AI & African Enterprises Newsletter: February 2025
Originally published February 24, 2025
🤿A Deep Dive into Deep Seek
In the aftermath of the feverish activity that followed the announcement of DeepSeek's R1 model delivering superior performance at a lower cost, relative to industry-leading large language models, we're taking a closer look at what DeepSeek means for African AI.
Background
DeepSeek is an offshoot of quantitative hedge fund, High-Flyer Quantitative Investment Management, founded by Liang Wenfeng. Wenfeng acquired 10,000 Nvidia A100 GPUs in 2023 before the first set of chip export restrictions imposed by the Biden Administration came into effect, providing sufficient compute to pursue large language model research.
The Innovation
DeepSeek introduces an innovation to the architecture used to train Large Language Models, to achieve significant cost reductions and superior performance. DeepSeek claim their standard-setting model was trained using $5.6M, a fraction of the cost of US counterparts.
DeepSeek seems to be challenging the notion that innovation comes from the US, and "China only copies"; some in the US questioned or downplayed DeepSeek's accomplishments. DeepSeek open-sourced the results of their research, allowing the public to examine and benefit from their efforts.
Implications for Africa
DeepSeek fits quite neatly into China's global AI policy, which advocates for broad and inclusive AI access. Cost is a key barrier to AI research and product development, especially in Africa where AI investment is at a fraction of global counterparts. Wenfeng has demonstrated how the combination of capital and well-organized use of local resources can give rise to significant innovations. Wenfeng completed all of his education in China, an only hired local engineers as a priority.
In the final week of Biden's presidency, the US president announced an executive order further restricting US chip exports to all but 18 US allies, none of which are African. The executive order requires a license for GPU orders comprising more than 1700 units.
Takeaway
DeepSeek achieved new levels of efficiency in the AI frontier by developing innovative ways to consume less compute. The achievement has demonstrated the scale of what can be achieved through diligent application of capital and local talent. If Africa is to achieve AI sovereignty, efforts will likely need to be made to organize local engineering talent and optimize limited access to US compute resources.