The team at Refiant AI took a 120 billion parameter open-source AI model, the kind that normally needs a data centre’s worth of computing power to run, and got it working on a MacBook Pro with 12 gigabytes of memory. That demonstration is the whole pitch of the company: AI models do not need to be as computationally expensive as the industry has assumed, and proving it is worth real money. Refiant AI has raised $5 million in seed funding led by VoLo Earth Ventures, a California-based climate technology investor, TechMoran reported.
Compression, not a smaller ambition
Refiant, founded in South Africa in 2025 by Viroshan Naicker, Siddharth Gutta and Mathew Haswell, builds software that compresses and restructures AI models to cut their computational requirements while keeping their performance intact, letting them run on local devices with far less power than a data centre needs. The team includes a former Google Cloud architect, a Cambridge University doctoral researcher and an engineer with NASA experience, Tech In Africa reported, a mix that reads like a company built specifically to solve a hard systems problem rather than to chase a trend.
In July, Refiant put the technology to a public test: TechCabal reported the launch of Protea, a large language model named after South Africa’s national flower, built with a 10 million token context window, large enough to process far more data in a single session than most consumer AI tools currently allow.
The problem this is actually aimed at
The investment case is a real infrastructure constraint, not a marketing angle. Global data centre power consumption is projected to double by 2028, and major technology companies are expected to spend close to $700 billion on AI infrastructure in 2026 alone. Every one of those data centres needs power, cooling and, increasingly, a grid connection that is not guaranteed. A technology that lets a meaningful AI model run on ordinary hardware instead of a hyperscale data centre is not a nice-to-have efficiency gain, it is a way around a bottleneck the entire industry is racing toward.
For South Africa specifically, that framing lands differently than it would in a market with abundant, cheap power. A country that spent years living with load shedding, and where reliable infrastructure is never something to take for granted, has an obvious stake in AI technology that does not assume unlimited data centre capacity is always available.
What it means for South African AI, beyond one company
VoLo Earth Ventures backing a South African-founded company is itself notable: a climate-focused Californian fund chose this team over the many AI-efficiency startups competing for the same capital in bigger, better-connected markets. Between Refiant’s $5m and Verascient’s $1.2m pre-seed round in a related AI infrastructure space, this is the strongest run of international AI investment into South African-founded technical teams this year, and a genuine data point against the assumption that serious AI infrastructure only gets built in San Francisco or London.
Protea’s naming choice is not incidental branding. A large language model built by a South African team, capable of holding far more context than most consumer tools currently offer, carrying a name drawn from the country’s own natural heritage, is a small but deliberate statement about where the company sees itself: not a local outpost of someone else’s AI industry, but a team building foundational technology on its own terms, for a global market that happens to also need it. Whether Refiant can keep that positioning as it scales past a seed round, competing against far better-funded rivals doing similar compression work in the US and China, is the real test still ahead of it.
