The U.S. and China control more than 90% of the world's AI data centers. Generative AI adoption in wealthy countries grew nearly twice as fast as in low- and middle-income countries last year, according to Microsoft Research. The gap is widening, but something interesting is happening on the other side of it.
Researchers and entrepreneurs across India, Indonesia, Kenya, Argentina, Brazil, and other countries are building what's being called "frugal AI" - models designed to do specific jobs well on minimal hardware, instead of trying to do everything on billion-dollar infrastructure.
A Language Model on a $50 Raspberry Pi
The Saving Voices Project, led by Arjuna Sathiaseelan (CTO of the Frugal AI Hub at Cambridge University) and Sarabani Banerjee Belur of IIIT Dharwad, built a text-to-speech AI for the Indigenous Soliga language of southern India. It runs on a Raspberry Pi - a computer that costs under $50 - using just five hours of recorded voice data. No cloud connection required. No GPU cluster. No API fees.
That's the frugal AI thesis in miniature: instead of feeding every problem into a massive general-purpose model, build something small and specific that runs where it's needed.
Why Smaller Can Be Smarter
China's DeepSeek releasing competitive open-source models has energized this movement. In India, companies like Sarvam (voice AI) and Adalat AI (legal services) are building products around smaller, locally-run models. Microsoft Research's FrugalGPT framework automates the process of picking the cheapest model that can handle a given task - routing easy queries to tiny models and only calling expensive ones when necessary.
Sebastián Uchitel, a computing professor at the University of Buenos Aires, represents the growing academic push to develop AI approaches that don't require access to expensive chips or massive data centers.
The advantages go beyond cost. Frugal models can run offline, which matters in regions with unreliable internet. They keep data local, which matters for sovereignty. And they consume far less energy and water than large language models (training a single large model can use as much water as a small town consumes in a day).
The Sustainability Argument
"The current trajectory is unsustainable economically, environmentally, and socially," Sathiaseelan told Rest of World, referring to the trend of ever-larger AI models.
He has a point. The race to build bigger models assumes that bigger is always better. But a Soliga language model doesn't need to write Python code or analyze financial reports. It needs to preserve a language spoken by a few thousand people, and it does that on hardware you could buy with pocket change.
For the AI tools market, this trend is worth watching closely. The assumption that everyone will eventually subscribe to the same handful of cloud AI providers looks less certain when viable alternatives can run on a device the size of a credit card. The frugal AI movement isn't competing with GPT-5 or Claude - it's solving problems those models were never designed for, at price points they can't touch.