As big tech firms in the U.S. race to spend hundreds of billions of dollars on artificial intelligence, away from Silicon Valley, startups and researchers who cannot access the most advanced chips are adopting a more frugal approach — building smaller AI models on open-weight systems with fewer tokens — to meet the needs of populations with fewer resources.
Generative AI adoption is rising quickly worldwide, but there is a wide gap: Adoption in wealthier countries grew nearly twice as fast as in low- and middle-income countries last year, according to data from Microsoft Research. The gap is likely to widen as computing power for building advanced AI systems is heavily concentrated, with U.S. and Chinese companies operating more than 90% of AI data centers that businesses and institutions rely on, according to researchers at Oxford University. Africa and South America have almost no AI computing hubs.
“The current trajectory of AI development is unsustainable economically, environmentally, and socially. Model sizes have exploded, leading to significant energy and water consumption, and yet billions of people remain excluded from AI’s benefits,“ Arjuna Sathiaseelan, founder of the Saving Voices Project nonprofit, and chief technology officer of the Frugal AI Hub at Cambridge University, told Rest of World. “Frugal AI addresses these failures.”
By design, the systems use less compute, less memory, and less energy, which directly translates into a smaller carbon footprint.”Arjuna Sathiaseelan, founder of the Saving Voices Project nonprofit, and chief technology officer of the Frugal AI Hub at Cambridge University
The Saving Voices Project recently built a speech AI system for the Indigenous Soliga tribe in southern India. As younger members migrated to the cities for jobs, elders in the community feared losing their language. With a small number of speakers, no written script, and no internet access, commercial speech technology was not an option. The Saving Voices Project, along with the Indian Institute of Information Technology, Dharwad, custom-built cheap text-to-speech AI models that run on low-powered devices, and can operate offline for long periods.
The model is replicable for Indigenous language preservation globally, Sathiaseelan said.
“With just five hours of voice data, we were able to build a voice model for the Soliga by prioritizing community ownership, and with frugal, deployable technology,” he said.
Unlike the compute-heavy AI models developed by Silicon Valley, the smaller models being built in India, Indonesia, and elsewhere can run on low-end devices and low-bandwidth networks, and be deployed in sectors such as agriculture, health-care, and education. The models are not only cost-efficient, they also have a lower impact on the environment, Sathiaseelan said.
“This is perhaps the most important dimension of frugal AI,” he said. “It is about building leaner, more efficient systems from the ground up. By design, the systems use less compute, less memory, and less energy, which directly translates into a smaller carbon footprint.”
The launch of DeepSeek in China last year energized advocates of frugal AI. China is developing its own AI cloud and semiconductor supply chain, and its open-source models have quickly become the foundation for developers and builders worldwide. Countries including India, Mexico, and Malaysia also aim to reduce their reliance on expensive chip imports.
In India, Microsoft, Google, and Amazon have all announced multibillion-dollar AI investments to tap the country’s vast data resources and market. But smaller open-source models trained on specific data for specific uses can be almost as effective as massive LLMs trained on general data, Indian tech entrepreneur Nandan Nilekani told Rest of World in an earlier interview.
Indian startups including voice AI company Sarvam and Adalat AI, which provides legal services, use a frugal approach. The Soliga voice models use Raspberry Pi hardware that runs on the Linux open-source operating system. The voice data never left community devices, “something that simply isn’t possible with closed cloud systems,” Sathiaseelan said.
Open-weight models are more amenable to frugal AI as they eliminate proprietary API margins, can run on any infrastructure, and enable data sovereignty, “which matters enormously in non-Western contexts,” he said.
Even Western startups can benefit from a more cost-efficient approach, Lingjiao Chen, a researcher in the AI Frontiers group at Microsoft Research, who wrote a paper on an algorithmic framework he calls FrugalGPT, told Rest of World.
With a growing number of LLMs to choose from, users often don’t know how to pick the one that suits their budget and accuracy goals. FrugalGPT can automate model selection, reducing cost significantly while improving accuracy, he said.
“Given the huge financial cost, energy consumption, and environmental impacts of LLMs, a major issue is how sustainable they can be in the long-term,” Chen said. “There is also a risk that AI models become unaffordable to more users due to their high cost. FrugalGPT and other frugal AI tech are thus increasingly important.”
How concentrated will compute become — will access to major AI models become similar to access to oil?”Sebastián Uchitel, a professor in the department of computing at the University of Buenos Aires
With supply chain constraints and crises such as the Iran war creating other bottlenecks, frugal AI is also vital for a country’s tech sovereignty as a growing number of economic activities come to rely on foreign AI models, Sebastián Uchitel, a professor in the department of computing at the University of Buenos Aires, told Rest of World.
“How concentrated will compute become — will access to major AI models become similar to access to oil?”
Still, frugal AI models can be restrained by data scarcity, compute limitations, and funding gaps that prevent scaling. Their success also depends on the domestic AI infrastructure, including access to efficient, cost-effective graphics processing units, and data centers.
“The performance trade-offs are real, though the art is in identifying which tasks genuinely require frontier capability — and the answer is far fewer than people assume,” Sathiaseelan said. The word error rate in the Soliga models is slightly high, but “what we gained is complete data sovereignty, offline deployment on sub-$50 hardware, and a governance structure that elders and community leaders actually trust,” he said.
Demand for this approach is growing. The Frugal AI Hub is setting up a lab in the Indian state of Andhra Pradesh, as well as across the country with universities and foundations. It is also in talks with officials in Kenya and Nigeria.
The Saving Voices Project aims to reach nearly 500 million Indigenous people in 90 countries. For the communities, their knowledge and culture are intricately tied to the language, and these need preserving, Sarabani Banerjee Belur, co-founder of the project and an assistant professor at IIIT Dharwad, told Rest of World.
It needs to be done without the exploitative data extraction practices of big tech companies, she said.
“For communities with a long history of having their land and cultural materials extracted by outsiders, that is a deal breaker,” she said. “Our aim is creating community data sovereignty and a scalable blueprint for language restoration and digital agency by bridging the gap between cutting-edge AI and Indigenous oral traditions.”
