A South Africa-founded startup is challenging the AI industry’s energy-intensive growth model, securing $5 million in seed funding to develop tools that compress and restructure AI models so they can run on smaller devices using significantly less power.
The round was led by California-based climate technology investor VoLo Earth Ventures. Refiant AI was founded in 2025 by Viroshan Naicker, Siddharth Gutta and Mathew Haswell, and builds software that reduces the computational requirements of AI systems while maintaining performance, allowing models to run on local infrastructure rather than large data centers.
The company’s headline result illustrates the potential scale of the approach: Refiant compressed a 120 billion parameter AI model — among the most powerful open-source models available — to run on a MacBook Pro with just 12 gigabytes of RAM.
The startup is entering a market under growing pressure. Global data center power consumption is expected to double by 2028, driven largely by the explosive growth of large language models and generative AI workloads. Major technology companies are projected to spend nearly $700 billion in 2026 alone on building infrastructure to sustain AI development.
Joseph Goodman of VoLo Earth Ventures said the fundamental constraint on AI is not demand but energy, and that Refiant’s approach replaces brute-force scaling with a more efficient architecture that lowers energy use while increasing capability.
Refiant’s team includes a former Google Cloud architect, a Cambridge University doctoral researcher and an engineer with NASA experience. The company is in discussions with several multinational technology firms exploring how its approach could reduce AI computing costs while maintaining data and energy sovereignty.
For Africa specifically, Refiant said the continent’s AI ambitions are currently constrained by limited data center infrastructure and dependence on foreign cloud providers, making the case for efficient, locally operable AI models particularly relevant to the region.
