The recent G20 Summit in Johannesburg turned its attention to the rising cost of artificial intelligence and the broader challenge of building a trusted, inclusive digital economy.
In response, says South African born AI company Lelapa AI, it is introducing an efficiency-first approach, shaped by the realities of limited computing power and complex multilingual environments.
AI systems are becoming increasingly expensive to build and maintain, with large-scale models demanding vast computing power, heavy energy use, extensive datasets, and substantial capital investment. These pressures widen the divide between countries with abundant digital infrastructure and those still working to expand affordable access, strengthen data governance, and build resilient innovation ecosystems.
The G20’s digital agenda highlights interconnected challenges: managing cross-border data flows securely, ensuring trustworthy AI development, supporting MSMEs in the digital economy, strengthening national capabilities, and closing the usage and coverage gaps that limit digital participation.
Lelapa AI’s approach, it says, offers a technically credible pathway that responds directly to these concerns by reducing infrastructure demands, lowering costs, and enabling broader language coverage without compromising performance. Big tech currently absorbs close to 90% of model serving costs, a burden that is neither sustainable nor scalable for the world.
“The G20 places sustainability and equity at the centre of its digital agenda. Efficiency is essential to both,” says Professor Vukosi Marivate, co-founder of Lelapa AI and knowledge partner to the G20 AI Task Force. “Our contribution at Lelapa AI highlights that low-resource approaches are necessary for global AI that can serve diverse economies and linguistic realities.”
Lelapa AI provided the following information about its efficiency-first approach:
Lelapa AI builds language technologies designed to operate in complex, multilingual, and data-scarce environments. By designing for constraint from the start, the company develops models that are affordable, high performing, and scalable across regions. Their approach benefits the Global South and provides a blueprint for high-resource markets that face rising inference costs and increasing pressure to reduce environmental impact.
Lelapa AI’s work includes:
- Vulavula, a real-world transcription and translation engine built to perform accurately in multilingual, code-switched environments. Its efficiency-first design lowers serving costs, making high-quality language intelligence more economical and accessible for governments, enterprises, and public institutions.
- InkubaLM, Africa’s first multilingual small language model engineered for low-resource and code-switching environments. Through the Buzuzu Mavi Challenge, the model was significantly shrunk by 75% without compromising performance, proving that efficiency and capability can coexist.
- The Esethu Framework, a pioneering sustainable data governance model that centres communities in how low-resource language datasets are built and maintained. Its license ensures that foreign users of African language data contribute to future dataset development, creating a self-sustaining ecosystem and economically sound path that supports underserved languages. This approach offers a blueprint that can be adapted and applied globally.
- ViXSD, the first dataset produced under the Esethu Framework. This open-source isiXhosa ASR dataset includes ten hours of high-quality speech across dialects, ages, and regions. Its community-led development ensures authenticity, while the Esethu License secures ongoing reinvestment into new datasets and the advancement of equity outcomes.
- Extensive Research and Applied Expertise. Lelapa AI integrates rigorous academic research with real-world product development, a rare combination in the industry where companies often specialise in one or the other. The company is a world leader in efficient model design, sustainable dataset creation, and scalable language AI.
Together, these innovations demonstrate that scalability and performance can be achieved without overbuilt systems or excessive compute.
Said Pelonomi Moiloa, CEO of Lelapa AI: “The future of AI depends on efficient design. Scarcity encourages sharper thinking, and our work at Lelapa AI shows that scalable and sustainable AI does not require boundless compute or heavy infrastructure. It starts with smarter foundations that serve real people and real contexts.”
Lelapa AI’s work sits at the intersection of real-life challenges faced by institutions that must do more with limited resources. In government offices, clinics, classrooms, and community centres, language remains one of the most significant barriers to access.
By designing AI systems that thrive in low-data and low-compute environments, Lelapa AI supports public-facing sectors that need reliable tools without the heavy cost of infrastructure. This approach strengthens digital public services, improves multilingual healthcare communication, and gives schools the language-aware tools they need to reach learners across diverse linguistic contexts.
Across the wider economy, resource-efficient AI creates room for participation that high-cost systems often make impossible. MSMEs can engage customers across languages, frontline workers can document interactions accurately, and public service teams can communicate clearly with communities that span dozens of linguistic realities.
A global contribution from Africa
Lelapa AI positions Africa’s resource-conscious innovation as a model for the world. With affordability, efficiency, and linguistic diversity at the centre, Lelapa’s work gives stakeholders practical pathways to build affordable, inclusive, and sustainable digital ecosystems. This perspective signals a future where AI strengthens societies without placing undue strain on the systems meant to serve them.