Frugal AI: Shifting from Silicon Valley Dependency to Sustainable Deployment
Serish Gandikota advocates for a design-first approach to AI implementation beyond hyperscale models.
1Redefining AI Innovation
Serish Gandikota, co-founder of the Frugal AI Hub at Cambridge University, argues against blindly replicating Silicon Valley's resource-heavy AI models. Instead of simply building cheaper versions of existing tech, he proposes a design philosophy that prioritizes specific organizational outcomes. This approach focuses on solving local problems under constraints of cost and capacity.
2Lessons from Fieldwork
Gandikota’s perspective is rooted in past failures, including a 2008 Arctic expedition where bent equipment rendered a project impossible despite significant resources. This experience highlighted that abundance does not guarantee success. He contrasts this with the Embrace baby warmer, a 200-dollar innovation that replaced expensive 20,000-dollar incubators through clever design.
3The Frugal AI Hub
Launched in April 2025 at the Judge Business School, the Hub collaborates with global partners to transition AI from pilots to sustained real-world use. The team recently partnered with the UN International Computing Centre to publish a report on the total cost of ownership and social impact for AI projects.
4Global Deployment Labs
The Hub is establishing deployment labs to build local capacity and co-create solutions. A notable initiative includes an MoU signed in November 2025 with the Andhra Pradesh government under the Amaravati Quantum Valley project. These labs aim to identify specific regional problem statements and design pilots that can eventually scale effectively.
5Design-First Architecture
Frugal AI is presented as a systems architecture rather than just a technical optimization. Before selecting a model, organizations must establish governance, workflows, and change management processes. This framework helps leaders decide whether they truly require frontier technology, open-source models, or small language models based on their specific operational needs.
6Addressing the Absorption Gap
Gandikota emphasizes that the primary constraint in AI is no longer intelligence, but absorption. While supply-side innovation is exponential, institutional ability to deploy these solutions lags behind. His forthcoming book provides a toolkit for leaders to make better investment decisions before dependencies become too difficult or costly to reverse later.
7Infrastructure Constraints
Even in developed regions like the UK and Europe, hyperscale AI faces physical limitations. Building massive data centers requires significant land, water for cooling, and electricity grid capacity. These regions are increasingly recognizing that they must operate within their own constraints, similar to the Global South, rather than chasing Silicon Valley.
8India's Strategic Position
India leverages its experience in building digital public infrastructure like Aadhaar and UPI to navigate the AI race. Sovereignty is a major driver, as reliance on foreign computing capacity poses risks. Furthermore, the country is focusing on foundational models to address the linguistic diversity of its massive population, which remains underserved.
9Practical Application Example
The MahaVISTAAR voice AI service for farmers serves as a primary case study. It manages complex inputs like background noise and regional dialects while ensuring information accuracy. The system relies on an orchestration layer where institutions maintain up-to-date databases for weather and fertilizers, proving that successful AI requires robust human-led coordination.
10Moving Beyond Reactivity
Organizations often suffer from a fear of missing out, adopting AI without clear objectives. Gandikota urges leaders to stop reacting to Silicon Valley trends and instead ask if AI is the necessary solution for their desired outcome. This shift in mindset is essential for achieving long-term, sustainable, and scalable digital transformation.
11The Future of Deployment
True competitive advantage will belong to organizations that can govern and absorb AI systems over time. By focusing on resource efficiency, data capture methods, and model building, entities can achieve significant results without needing hyperscale resources. Deployment remains the ultimate test for any AI strategy, requiring careful planning and structural alignment.