Blog: Sovereign in Name

A peculiar vocabulary forms the basis of India’s Artificial Intelligence (AI) ambitions today. At the massive AI Impact Summit that was organised by the Indian Government this year, there was one word that saw itself recurring through multiple sessions during the five-day summit: Sovereign. This word, in its most basic form, conveys autonomy. And in the summit’s context, staying true to its definition, it signalled a national capacity to build and control the systems on which economic and political life increasingly depends. This essay uses the word “sovereign” in a stricter sense and tries to analyse whether India owns, controls, and can independently sustain the substrate on which its AI ambitions rest.

To better understand the current scenario, we first turn to the thesis of technofeudalism introduced by the economist Yanis Varoufakis. His core argument rests on a distinction drawn from the Marxian interpretation of rent, understood in a context where profit and rent are two distinct kinds of income drawn from distinct sources. Profit is the return generated by ownership of the means of production and the exploitation of labour. Rent, on the other hand, is acquired through control of a scarce asset that others must use in order to produce, but the nuance here is that the owner of that asset produces nothing new. They simply control access and collect a toll on it. Varoufakis cleverly applies this distinction to cloud platforms, highlighting that firms that currently own the digital substrate collect rent on access rather than the profit on production. They are viewed as feudal by Varoufakis because they exercise power through fiefdoms that displace the two features central to a capitalist economy, i.e. markets and profits.

Before moving ahead, it is important to highlight that the above thesis is contested. Critics have rightly argued that what Varoufakis describes is not really the death of capitalism and should be looked upon as a form of mutation into a more concentrated and rent-heavy form. They also point out that markets still very much plainly exist, and the big tech firms he talks about remain embedded in ordinary industrial production contrary to being mere rent collectors. These arguments are logically sound and well noted. This essay’s focus, therefore, is on an analysis of the relation of access without ownership. This relation, the author believes, is the defining feature of the feudal bond, and exhibited quite overtly in India’s present AI position.

As discussed above, the tenet of a feudal relation is access without ownership. Traditionally, the serf worked on land he did not own and surrendered a portion of his surplus to a lord who laid claim over the land. Now if we look at this within the domain of AI today, a parallel can be observed. The present-day substrate of production is computational capacity, and control over this substrate is concentrated to a degree that is similar to some of the great monopolies of the industrial era. Nvidia held an estimated 70–80% of the AI Accelerator market by revenue across 2025, although this share is expected to soften through 2026 as competitors scale. So while the monopoly exists, it is not static. The most serious pressure here comes from the custom silicon of the American hyperscalers themselves, i.e. Google, Amazon, and Microsoft, rather than rival chipmakers.

India’s response to the above structure is quite substantial. The IndiaAI Mission was approved with an outlay of Rs. 10,371.92 crore (~$1.25 billion), for a period of five years and spread across seven pillars. On compute particularly, the Indian Government has stated its ambition of deploying tens of thousands of Graphics Processing Units (GPUs) at subsidised prices. The Indian Government informed the Lok Sabha on 1 April 2026 that more than 38,000 GPUs had been empanelled through 14 service providers, with another 20,000 being added. Apart from select premium GPUs, the average subsidised rate was about Rs. 65 per GPU per hour. Available secondary research places the current deployment at around 34k to 38k units, reported between late 2025 and early 2026, with ambitions to take this forward to 100k by the end of 2026. These units have been offered to startups and academic institutions at Rs. 115 to Rs. 150 per GPU per hour, rates described by Government officials as approximately 42% below the ongoing market rates.

However, while the Government’s ambitions sound promising, there is a gap between this ambition and actual expenditure. In a reply to the Rajya Sabha during the 2026 Budget Session, the Ministry of Electronics and Information Technology (MeitY) disclosed that it had released only Rs. 21.79 crore in 2024-25 against revised estimates of Rs. 173 crore, and Rs. 379.15 crore in 2025-26 against revised estimates of Rs. 800 crore, as of February 2026. No funds had been released for the 2026-27 financial year. The total disbursed across the first two years is just under Rs. 401 crore, which is under 4% of the approved outlay of Rs. 10,371.92 crore. This also points to the fact that while the architecture of dependence might exist, the value transfer through subsidised compute seems to be limited. India’s ambitions and goals are very much within the contours of an import-dependent design, although the rent that it might one day channel abroad remains mostly prospective for now.

The India AI Mission’s most notable achievement is the migration of Bhashini, India’s national multilingual platform, to an indigenous cloud infrastructure. According to an Indian Government press release, the platform was transitioned to Yotta’s Government Community Cloud and Shakti Cloud with zero data loss across 3.5 billion files. This migration, however, highlights a contradiction. The press release also claimed performance gains of up to 40%, cost savings of 20–30%, and a sustained uptime of 99.99%. These figures should be looked at cautiously, particularly because they are unaccompanied by disclosed benchmarking methodology. Additionally, beyond these claims, there’s a strange contradiction that seems to unearth itself. The Bhashini platform, which now runs on Yotta’s Shakti Cloud, is powered by Nvidia H100 GPUs and Yotta’s wider sovereign build is set to run on Nvidia’s Blackwell Ultra hardware. This begets the question of whether India’s AI sovereignty rests on American silicon. The reason for this, as certain news reports online have pointed out, is that there is no legal standard or definition of the term “sovereignty”, so it is easy for firms and governments to self-certify, while the data held by a provider that falls under a foreign jurisdiction may remain reachable by that jurisdiction regardless of where it is physically stored. Under the US CLOUD Act, the focal point is whether the provider or its parent is subject to US law, with the location being irrelevant. The asymmetry cuts against India in particular because the US has signed CLOUD Act agreements with the UK and Australia but not with India, so Indian agencies are left with the slower mutual legal assistance route.

The technofeudal lens adopted in this essay is to highlight the above contradiction. The basic idea is that India has advanced quickly on the layers it can build – language models, data, applications, etc. However, the layer that actually constitutes the means of production, the compute substrate in the contemporary era, remains under foreign control. The serf may improve the land, build well and even brilliantly upon it, but the title to the land is held elsewhere. Improvement and advancement of this land raises its value without changing the ownership. Therefore, what we see as autonomy right now is basically advancement at the application layer, which is resting on a borrowed substrate.

The same relation of access without ownership can also be viewed through labour. A 2023 World Bank report highlighted that hundreds of millions of people worldwide engage in microwork, such as image tagging, annotation, and classification tasks to train models, with most of these workers having a bachelor’s degree. India stands out to be one of the largest suppliers of this work. And while microwork offers another point of entry into the global AI value chain, it is poorly paid and eventually vulnerable to automation, making it quite a fragile foundation for development. Structurally, the risk is that while countries like India supply inexpensive human labour to train models, accessing these models requires foreign interfaces. It is the inference economy concern in a different form. To simplify – if the chips, the foundational models, and

the high-margin inference layer remain largely foreign or external, the domestic economy is only able to capture lower-margin segments, while durable rent settles upstream and stays there.

As mentioned at the start of this essay, sovereignty in its literal sense means the absence of dependence on an external will. And a polity that leases the substrate of its own intelligence production manages to secure the form of autonomy while conceding a part of its substance. Hence, the rent that Varoufakis describes presents itself in the contemporary era in the form of access, affordability, and even partnership. This naturalises a relation that the language of sovereignty is meant to deny.

None of the above claims binds India to a permanent condition of dependence. The India AI Mission is still early in its life and the hardware layer is itself under immense competitive pressure. Domestic fabrication, however, is a decade-long project, so India is unlikely to close the gap with the US or China in the near term. Studies on sovereign AI hold the same view, highlighting that full-stack sovereignty is structurally infeasible for almost any country, and that the effort should be put towards managed interdependence through partnerships. If partnership is the form the rent now takes, India’s task is to negotiate its terms. A base for the same already exists on paper. For instance, in 2025, India and the US launched the Transforming the Relationship Utilizing Strategic Technology (TRUST) initiative, promising a roadmap for US AI infrastructure in India by the end of the year. However, officials still describe this as under development in mid-2026. In February 2026, the two countries signed the AI Opportunity Partnership, laying out commitments for access to compute, although none of it is binding. The risk of relying on access without a formal partnership or binding terms became a lot more visible in June this year, when the US Department of Commerce directed Anthropic to suspend access to its most advanced AI models for all foreign nationals. The order was lifted after 18 days.

Now, looking at the overall picture, India’s core interest ideally lies in a binding compute agreement and guarantees of chip and model allocations over several years. This can be paired with a CLOUD Act executive agreement permitting legal access on both sides. The focus for India then can shift to the application layer, where digital public infrastructure and deployment at population scale are the strengths India brings to the partnership. While such an agreement would not technically transfer the ownership of the substrate, at the very least, it would put the terms of India’s access in writing with grounds to contest a sudden change to them.

Authors:

Senior Research Associate