Three years ago, when I moved to Singapore to focus on building a business, I assumed the most interesting AI story would keep unfolding in San Francisco and Shenzhen. The money was there. The chips were there. The talent pipeline flowed through a handful of zip codes. I was wrong about where the story was heading, even if I wasn’t wrong about where the capital was pooling. The most interesting AI work I’ve encountered recently isn’t happening in gleaming data centers cooled by municipal water supplies. It’s happening on Raspberry Pi boards in southern India, running speech models built from five hours of recorded voice data, preserving a language that has no written script.
The conventional wisdom says AI is a scale game. Bigger models, bigger budgets, bigger compute clusters. Major tech companies are collectively spending hundreds of billions of dollars on infrastructure. The assumption, repeated so often it feels like natural law, is that the nations and institutions that can’t match this spending will simply become consumers of AI rather than producers of it. They’ll rent access to someone else’s intelligence. They’ll pipe their data into someone else’s cloud. They’ll pay API fees denominated in someone else’s currency.
But that assumption misses something fundamental about how technology actually diffuses through societies. And a growing movement of researchers, startups, and governments is proving it wrong in real time.
The concentration problem nobody wants to name
The numbers are stark. Research suggests that U.S. and Chinese companies operate the vast majority of the AI data centers that businesses and institutions worldwide depend on. Africa and South America have almost no AI computing hubs. Generative AI adoption in wealthier countries grew substantially faster than in low- and middle-income countries last year.
These are not statistics about a temporary lag. They describe a structural arrangement. Sebastián Uchitel, a professor in computing at the University of Buenos Aires, has raised concerns about whether access to major AI models could become as concentrated as access to oil. If advanced AI becomes essential infrastructure for healthcare, agriculture, education, legal systems, and government services, then a world where the vast majority of AI compute sits in two countries looks less like a market structure and more like a dependency relationship. I wrote about a version of this dynamic in Nigeria’s creator economy, where African talent generates value that platform economics routes back to Silicon Valley. AI threatens to replicate that extraction at a deeper level, not just capturing economic value, but capturing the data, language, and decision-making infrastructure of entire populations.
The standard response from Big Tech is investment. Microsoft announced $17.5 billion in India. Google committed $1.5 billion. Amazon pledged $35 billion. These are real commitments. But the model is still one where a handful of foreign corporations build the infrastructure, set the terms, and control the stack. Countries become markets, not makers. And that distinction matters, because the entity that controls the AI stack controls the terms on which everyone else participates in the economy it enables.
The frugal alternative
Against this backdrop, a different approach is taking shape. Researchers and builders in India, Indonesia, Argentina, Kenya, and elsewhere are constructing AI systems designed from the ground up to work within severe resource constraints. Smaller models. Open-weight architectures. Offline capability. Hardware that costs less than a restaurant meal in Manhattan.
Arjuna Sathiaseelan, founder of the Saving Voices Project and associated with the Frugal AI Hub at Cambridge University, has argued that current AI development is unsustainable and excludes billions of people, advocating for frugal AI approaches that address these challenges.
The word “frugal” matters here. It’s not a euphemism for inferior. It’s a design philosophy rooted in a specific thesis about technological sovereignty: the communities and nations that control their own AI infrastructure, however modest, retain more power than those that rent access to superior systems they can neither inspect, modify, nor govern. In my recent piece on the $50 AI revolution, I explored how smaller models built for sovereignty may matter more than the trillion-dollar arms race. The frugal AI movement is the practical, on-the-ground expression of that argument.
The core insight is deceptively simple: most real-world AI tasks don’t require frontier-capability models. A farmer in Karnataka checking crop disease patterns doesn’t need GPT-5. A community health worker in rural Kenya running diagnostic screening doesn’t need access to a model trained on the entire internet. What these users need are models trained on specific data for specific purposes, running on hardware they can afford and maintain, storing data they control.
Indian tech entrepreneur Nandan Nilekani has made this point directly: smaller open-source models trained on specific data for specific uses can be almost as effective as massive LLMs trained on general data. The Open LLM Leaderboard on Hugging Face has made comparing these models increasingly accessible, and the performance gaps on task-specific benchmarks are narrower than most people assume.
Five hours of voice data and a $50 computer
The Saving Voices Project offers the most vivid example of what frugal AI looks like in practice — and why the sovereignty question isn’t abstract. The project built a speech AI system for the Soliga, an Indigenous tribe in southern India whose language has no written script. Younger Soliga members have migrated to cities for employment. Elders feared the language would die with them. Commercial speech technology was useless: no internet access in the community, no existing corpus of Soliga text, no way to feed data into a cloud platform.
Working with the Indian Institute of Information Technology in Dharwad, the team custom-built text-to-speech AI models using just five hours of recorded voice data. The models run on Raspberry Pi hardware costing under $50, operating on the Linux open-source operating system, capable of functioning offline for extended periods. The voice data never left community devices.
Sathiaseelan emphasized that the approach provides data sovereignty, offline deployment on inexpensive hardware, and governance structures trusted by community leaders.
That last phrase is the one that sticks with me. A governance structure that elders actually trust. When I ran Ideapod, a platform that once reached hundreds of thousands of users before closing, one of the hardest lessons I learned was that grand technological visions mean nothing if the people you’re building for don’t trust the system. Trust isn’t a feature you bolt on. It’s a function of who controls the data, who sets the terms, and who can walk away.
For Indigenous communities with generational experience of extraction (of land, of resources, of cultural materials), the question of who controls AI data isn’t abstract. It’s existential. And it illustrates a crucial point about why technological sovereignty can’t be separated from cultural survival. When a language dies because no technology was built to preserve it, that’s not a market failure. It’s a consequence of an AI development model that only builds what scales profitably. Sarabani Banerjee Belur, co-founder of the Saving Voices Project and assistant professor at IIIT Dharwad, emphasizes that communities with histories of extraction need data sovereignty, with the project aiming to create scalable models for language restoration and digital agency.
The Saving Voices Project aims to reach Indigenous communities across dozens of countries. That’s not a niche use case. That’s a significant portion of humanity — and a direct challenge to the assumption that AI’s value is determined solely by the markets it serves.
Why data sovereignty is the real game
The Soliga example makes a specific case visible: technological sovereignty isn’t just about who builds the hardware or trains the model. It’s about who decides what gets built, whose knowledge gets encoded, and whose interests the system ultimately serves. When a community in southern India builds its own speech model on local hardware, stores all voice data on community devices, and governs the system through its own leadership structures, something happens that no amount of Big Tech investment can replicate. The community retains control — not just over a technology product, but over the terms of its own participation in the digital economy.
Sathiaseelan described why open-weight models are essential to this: they eliminate proprietary API margins, can run on any infrastructure, and enable data sovereignty, which matters enormously in non-Western contexts.
This matters because the AI supply chain is becoming a geopolitical pressure point. Supply chain constraints and conflicts like the Iran war are creating bottlenecks in chip supply. Countries that depend entirely on foreign AI infrastructure are exposed to disruptions they cannot control. For Argentina, for Kenya, for Indonesia, the ability to run locally built models on locally maintained hardware isn’t just an efficiency play. It’s a sovereignty strategy.
Indian startups are already operationalizing this. Voice AI company Sarvam and legal services provider Adalat AI both use frugal approaches, building specific-purpose models that work within Indian data and compute constraints rather than attempting to compete with frontier labs on their own terms. The Frugal AI Hub at Cambridge is expanding too, setting up a lab in the Indian state of Andhra Pradesh and entering discussions with officials in Kenya and Nigeria.
Silicon Canals has reported on Thales’ work on sovereign AI in a military and civil context for France, which reflects a different version of the same underlying anxiety: even wealthy nations are questioning whether critical AI infrastructure should depend on foreign providers. The frugal AI movement takes that concern and applies it to contexts where the resources to build alternatives are dramatically more constrained.
The DeepSeek effect
The launch of DeepSeek in China last year changed the conversation. Here was a model that achieved strong performance without matching the compute budgets of its American counterparts. China is building its own AI cloud and semiconductor supply chain, and its open-source models have become foundational for developers worldwide. Countries including India, Mexico, and Malaysia have been energized by the proof that frontier-adjacent capability doesn’t require frontier-level spending.
This connects to a broader pattern I keep watching from Singapore: the assumption that technological progress runs along a single track, defined by whoever spends the most, keeps breaking down. DeepSeek didn’t just compete on price. It suggested that the relationship between compute expenditure and model capability isn’t as linear as the dominant narrative implies.
Lingjiao Chen, a researcher in the AI Frontiers group at Microsoft Research, has been working on this problem from a different angle. His algorithmic framework, FrugalGPT, automates model selection to reduce cost significantly while improving accuracy. With a growing number of LLMs available, most users don’t know how to pick the model that suits their budget and accuracy requirements. FrugalGPT addresses that.
Chen framed the stakes clearly: “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. 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.”
What’s notable is that this isn’t just a concern for developing nations. Even Western startups can benefit from more cost-efficient approaches. The frugal philosophy applies up and down the economic spectrum. But its implications are most dramatic for the countries and communities that Big AI’s current trajectory would otherwise leave behind.
The limits are real, but the limits aren’t where you think
Frugal AI is not a magic solution. The constraints are genuine. Data scarcity, compute limitations, and funding gaps prevent scaling. Success depends on domestic AI infrastructure, including access to efficient GPUs and local data centers. A Soliga speech model with a slightly high word error rate is a real trade-off.
But Sathiaseelan has noted that while performance trade-offs exist, identifying which tasks truly require cutting-edge AI capabilities reveals that far fewer applications need them than commonly assumed.
This is the key insight. The AI industry has a massive incentive to convince every potential customer that they need the biggest, most expensive model available. The business model of OpenAI, Google, and Anthropic depends on enterprise clients believing that nothing less than a frontier model will do. But for the vast majority of practical applications (crop disease detection, basic legal research, language preservation, community health screening), a well-trained small model running on inexpensive hardware may be not only sufficient but superior, because it’s available, affordable, and controlled by the people who use it.
I learned a version of this lesson running Ideapod. For years, I chased the grand vision: build the platform, attract the users, scale to millions. When I finally shifted focus to doing small, specific things that actually helped the people using the platform, the work became more meaningful even as the business model proved unsustainable in an AI-dominated content world. The parallel isn’t perfect, but the principle transfers. The most valuable technology is often the technology that fits the actual problem, not the technology that impresses investors.
A different map of AI’s future
If you only follow AI through the lens of earnings calls and chip wars, the future looks like a two-player game between the U.S. and China, with everyone else renting access. But the map looks different from ground level.
In Andhra Pradesh, the Frugal AI Hub is building a new lab. In Buenos Aires, computer science professors are asking whether compute will become the new oil. In Kenya and Nigeria, officials are in discussions about how to build locally. In southern India, Soliga elders are preserving their language on devices that cost less than a pair of sneakers, using voice data that never touches a cloud server.
None of these efforts will produce the next model that tops the benchmarks. That isn’t the point. The point is that useful, deployable, sovereign AI is being built by people who were supposed to be passive consumers of someone else’s technology. And they’re building it precisely because the trillion-dollar approach wasn’t built for them.
The environmental dimension adds another layer. Sathiaseelan described the sustainability argument as among the most important dimensions of frugal AI. Smaller systems use less compute, less memory, and less energy, which translates directly into a smaller carbon footprint. At a moment when frontier AI labs are securing power contracts that rival small cities, the ecological case for efficiency-first design is hard to dismiss.
What changes if this works
I think there’s a broader lesson here about how power actually works in technology. The narrative of inevitable concentration (that whoever builds the biggest model wins the future) serves the interests of the people building the biggest models. It’s not wrong as a description of one dynamic. But it’s incomplete as a map of where value gets created and where agency resides.
So what specifically changes if frugal AI succeeds at scale?
First, the dependency pipeline breaks. If dozens of countries can build and deploy task-specific AI on local infrastructure, the leverage that comes from controlling cloud access and API pricing diminishes. Big Tech doesn’t lose its market, but it loses its monopoly on the terms of participation. Countries that build domestic AI capacity gain negotiating power — the ability to walk away from a deal they don’t like, which is the most fundamental form of sovereignty there is.
Second, the data extraction model reverses. Right now, data flows from the Global South into training sets controlled by Northern companies, who sell the resulting capabilities back at a markup. Frugal AI, built on local data stored on local devices, keeps that value where it originates. Over time, this creates a different kind of AI ecosystem — one where a Kenyan agricultural model trained on Kenyan soil data is owned by Kenyan institutions, not licensed from a California startup that scraped the data from a development agency’s open dataset.
Third, the cultural cost of AI adoption drops. When the only AI tools available are built on English-language, Western-context training data, every community that adopts them accepts a quiet erasure of its own knowledge systems. Frugal AI built for Soliga, for Quechua, for Yoruba doesn’t just preserve languages. It preserves the worldviews encoded in those languages. That’s not sentimental. It’s a form of intellectual sovereignty that determines whether AI amplifies human diversity or flattens it.
None of this is guaranteed. The funding gaps are real. The technical limitations are real. And the gravitational pull of Big Tech’s ecosystem is enormous — it’s far easier to plug into someone else’s API than to build your own model from five hours of voice data. But the frugal AI movement suggests that the most consequential AI work of the next decade might not come from the labs with the most GPUs. It might come from the communities with the clearest understanding of what they actually need.
Concentration breeds dependency. Frugality, when it’s intentional and well-designed, breeds something closer to self-determination. The choice between those two futures isn’t purely technical. It’s political. And the people making that choice most urgently aren’t in San Francisco or Singapore. They’re in the communities that the global AI economy was never designed to serve, who are deciding to build anyway — and in doing so, are quietly redrawing the map of who gets to shape what intelligence means in the twenty-first century.
Feature image by Zeal Creative Studios on Pexels
