Augmented Generation (RAG). This technique allows an AI to pull information from a specific, trusted local database, ensuring the output is accurate and culturally relevant without needing to rebuild the model’s core architecture.
This approach mirrors Africa’s success in mobile finance. The continent did not build the underlying mobile hardware or operating systems, but it leapfrogged traditional banking by building innovative applications like M-Pesa on top of existing telecommunications infrastructure.
The focus, according to Ehsan, should be on data sovereignty. While Africa may not need to build the engine, it must own the fuel—the high-quality, curated African datasets that make AI tools actually work for local users.
Investment in data curation and local language datasets is seen as a more sustainable way to ensure that AI does not simply reflect Western biases. This involves digitising local archives and partnering with community linguistic experts.
Venture capital interest in the region is already beginning to shift. Investors are increasingly scrutinising “wrapper” startups that simply repackage GPT-4, instead seeking companies that possess unique proprietary data and deep integration into local business workflows.
The next phase for African AI development will likely depend on the availability of affordable cloud compute and the ability of governments to create regulations that protect local data without stifling innovation. Regulatory frameworks from the African Union will be critical in determining how cross-border data flows are managed to support this application-first strategy.
Explore more Tech stories and analysis from Business Elites Africa.
