A handful of Global South companies now run their own AI compute instead of renting it abroad - though the biggest builders are corporate-backed rather than startups. In Africa, Cassava Technologies (Econet group) is installing what Nvidia calls the continent's first AI factory: 3,000 Nvidia GPUs in South Africa, scaling to 12,000 across five countries. In India, Yotta (Hiranandani group) runs Shakti Cloud, more than 16,000 Nvidia H100 GPUs. Genuine startups are further along on the models than on the metal underneath them.
Building compute is not building models. Owning GPUs and data centres is a different act from building the models that run on them. Cassava and Yotta own the metal; startups like Lelapa AI (a five-language African model) and Sarvam (India's sovereign LLM) build the models. Treating them as one list is the most common error in answering this question.
Owning the build-out vs owning the scale. The two compute leaders are not doing the same thing. Cassava is building physical AI-factory capacity from the ground up across five countries - owning the build-out. Yotta is already operating the single largest deployment - owning the scale. Both keep the compute on home soil, and that is the point: rent foreign cloud and your models, data, and roadmap sit in another company's jurisdiction. Owning the GPUs costs far more upfront, but that capex-versus-rent choice shapes sovereignty more than any GPU count.
A third pole in the Gulf. Abu Dhabi's TII builds and open-sources the Falcon models and runs institutional-scale compute; India's Krutrim is standing up a domestic Nvidia GB200 cluster. Expect the map to keep filling in.
| Company | Base | Owns its compute | What it builds | Source |
|---|---|---|---|---|
| Cassava Technologies | South Africa | Confirmed - building the AI factory | GPU capacity (build-out) | DataCenterDynamics |
| Yotta | India | Confirmed - operating at scale | Largest GPU cloud in India | Nvidia |
| TII / Falcon | UAE | Confirmed - institutional | Open models + compute | TII |
| Lelapa AI | South Africa | No - runs on others' hardware | African-language model | Lelapa |
| Sarvam | India | No - runs on others' hardware | Sovereign LLM | Sarvam |
Method. "Builds their own AI compute" here means owning or operating GPU clusters and data-centre capacity in-region, not renting foreign cloud. Companies that build models but run them on others' hardware are listed separately and marked as such. Every row links to a primary source, and superlatives are attributed to their claimants rather than asserted.