India must pursue AI sovereignty—targeting key layers like semiconductors, compute, and models—balancing autonomy with global collaboration
As artificial intelligence (AI) becomes a central driver of economic and geopolitical competition, governments across the world are increasingly pursuing AI sovereignty. In India’s case, AI sovereignty is a crucial part of its strategic autonomy. AI adoption could contribute an additional US$500–600 billion to the national GDP by 2035. However, for India, sovereignty across the entire AI stack—from chips to applications—may neither be feasible nor necessary.
The end-to-end cost of building sovereign AI systems runs into the hundreds of billions of dollars, a scale difficult for India to sustain given constraints in private capital and the entrenched dependence on global supply chains. Frontier capabilities today remain concentrated in the United States (US) and China. Thus, attempting to outcompete them is likely to be economically inefficient, especially as long development timelines risk lagging behind rapidly shifting technological frontiers. Consequently, the strategic challenge lies in prioritising which layers of the AI stack India should seek to control rather than pursuing full-stack sovereignty.
The global AI ecosystem is rapidly consolidating into a US–China duopoly across critical layers of the stack—advanced semiconductors, hyper-scale compute, and foundation models—concentrating both technological capability and economic value. China’s scale-driven approach is evident in its research output, producing more AI-related papers than the United States, European Union, and United Kingdom combined. Its model innovation is also accelerating: the release of DeepSeek V-3, reportedly built at 6 percent of the cost of GPT-4 while achieving competitive benchmark performance, has catalysed a new wave of domestic AI startups. This has significantly expanded the market, now on track to surpass US$200 billion by 2029.
The US has achieved AI dominance through a combination of proactive action, profound private capital investments and frameworks to promote speed, efficiency and scale.
In addition to this, China recently developed a working prototype of the Extreme Ultraviolet (EUV) lithography machine. This has dismantled the monopoly of ASML over producing advanced semiconductor chips (below 7 nm to 2 nm nodes). China has also invested in indigenous semiconductor capabilities through Huawei, partly catalysed by US export controls imposed since 2022—signalling broader technological self-reliance. Policy direction from the State Council further targets rapid AI diffusion across a majority of economic sectors within the decade. This state-backed push has been matched by capital mobilisation: China’s share of global private AI investment surged from 11 percent in 2016 to 48 percent in 2017, briefly surpassing the US.
The US has achieved AI dominance through a combination of proactive action, profound private capital investments and frameworks to promote speed, efficiency and scale. To maintain its dominance, capturing about 50 percent of the global semiconductor market, export controls for advanced chips were introduced and later reviewed, citing national security concerns. The US also leads in the number of data centre facilities, representing 40 percent of the world’s total capacity. It also has significant advantages in terms of integrating into already established big tech ecosystems, across multiple layers of the AI stack. This has enabled it to quickly become a provider of a full-stack ecosystem, thanks to its deep tech research and development (R&D)investments, which exceeded US $109 billion in 2024, compared with US$1.4 billion in India.
This concentration has significant implications for India. Despite accounting for approximately 17 percent of the global IT services market, India captures only around 1 percent of high-value global technology value pools, including AI, semiconductors, and hyper-scale infrastructure. This reflects a structural imbalance: India participates in AI deployment but remains largely absent from the layers where value is created and retained.
AI sovereignty, therefore, is not merely a question of security, but of economic strategy. Without domestic capabilities in compute infrastructure, foundational models, and data governance, India risks becoming a net importer of intelligence, capturing limited downstream gains while exporting economic rents. Strengthening data sovereignty and investing in domestic AI infrastructure will be critical to ensuring that value generated from India’s data and market scale is retained domestically, while enabling the country to shape governance norms for the Global South and safeguard the broader digital ecosystem.
Training frontier AI models has become exponentially more expensive, with overall costs of training increasing by roughly 2.4 times per year since 2016. A significant chunk of these costs for top models, comes from R&D (between 29 and 49 percent of the total amortised cost) and compute hardware (47–64 percent). Further, with the US and China leading frontier model developments, similar global penetration of Indian sovereign models is likely to remain limited due to entrenched disadvantages in compute, capital, and platform ecosystems. Frontier AI capabilities also continue to concentrate among a small number of firms, and late entrants face structural barriers to achieving comparable scale and cross-border diffusion.
Developing state-of-the-art LLMs domestically for India and the world is a crucial long-term aim to capture both the socio-economic and geoeconomic benefits of AI.
India’s response has therefore centred on purpose-built small language models (SLMs), as was demonstrated during the recent New Delhi AI Impact Summit. These enable cost-effective deployment, localisation, and integration across sectors. However, while SLMs are critical for diffusion, their economic impact remains limited as compared to general-purpose technologies built on large language models (LLMs), which generate cross-sector productivity gains. Generative AI alone is estimated to contribute US$2.6–4.4 trillion annually to the global economy, with India projected to see productivity gains exceeding 40 percent in its IT services sector. However, this creates an asymmetry between value creation and value capture: while India can drive widespread AI adoption through SLMs, a disproportionate share of economic rent is likely to go to firms controlling foundational LLMs, compute infrastructure, and global platforms. Thus, developing state-of-the-art LLMs domestically for India and the world is a crucial long-term aim to capture both the socio-economic and geoeconomic benefits of AI.
India launched its first fully sovereign multimodal LLM, BHARATGen, in June 2025, with support from the Department of Science and Technology (DST). Multiple startups such as SarvamAI, Gnani AI and Socket AI have also developed sovereign models, taking advantage of Nvidia GPUs procured through the India AI mission. Sarvam models, unlike China’s DeepSeek or US’s ChatGPT-5, run on 105 billion parameters, with 10.3 billion active parameters per token. Indian models have prioritised efficiency and localisation over scale, reflecting constraints in compute and training resources, but they allow for greater flexibility at the application layer.
| Country | Models | Open vs Closed | Benchmark position | Training Cost | Compute Scale | Capability level |
| India | Sarvam, BharatGPT, Soket AI | Mostly open/hybrid | Mid-tier / limited | <$5M–$30M | 10^2–10^3 GPUs | Domain-specific |
| US | ChatGPT-5 Gemini-3, Claude | Mostly closed-source | State-of-the-art | $30M–$100M+ | 10^4–10^5 GPUs | General-purpose (frontier) |
| China | DeepSeek V3.2 Qwen, ERNIE | Hybrid (open + closed) | Near-frontier | $10M–$50M | 10^3–10^4 GPUs | Near-general-purpose |
Source: Author’s Compilation
India’s dependence on global supply chains for computing has prompted a push toward domestic semiconductor capabilities under initiatives such as the India Semiconductor Mission 2.0. The recently established Micron facility in Gujarat, focused on Assembly, Test, Mark, and Pack (ATMP), represents an initial step in building manufacturing capacity, while the ₹91,000 crore Tata–PSMC fab aims to develop commercial fabrication capabilities at mature nodes. Together, these efforts seek to position India as an alternative destination for semiconductor manufacturing and reduce import dependence.
The growing adoption of specialised accelerators, such as Tensor Processing Units and other domain-specific chips, reflects a shift toward more efficient, application-specific compute architectures.
However, the rapid evolution of AI hardware complicates this trajectory. The growing adoption of specialised accelerators, such as Tensor Processing Units and other domain-specific chips, reflects a shift toward more efficient, application-specific compute architectures. At the same time, physical limits to silicon scaling below advanced nodes and the emergence of alternative paradigms, including photonic and neuromorphic computing, indicate that the technological frontier is continuously shifting. In this context, while the West has committed US$40-280 billion toward semiconductor ecosystems, India’s allocations are modest, underscoring a significant gap in deep-tech R&D investment.
A sovereign approach to AI is critical for securing India’s national interests and enabling population-scale benefits. A strategy of selective AI sovereignty—targeting priority sectors such as AI-enabled services, cybersecurity, aerospace, defence, and semiconductors—can better align investments with policy priorities while maximising long-term economic and technological gains. Sustained investments in semiconductor innovation and sovereign compute can generate economy-wide spillovers, reinforcing capabilities across strategic sectors such as defence, advanced manufacturing, and critical infrastructure. Complemented by mutually beneficial partnerships with like-minded countries, such an approach can help bridge gaps in R&D, compute and talent while maintaining strategic autonomy.
India must move beyond being an adoption accelerator for the Global South to becoming a selective leader in key sectors aligned with its core economic strengths.
In this context, AI sovereignty should be understood not as full-stack self-sufficiency, but as strategic control over critical layers of the AI ecosystem. India must move beyond being an adoption accelerator for the Global South to becoming a selective leader in key sectors aligned with its core economic strengths. Done right, this approach can translate into both economic and geopolitical leverage, positioning India as a credible ‘third way’ in AI, where scalable innovation is aligned with responsible, human-centric governance.
Ishita Deshmukh is a Research Intern with Observer Research Foundation, Mumbai.
The views expressed above belong to the author(s). ORF research and analyses now available on Telegram! Click here to access our curated content — blogs, longforms and interviews.
Ishita Deshmukh is a Research Intern with Observer Research Foundation, Mumbai. ...
Read More +
PREV
