> ## Content Index
> Fetch the complete content index at: https://aipriorities.digitalfutureslab.in/llms.txt
> Use this file to discover other available public pages before exploring further.

# Walking the Sovereignty Talk: Public Linguistic Infrastructures for India’s Language Commons
- URL: https://aipriorities.digitalfutureslab.in/walking-the-sovereignty-talk-public-linguistic-infrastructures-for-indias-language-commons/
- Published: 2026-10-03T17:56:59.000Z
- Updated: 2026-10-03T17:56:59.000Z
- Author: Hari
- Tags: #Import 2026-10-12 10:52

> **Anita Gurumurthy and Nandini Chami** 

**Abstract**

The development and diffusion of LLMs represents a disruptive watershed following the print, mass media, and social media revolutions. As digital artefacts acquire everyday materiality and societal-scale impacts, questions of “what next” become increasingly urgent.

As diverse linguistic cultures are appropriated within the flawed and homogenising architectures of Big Tech’s Generative AI, developing countries face new risks – from infrastructural dependency to local self-determination and cultural autonomy. This essay examines India’s emerging policy response to this challenge, distilling takeaways for public linguistic infrastructures. It argues that democratised innovation futures depend on understanding the changing landscape of language infrastructures and developing public responses that can sustain a flourishing language commons.

***A living language is a throbbing, vital thing, ever changing, ever growing and mirroring the people who speak and write it. It has its roots in the masses, though its superstructure may represent the culture of a few… If a language loses touch with the people, it loses its vitality and becomes an artificial, lifeless thing instead of the thing of life and strength and joy that it should be. Attempts to force the growth of a language in a particular direction are likely to end in distorting it and crushing its spirit.*** 

***\- Jawaharlal Nehru, the first Prime Minister of Independent India, at a pre-independence convening in 1937***

**1\. LLMs as the Infrastructure of Empire** 

The LLM moment has generated considerable anxiety and speculation about what it means for humans to encounter the exteriorisation of what Noam Chomsky referred to as the “mental organ” of language. A vital component of the complex of cognitive structures, the internalised syntax of language learning has long been considered uniquely human. The critique of the computational revolution in text generation and language processing is therefore unsurprising – LLMs, it is argued, lack the human intentionality integral to artistic expression; they are “stochastic parrots” not capable of true cognition; and their abilities are trivial when contrasted with the human mind. Hence, the alarm bells that these technologies may erode critical thinking.

This paper brings a different entry point to the debate. Acknowledging the everyday materiality of digital artefacts and their societal-scale impacts, we ask the “what next” question. How should we understand this phase of language infrastructures in human civilisation, and what responses are necessary to preserve the thriving unruliness of the language commons? We situate the discussion in the context of India’s emerging policy vision and strategy in relation to linguistic AI infrastructures. 

The development and diffusion of LLMs represents a foundationally disruptive watershed after the revolutions brought about by print, mass media, and social media. Every stage in the technology of language has materially altered the structures and patterns of social communication. Each has had implications for the production of language as a commons – that is, the protocols and processes through which communities co-construct language, engage in contextual meaning-making, and negotiate the fine balance between the impulses towards unification and diversification in preserving shared linguistic heritage and cultural memory. To paraphrase Marshall McLuhan, the medium has always mattered.

From this perspective, LLMs can be understood as infrastructures of modern-day empire, enabling a new form of colonialingualism –the perpetuation of imperialistic and neoliberal worldviews through the subsumption of indigenous and minoritised languages into a dominant language monoculture. Generative AI in text marks a new point of departure in the colonial geopolitics of knowledge and its political economy. Currently, the LLM race is dominated by firms from just two countries: the United States and China. Research published in early 2026 suggests that while US models have historically had an edge in market dominance, Chinese models are rapidly catching up, with some reports also claiming that they are more efficient. In both countries, a few firms dominate the innovation landscape, enjoying control of data centers, cloud services, chip supply chains, and energy exposure. The OpenAI-NVIDIA-Microsoft triangle in the U.S. exemplifies this oligopolistic trend, as does the investment duopoly of Alibaba and Tencent in China’s six “AI tiger” start-ups: Moonshot, Minimax, Zhipu, Baichuan, 0.1AI and Stepfun. Firms in both countries are developing massive multilingual models. Google, Microsoft and Meta support more than 100 languages, as does DeepSeek, while Alibaba’s Qwen3 offers support for 119 languages. Baidu is catching up rapidly through its LLM model ERNIE, which supports 22 languages, and is complemented by AI-powered translation services for over 200 languages.

Less than 5 per cent of the roughly 7,000 languages spoken around the world today have meaningful online representation. Technology adoption among a linguistic community depends, among other factors, on the availability of infrastructures, including the incorporation of a language into script-encoding standards such as Unicode or in usable typefaces and keyboards, and the availability of computational expertise/labs for natural language understanding and machine translation. 

The absence of computational infrastructures affects both the availability and quality of training data for building language models. This is known as the “resourcedness gap” in languages. From this perspective, most of the world’s languages can be classified as “low-resource languages”, characterised by inadequate and unrepresentative training data to guide LLM development. At the other end of this spectrum are high-resource languages, with English (low complexity) and Mandarin (high complexity) emerging as dominant due to the global political economy of AI innovation. 

Massive multilingual LLM models aim to improve performance for more languages through cross-lingual transfer learning – a technique through which a “model improves its low-resource language performance by inferring universal or shared linguistic patterns from high-resource languages”. However, such cross-lingual transfer learning has a range of limitations. 

First, transfer learning produces an “illusion of locality” wherein multilingual LLM outputs appear culturally situated at first glance due to continuous learning from user input, but organise meaning through epistemological assumptions and worldviews embedded in the high-resource, dominant language. Such language modelling biases promote “a simplistic understanding of diversity which does not do justice to the more profound differences that languages, and ultimately the communities that speak them”, embody. LLMs trained predominantly on English data “make key decisions in a representation space closest to English, regardless of their input and output language”. This can lead to cultural misalignment when applied to their contexts of operation and a range of harms, such as “disparate access due to performance gaps, imposition of hegemonic classifications, violation of cultural values, misinformation, or stereotypes about cultures, misrepresentation of cultural experiences, and outright erasure of cultures.” The problem also extends to LLMs in other languages. For example, there is evidence of LLMs in Mandarin (a high-resource Chinese language) failing to respond appropriately to the cultural nuances of Cantonese (a low-resource Chinese language). 

Second, LLMs are shaped by the ideologies of their creators. As scholars Kerche, Zook, and Graham highlight, bias “is not a correctable anomaly but an intrinsic feature of generative AI, rooted in historically uneven data ecologies and design choices.” Their research on LLMs in English demonstrates how a “silicon gaze” in these models entrenches geographies of coloniality, perpetuating shallow cultural stereotypes. Design choices in LLMs are based on performance benchmarks set by a few powerful firms that reward guessing. In the AI race for market domination, these models are designed as test-takers, programmed to hallucinate rather than to honestly admit uncertainty. In transfer learning, the model is bound to inherit and even intensify “hallucinated” information. 

Third, transfer learning tends to run into a computational wall, requiring substantial memory for fine-tuning and long training times. While it lowers initial costs, it does not entirely eliminate the high computational complexity of deep learning. As the number of languages supported by the model increases, a transfer-dilution trade-off occurs, diminishing per-language capacity and consequently impacting model performance. 

Major technology firms are actively pushing LLMs and generative AI in the Global South through massive infrastructure investments, partnerships, and customised, low-cost services. The dominant discourse also nudges governments in the South to pursue LLMs as the key driver of enterprise productivity and technological leapfrogging. Given that local languages in the developing world often lack the training data required for AI development, and building monolingual LLM models attuned to cultural nuances requires resource investments of a high order, many countries in the majority world face a Hobson’s choice. Dependencies on American and Chinese Big Tech infrastructures become inevitable, creating heightened risks not only for national political and economic sovereignty, but also for local cultural autonomy, as linguistic cultures are appropriated into technological enclosures that may not adequately reflect their complexity and nuance.

We now proceed to examine India’s policy response to this challenge.

**2\. India’s AI Sovereignty: Uneasy Negotiations Between Foundation and Superstructure** 

Digital sovereignty in general, and LLM sovereignty in particular, are recurring themes in India’s policy discourse. The India AI Mission was launched in March 2024 with an outlay of INR 10,372 crore (approximately USD 1.12 billion). The Mission adopts a four-pronged approach, with allocations for:

- Compute Capacity: Significant funds for over 10,000 Graphics Processing Units (GPUs) via public-private partnerships, expanding to over 58,000 GPUs.
- Startup Funding: Dedicated financial support for AI startups, covering product development and commercialisation.
- AI Models & Datasets: Development of indigenous foundational models and curated datasets.
- Centers of Excellence: Establishment of AI Centers of Excellence (CoEs) for healthcare, agriculture, and sustainable cities, with additional focus on education.

Financial assistance, including actual costs of compute, have been provided to twelve teams. BharatGen – set up as a not-for-profit foundation through a consortium of India’s leading public technology institutions and supported by the Department of Science and Technology (DST), Government of India – received more than Rs 1,000 crore from the Mission for developing India’s sovereign, multimodal LLM. BharatGen reflects India’s aspiration for building a DeepSeek-like AI infrastructure that can compete with global models. Other major recipients of financial assistance, such as Sarvam AI, are focused on developing large-scale open-source and multilingual models.

The Mission also offers a 40 per cent subsidy on AI compute services on the cloud for approved projects, startups, and researchers. Subsidised compute is provided at roughly one-third of the global cost (around USD 0.65 per GPU hour vs USD 2.50-3.00 globally). While the government has sought to make high-performance compute accessible, the challenge remains one of streamlining access and generating enough demand to fully utilise the infrastructure. 

India is funding both compute and models, but many start-ups prefer more efficient full-stack systems – integrated ecosystems of hardware, software, libraries, and networking technologies – designed for the efficient training of LLMs. Start-ups are also choosing Microsoft Azure, Amazon Web Services (AWS), or Google Cloud Platform (GCP), thanks to VC funding that includes bundled cloud credits. These procurement lock-ins within hyperscaler ecosystems create a significant challenge for India. The shortage of demand for subsidised compute infrastructure presents a real conundrum: the existing demand is already captured by foreign players. While government programs can help in creating baseline utilisation, enabling the Indian private sector to compete with hyperscalers on both price and reliability remains an uphill task. Meanwhile, Big Tech players have also pledged investments worth over USD 67 billion in India since October 2025 for local data processing and a wider compliance zone, packaging this as an offering towards “sovereign” AI. Particularly noteworthy is the rise of NVIDIA’s Nemotron open models to build multilingual AI systems for government, finance, and enterprise use in India. As a full-stack AI provider, NVIDIA has partnered with BharatGen to train Param-2 on its AI Enterprise software and infrastructure, including NVIDIA NeMo open libraries and Base Command Manager for scalable model training and deployment. In early April 2026, NVIDIA and Amazon joined other big investors to back Sarvam AI’s mega fundraising round. Sarvam AI has also joined NVIDIA’s global open AI coalition.

The enthusiasm surrounding these collaborations and the prospect of AI built on Indian soil for Indic languages may be misplaced. Jurisdictional or operational control does not necessarily translate into sovereignty in the sense of technological independence. Partnerships in which the underlying training stack remains foreign-owned tend to transfer value from and control over local innovation infrastructures, eventually strengthening the capability and market power of Big Tech despite generating immediate benefits for the local ecosystem. Also, the Indian government’s 20-year tax holiday for foreign cloud-service providers using India-based data centres is, ironically, a concession that benefits foreign corporate shareholders of Indian start-ups, thus subverting any policy intent to protect and promote national sovereignty.

The adoption of AI applications through public-private partnerships across a range of contexts in India presents important sovereignty dilemmas that need to be carefully evaluated. Several initiatives in rural India, for instance, deploy AI-powered conversational agents of American Big Tech for English language learning. Built on datasets that are largely foreign, these tools may not meet learners where they are. Second, in the case of children, these interactions give platforms a degree of influence over learning and development that could lead to unhealthy cognitive and affective dependencies of the children on the AI. Third, the data generated through these interactions that are harvested by the platform may not meet ethical standards of prior and informed consent, with purpose limitation. Fourth, and in the case of partnerships involving public institutions, the proprietary tech model creates a vendor lock-in, potentially increasing the public system’s dependence on the vendor.

The sovereignty debate also raises the abiding question of “we-the-people.” From this standpoint of the centrality of the will of the people conferred by the Constitution, the development and deployment of AI in India without an independent, techno-legal evaluation architecture poses important concerns. At present, India lacks a robust institutional mechanism to monitor and address representational bias and epistemic injustice in AI, particularly in its indigenous LLM ecosystem. Consistent with its stated policy approach of relying on voluntary guidance for AI safety, the Ministry of Electronics and Information Technology issued guidance in December 2025, urging LLM developers supported by the India AI Mission to ensure that “sensitive connotations linked with caste, gender, food practices, regional and linguistic stereotypes, as well as ethnic and religious differences” be handled with utmost care. A few months earlier, the Ministry had also invited expressions of interest for the stress testing of AI models through adversarial prompts.

Experience from other contexts, particularly the EU, makes it clear that AI governance will likely require much more than benevolent volunteerism. The prevalence of caste bias in global multilingual LLMs, for instance, is an issue of legal-constitutional concern in India that indigenous language models need to address. 

In a winner-take-all digital economy, true AI self-sufficiency is difficult. As a World Economic Forum article argues, tech sovereignty is multi-dimensional, encompassing data, operational, technical and even legal dimensions. While India’s policy sentiment reflects a justifiable ambition to navigate this complex political economy, empirical realities point to uneasy manoeuvres on the AI road. The continuing dominance of large foreign technology firms and India's dependence on external infrastructures complicate claims of techno-economic independence. 

Moreover, despite enormous capital investment, the global generative AI boom is yet to trigger a real innovation economy. In this state of suspended animation, India’s policy may be veering towards securing a possible share in tomorrow’s global market. Meanwhile, local partnerships with Big Tech seem like a win-win situation in otherwise under-resourced institutional settings. At the same time, equal attention must be paid to building the institutions of the future – the social arrangements capable of democratising innovation, protecting the public interest, and nurturing the country’s rich civic and cultural foundations.

We now turn to the question of what a forward-looking AI policy must do to address these challenges.

**3\. Imagining India’s Linguistic Infrastructures for Democratised Innovation Futures** 

Language is humanity’s foundational intangible heritage, serving as the primary vehicle through which culture, values, and knowledge are transmitted across generations. A flourishing language commons for India must be aligned with the constitutional vision of linguistic secularism and freedom, ensuring a pluralistic, non-discriminatory approach to the country's diverse linguistic landscape. 

Historically, developments in technology have always opened up new possibilities – transforming linguistic cultures and producing linguistic infrastructures that mediate meaning-making. History reveals not only an evolution of the disciplining structures of text, but also the often-subversive intent of human creativity that challenges these structures. The vitality of communities in the Gen AI moment depends on a new dialectic between the techno-social conditions that computational language models forge and the language commons of human creativity.

India’s policy challenge lies in mapping and responding to this challenge. Through publicly debated norms and standards for linguistic infrastructures, AI policies must strengthen the Nehruvian ideal of language as “the thing of life and strength and joy” for all. 

The following considerations are key in this regard:

**a. Incentives for Community-Led Co-Creation of Innovation Cultures**

State-led initiatives like AI Kosh – an anonymised dataset platform in multiple regional languages – and Bhashini, a program which provides open source translation APIs in over 22 Indian languages, are attempting to fill an important resource gap in the AI ecosystem. Similarly, the Ministry of Tribal Affairs released Adi Vaani, India’s first AI-Powered Translator for Tribal Languages. The initiative seeks to support the digitisation of endangered languages while facilitating tribal entrepreneurship and access to services. 

The real benefits of AI, however, depend on how such crucial public infrastructures are leveraged by communities to drive innovation that serves their local economy and society. This calls for sectoral and domain-specific policies and programs designed for application-layer innovations, especially inference AI services, built on top of these state-led foundations. 

As the history of the internet demonstrates, lasting innovation cultures are born from and sustained through community-led peer production. In the realm of language and culture, community enterprise and co-creation may be the only antidote to the homogenising impulses of top-down state-led infrastructure systems. The experience of China’s state-led LLM and generative AI ecosystem demonstrates that an overreliance on state-led curation of linguistic corpora can reinforce hegemonic standpoints. 

Policies must therefore incentivise community-led co-creation. The Chandamama Kathalu Telugu SLM corpus preparation drive, for instance, brought together college students to develop language corpora through a community-led effort. Public support for such projects can encourage local innovation cultures, also providing possible access to valuable training data for foundation models such as BharatGen. Low-resource and morphologically complex languages like Tamil – spoken across many countries and enjoying a wealth of non-digital resources – have benefitted from projects like Aalamaram, a large-scale, open-source Tamil language treebank initiative designed to advance Natural Language Processing (NLP) and AI for the Tamil language. Featuring nearly 10,000 annotated sentences, the project is considered the largest publicly available treebank for Tamil and was led by a team of researchers from various institutions in India, Singapore, and Sri Lanka. Aalamaram demonstrates the potential of transnational scholarly collaboration in creating and refining the vital computational resources needed to support foundational language models. 

**b. Prioritising Localised Models and Building for Resource-Constrained Contexts**

In countries like India, where governance priorities are shaped by urgent socio-economic and development challenges, AI models attuned to the local context can potentially support locally led solutions. 

Small Language Models (SLMs) and decentralised compute are converging to enable a new paradigm of AI, shifting away from massive, centralised cloud models towards localised, privacy-preserving, and efficient intelligence. While these developments remain at an early stage, SLMs — compact models designed for specialised tasks or specific domains — offer significant potential for deployment on resource-constrained hardware, including mobile devices.

India’s Economic Survey 2025-26 notes that several constraints render the pursuit of foundational models as the centrepiece of India’s AI strategy challenging. According to the survey, access to cutting-edge compute infrastructure is limited, financial resources for large-scale model training are scarce, and private participation in foundational AI research remains relatively modest. It therefore recommends a bottom-up approach to AI development that is aligned more closely with these realities.

A turn towards small models, however, will not automatically address the deficits of India’s LLM strategy discussed in this essay. Nor should LLMs and SLMs be viewed as binary choices. India’s policy process therefore requires a strategic stocktaking of its AI infrastructure strategy. The balance hinges on navigating the trade-offs between developing domestic AI infrastructure, leveraging global, open-source, or cost-effective foreign technologies, and pursuing a path towards SLMs. The guiding principle in this regard should be the democratisation of innovation dividends – embracing choices that add to public value.

**c. Adopting a (Gen)AI Governance Approach Rooted in Constitutional Rights**

The oft-invoked dichotomy between innovation and human rights perpetuates a laissez-faire economic logic – that free markets protect human freedoms. This obscures a more fundamental truth – that innovation is inherently a public good and must pass the litmus test of contributing to societal wellbeing. Given the reproduction of systemic injustice in LLMs and their behaviour, which renders traditional safety approaches inadequate, public evaluation standards for AI architectures are essential. In the context of today’s LLMs, colonial worldviews, racial prejudice, and gender bias are often embedded within the systems. Localisation through culturally attuned training data does not, by itself, solve these problems. Flawed AI architectures present a serious threat to the social, economic, and cultural rights of individuals, especially those from marginalised groups. In doing so, they threaten to undermine India’s longstanding constitutional and legal commitments to anti-discrimination, equality, and justice. 

Big Tech-led industry lobbies are prone to safety-washing. India’s forays into language models need a strong oversight system – consolidated authority, licensing regimes, mandated training data and modelling disclosures, formal verification of system behaviour, and the capacity for rapid intervention. A national registry of large models, along with independent monitoring mechanisms for effective enforcement of architectural standards, is necessary. This system needs to be grounded firmly in the rights enshrined in India’s enduring constitutional legacy, drawing upon specific rights in various domains – the right to education, gender equality, and more. The cultural superstructures of Gen AI are contingent on the values encoded into their foundational technologies. From this perspective, AI governance acquires indisputable urgency; it is not merely a technical challenge but a constitutional imperative.

**d. Leading from the South with a Value-Based, Anti-Colonial Approach**

The balancing act required to pursue technological sovereignty in deeply unequal digital markets presents India with a diplomatic opportunity: to lead the discourse on the global digital commons. UNCTAD’s Technology and Innovation Report 2025 has proposed a decentralised “CERN” model for compute infrastructures, highlighting the potential for pooling resources and expertise based on the principles of international cooperation and open science. Working alongside other countries of the Global South, India can push for a CERN-like collaborative international mechanism to enable equitable AI infrastructures.

Development futures may be tied more closely with task-specific SLMs, domain-specific applications, and inference models that enhance domestic productive capacity in key sectors. While strategic considerations may justify investments in home-grown LLMs, India must concurrently incentivise domestic AI expertise for inference models that can be adapted to other contexts in the Global South. Such an approach could also enhance exports in AI services. 

The global race for dominance in the age of the LLM represents an unprecedented challenge. Existing international governance frameworks have struggled to respond in a just and effective way to the nexus between technology behemoths and finance capital. The resulting crisis is not only one of the distributive integrity of the knowledge economy, but also of cultural self-determination of all societies. There is an opportunity here for rebuilding the world order with a South-led perspective in which India, as a cradle of civilisational pluralism, has a significant stake. 

As we continue to grapple with the implications of the LLM moment, the *Thiruvananthapuram Declaration on A New Innovation Ecosystem for Our Collective Digital Futures* – co-authored by public interest technologists, cooperatives, scholars and civil society – offers a timely reminder: the infrastructural public goods of today will determine the cultural commons of tomorrow.