Mr. Tanishq Soni
IXth Semester B.A. LL.B (Hons)
Department of Legal Studies and Research, Barkatullah Vishwavidyalaya
The Urgency of Now
In the twenty-first century, the definition of sovereignty has shifted fundamentally, since it is no longer determined solely by borders or standing armies, but by FLOPs, or Floating Point Operations. We are currently operating within a critical historical window and experts estimate we have perhaps no more than thirty-six months before the global AI supply chain hardens into a permanent hierarchy. The world is rapidly bifurcating into two distinct classes of nations in the present regard. There are the "Compute Haves," i.e., the nations that own the infrastructure to train the intelligent algorithms (including AIs), and the "Compute Have-Nots," i.e., the nations reduced to the status of API consumers. The capacity is determined by a term, namely, ‘compute hours’. Simply put, it is the standard currency for renting supercomputing power. To say, instead of buying a $30,000 physical GPU, researchers and developers rent one from a cloud provider (like AWS, Google Cloud, or Lambda Labs) and pay only for the time they use it.
These consumer nations pay rent to access intelligence they cannot create, run their digital lives on servers they do not own, and are subject to algorithmic rules they did not write. For a glimpse, the inconspicuous absence of the triumvirate as mentioned above, constitutes as the rationale for this article.
It is also pertinent to quote that if India does not establish firm control over its compute infrastructure within this period, we risk becoming a digital colony. The statement is a fact begotten in the womb of economic certainty. The cost of inaction extends far beyond temporary economic stagnation, and it is probable that such an inaction can represent permanent marginalization of a billion people from the defining technology of our age.
What is the Core Conflict?
To understand why Compute Governance is the solution, it is put forth that we must first objectively define the problem we face. We are witnessing the emergence of a "Compute Divide." It is exclusively distinct from the "digital divide," of the 2010s, in terms of not merely the lexology, but the concept. That previous crisis was about connectivity and hardware availability, while the new divide is a crisis of capacity. It asks who has the GPU hours and processing power to train a model. For the sake of brevity, 1 GPU hour means 1 specific GPU running for one hour. Consequently, if you rent a server with 1 GPU and run a task for 2 hours, you have consumed 2 GPU Hours.
Now, since 2012, the year following which the AI research field turned un-democratic and was left to the free market, the raw processing power required to train and deploy AI naturally concentrates in the hands of a few global hyperscalers. Entities like AWS, Google Cloud, and Azure dominate this space due to the massive capital expenditure required to build modern data centers. The barrier to entry for training a frontier AI model has risen from thousands of dollars to hundreds of millions. This pricing structure systematically excludes the marginalized, for instance, the rural startups, researchers, and other enthusiasts, working in non-English languages, and public sector innovators are priced out of this innovation economy.
Demarginalization requires the state to shift its perspective. We must move from viewing compute as a private commodity to managing them as a public utility and necessity. To apprehend this transformation, an analogy of three concentric circles is roughly presented as follows.
Circle 1 - The Economic Necessity, or, The Outer Layer
The argument must begin with the "Why." We must recognize that hardware without access is merely silicon in a box.
The India Semiconductor Mission (ISM) has been a monumental step in our national strategy. With an outlay of ₹76,000 crore, it has successfully incentivized the physical production of chips. As of late 2025, we are seeing the fruits of this labor with the unveiling of indigenous developments and the groundbreaking of Silicon Carbide facilities in states like Odisha, and India is proving it can build the engine.
However, there is a substantial risk of market failure regarding distribution because producing a chip is only the first step. Deploying it accessibly comes second. Consider a rural agri-tech startup in Vidarbha. If they cannot afford the H100 GPU hours to train a local weather prediction model because the capacity is hoarded by high-frequency trading firms or ad-tech giants, the infrastructure has failed its public purpose. Right at this point it becomes clearer as to where the necessity of Digital Demarginalization arises. It must be foreseen as an active process of deconstructing the power structures that push communities to the periphery. Precisely, these structures can be listed as linguistic, economic, and algorithmic. Linguistic marginalization occurs when compute is expensive. In a high-cost environment, resources are only used to train models on English data because it is abundant and cheaper to process. Low-resource languages like Maithili or Santali become "digitally invisible" because they are not profitable to compute. Economic marginalization follows a similar path. If AI training costs ten million dollars, only unicorns can innovate, and the Small and Medium Enterprise sector is wiped out. Simply put (and empirically backed), the innovation-narrative is pushed forth by the fewer actors in exponential magnitude, and hence, it becomes the deciding factor of the wave.
The "Necessity" circle demands that the value created by our national missions flows to the edges of the grid rather than remaining trapped in urban hubs. It is obvious that we cannot build a sovereign Digital India on the rented foreign lands.
Circle 2 -The Mechanism, or, The Middle Layer
It is a settled principle in multiple disciplines, most notably, the law - we cannot govern what we do not have. Which brings us to the "Mechanism" layer, or, the creation of a sovereign compute stack.
The objective is currently being pushed through the IndiaAI Mission; with a budget of
₹10,372 crore, the mission has explicitly targeted the procurement of over 10,000 GPUs, which is intended to scale toward a target of roughly 38,000 GPUs to build a national AI infrastructure. By mid-2025, reports indicated that significant progress had been made under the mission's tender process to install these units.
The mechanism becomes pivotal because it breaks the monopoly of the hyperscalers. By creating a public-private partnership (PPP) model, and empanelling providers like Yotta and Jio, the government is approaching the thresholds of creating a public cloud option.
However, an objective analysis must also address the physical constraints associated with the PPP synergy and the governmental efforts via public policies. Compute is energy. Training a single large language model (LLM) can consume as much electricity as one hundred homes use in a year. Here, India must look to global precedents, because if not, then the entire premise is obligated to succumb. China’s "Eastern Data, Western Computing" plan is physically relocating data centers from the crowded, energy-poor east coast to the wind-and-solar-rich western provinces.
India’s mechanism must similarly integrate its energy policy with its compute policy. Our High-Performance Computing (HPC) facilities, such as AIRAWAT and Param Sankul, must be linked to our renewable energy grid. Fascinatingly, if we power our AI revolution on coal, we are merely trading digital marginalization for environmental marginalization. Climate change disproportionately affects the poor, and it is no hidden secret that a truly sovereign mechanism is one that is sustainable.
Circle 3 - The Solution of Governance, or The Inner Layer
Only the inner-most of these three concentric circles deals with the aspect on which the topical premise of this article rests. This is the Governance layer, where the strategy is executed. Considerably, and relatively, it must be noted that building the GPUs is a simpler affair. Deciding who gets to use them, and for what purpose, is where the challenge lies.
In order to strictly track our goals, India must implement scientifically tempered Compute Governance. The world offers a lot of precedents. We can adapt global frameworks, but with a goal of improvement as well. The US Executive Order 14110 focuses heavily on national security. It sets a reporting threshold for models trained on more than 1026 FLOPs. Similarly, the EU AI Act focuses on safety and sets a "systemic risk" threshold at 1025 FLOPs. Essentially, what this means is that the limit set by, for instance, the US Order, practically restricts building LLMs or other models which are five to ten times more powerful than existing (OpenAI’s) GPT-4. Such a deployment would necessitate the deployer to seek permission from the government.
India’s governance model must be different. It must focus on Development and Equity. It should rest on three specific pillars, i.e., to say, visibility, whereby, the state must know what is being built on its infrastructure, such that we know if the compute capacity in the country is being used to train a predatory lending algorithm or a tuberculosis diagnostic tool; secondly,
allocation, whereby, it is ensured that compute is subsidized for public good, and the allocation of the compute resources is not left largely with the markets; and lastly, enforcement, which in essence would mean that access to sovereign compute must be a privilege, is conditional on responsibility. If an entity wants to use the subsidized public cloud, they must adhere to the Digital Personal Data Protection Act or any other legislation in force for the time-being.
Conclusion
We are witnessing the final evolution of the Indian digital state. The journey can be mapped in three distinct phases.
Phase one was Connectivity. In the last decade, we focused on fiber optics and 4G networks. We successfully connected the village to the network. Phase two was Platforms. We built the Digital Public Infrastructure Systems, such as the UPI, Aadhaar, and ONDC were created to move value and identity across those roads. Phase three is Compute Governance, or, the missing layer. Governance allows the state to deny compute to bad actors. By controlling the sensitive point of AI-powered nuisance and anomalies, which is the compute itself, the state can enforce safety standards that are impossible to enforce through legislation alone.