AI development is often framed as a compute race, dominated by well-funded labs. However, India is proving that constraint can be a design philosophy, leading to AI models that outperform expensive ones in specific contexts.
A significant, overlooked issue is the "tokenization tax" affecting non-English speakers. Large language models break text into tokens. English, dominant in training data, is tokenized efficiently. Indian languages, however, require significantly more tokens per query, potentially increasing costs up to five times for users. This makes AI interactions prohibitively expensive for many.
Startups like Sarvam AI are addressing this by building better tokenizers for Indian languages, reducing per-interaction costs at the source. This enables vital applications like voice-enabled medical triage in rural areas and affordable AI tutoring.
Instead of training foundation models from scratch, India's approach focuses on adapting existing open-source models. Sarvam AI's OpenHathi project enhances large language models with Indian language capabilities. This layered approach, focusing on domain-specific fine-tuning, is more efficient and cost-effective than building massive, general-purpose models from the ground up.
Krutrim, another Indian AI venture, is designing models optimized to run on modest hardware, assuming limited infrastructure. This contrasts with frontier AI development that relies on extensive data centers. Initiatives like AI4Bharat are building lightweight systems for low-end smartphones and unreliable networks, making AI usable for populations with limited connectivity and budget.
This "frugal AI" movement offers a blueprint for resource-constrained nations. The principles of using open-source models, prioritizing local languages, and designing for low-bandwidth environments are transferable globally. This approach can create new economic roles, such as AI system administrators in local hospitals and regional language content curators, rather than solely displacing existing jobs.
Sarvam AI and Krutrim emphasize "sovereign AI," not as nationalism, but as structural independence. Local models keep data within national jurisdiction, are fine-tuned for local contexts, and operate on local economic principles, avoiding dependency on foreign companies' priorities and pricing.
The success of India's approach hinges on core principles: build open, build cheap, build at scale, mirroring the logic behind India's Unified Payments Interface (UPI). While not every country can replicate India's scale or talent pool, the core principles of adapting open-source models for local languages and low-resource environments offer a viable path forward.
The future of AI is not a single, dominant model, but thousands of specialized models shaped by local needs and constraints. India is demonstrating that this inclusive, constraint-driven approach is not only feasible but offers a competitive advantage, providing more valuable AI solutions for the majority of the world's population.