“ROI doesn’t come from creating a very large model; 95% of work can happen with models of 20 or 50 billion parameters.” – Ashwini Vaishnaw – Minister of Electronics and IT, India
Delivered at the World Economic Forum (WEF) in Davos 2026, this statement by Ashwini Vaishnaw, India’s Minister of Electronics and Information Technology, encapsulates a pragmatic approach to artificial intelligence deployment amid global discussions on technology sovereignty and economic impact1,2. Speaking under the theme ‘A Spirit of Dialogue’ from 19 to 23 January 2026, Vaishnaw positioned India not merely as a consumer of foreign AI but as a co-creator, emphasising efficiency over scale in model development1. The quote emerged during his rebuttal to IMF Managing Director Kristalina Georgieva’s characterisation of India as a ‘second-tier’ AI power, with Vaishnaw citing Stanford University’s AI Index to affirm India’s third-place ranking in AI preparedness and second in AI talent2.
Ashwini Vaishnaw: Architect of India’s Digital Ambition
Ashwini Vaishnaw, a chartered accountant and IAS officer of the 1994 batch (Muslim-Rajasthan cadre), has risen to become a pivotal figure in India’s technological transformation1. Appointed Minister of Electronics and Information Technology in 2021, alongside portfolios in Railways, Communications, and Information & Broadcasting, Vaishnaw has spearheaded initiatives like the India Semiconductor Mission and the push for sovereign AI1. His tenure has attracted major investments, including Google’s $15 billion gigawatt-scale AI data centre in Visakhapatnam and partnerships with Meta on AI safety and IBM on advanced chip technology (7nm and 2nm nodes)1. At Davos 2026, he outlined India’s appeal as a ‘bright spot’ for global investors, citing stable democracy, policy continuity, and projected 6-8% real GDP growth1. Vaishnaw’s vision extends to hosting the India AI Impact Summit in New Delhi on 19-20 February 2026, showcasing a ‘People-Planet-Progress’ framework for AI safety and global standards1,3.
Context: India’s Five-Layer Sovereign AI Stack
Vaishnaw framed the quote within India’s comprehensive ‘Sovereign AI Stack’, a methodical strategy across five layers to achieve technological independence within a year1,2,4. This includes:
- Application Layer: Real-world deployments in agriculture, health, governance, and enterprise services, where India aims to be the world’s largest supplier2,4.
- Model Layer: A ‘bouquet’ of domestic models with 20-50 billion parameters, sufficient for 95% of use cases, prioritising diffusion, productivity, and ROI over gigantic foundational models1,2.
- Semiconductor Layer: Indigenous design and manufacturing targeting 2nm nodes1.
- Infrastructure Layer: National 38,000 GPU compute pool and gigawatt-scale data centres powered by clean energy and Small Modular Reactors (SMRs)1.
- Energy Layer: Sustainable power solutions to fuel AI growth2.
This approach counters the resource-intensive race for trillion-parameter models, focusing on widespread adoption in emerging markets like India, where efficiency drives economic returns2,5.
Leading Theorists on Small Language Models and AI Efficiency
The emphasis on smaller models aligns with pioneering research challenging the ‘scale-is-all-you-need’ paradigm. Andrej Karpathy, former OpenAI and Tesla AI director, has advocated for ’emergent abilities’ in models as small as 1-10 billion parameters, arguing that targeted training yields high ROI for specific tasks1,2. Noam Shazeer of Character.AI and Google co-inventor of Transformer architectures, demonstrated with models like Chinchilla (70 billion parameters) that optimal compute allocation outperforms sheer size, influencing efficient scaling laws1. Tim Dettmers, researcher behind the influential ‘llm-arxiv-daily’ repository, quantified in his ‘BitsAndBytes’ work how quantisation enables 4-bit inference on 70B models with minimal performance loss, democratising access for resource-constrained environments2.
Further, Sasha Rush (Cornell) and collaborators’ ‘Scaling Laws for Neural Language Models’ (2020) revealed diminishing returns beyond certain sizes, bolstering the case for 20-50B models1. In industry, Meta’s Llama series (7B-70B) and Mistral AI’s Mixtral 8x7B (effectively 46B active parameters) exemplify mixture-of-experts (MoE) architectures achieving near-frontier performance with lower costs, as validated in benchmarks like MMLU2. These theorists underscore Vaishnaw’s point: true power lies in diffusion and application, not model magnitude, particularly for emerging markets pursuing technology strategy5.
Vaishnaw’s insight at Davos 2026 thus resonates globally, signalling a shift towards sustainable, ROI-focused AI that empowers nations like India to lead through strategic efficiency rather than brute scale1,2.
References
3. https://www.youtube.com/watch?v=3S04vbuukmE
4. https://www.youtube.com/watch?v=VNGmVGzr4RA
5. https://www.weforum.org/stories/2026/01/live-from-davos-2026-what-to-know-on-day-2/
