Soham Gupta is a first-year student at Gujarat National Law University, Gandhinagar
China recently imposed a regulation that bans the use of foreign AI chips by state-funded development institutes. The United States put up export laws that regulate the export of high-end GPUs such as NVIDIA’s A100, H100, specifically limiting access to Chinese entities. Further, the European Union released its Chips Act, promoting the development of domestic manufacturing units of high-end semiconductor chips. Around the globe, geopolitical powers are in a race to make the most effective AI governance regulation, and as this continues, they have started to recognise the existence of one of the major parts in an AI model’s development, compute. Compute is the hardware required to build complex artificial intelligence models, particularly including the high-end GPUs and powerful semiconductors.
Artificial Intelligence governance has become a widely debated topic in society, with multiple ideas on how it should be regulated, and many countries stepping forward to release regulations governing this sector. Yet, global governance of such models is increasingly being reliant on the model’s development, style, and capacity, and has overlooked the rising complexities of controlling advanced semiconductor hardware. Over the past 2 years, compute has become quite a site of geopolitics and intervention by industrial policy. These developments illustrate how access to computing is shaping the global development trajectory of frontier AI systems. Further, digital sovereignty is quickly becoming a concern among countries, revealing a fatal pattern in the flow of important AI infrastructure in the world. These developments all reveal one common thing: they show that the countries are coming around to regulating compute, but quite often, this is being done indirectly and without major public debate.
Despite all this, current research on AI governance majorly focuses on the outputs of AI models, talking about model safety, transparency, biases, or harmful content. All of these trends assume that Artificial Intelligence, as a resource, continues to be openly available to all. Yet, this can be strongly refuted, as the world shows us trends where compute is becoming scarce and concentrated in a few major entities. This leads to the development of a new political angle, and needs to be analysed. This blog thus comes to map how compute is quietly, yet definitely becoming an avenue for global regulation, and outlines the gross need of treating compute as an asset requiring its own rules and regulations.
1. The Hidden Assumption in AI Governance
Many of the recent regulations in the sphere of AI policy hold the implicit assumption that resources for the development of AI, i.e. compute, are functionally unlimited for the entities under the regulations. These codes often focus on the behaviour of the model, ensuring rules around how the models are trained, what data is used, how accountability is measured for algorithms, how the content is being moderated, etc. Yet, all the development of AI relies on one foundational input that is, the availability of accelerated compute. This resource is seen to be quite unevenly distributed across countries, institutions, and firms, with a small number of cloud providers and semiconductor manufacturers determining who can train frontier-scale models, the cost they need to pay to train such models, and what are the conditions for these models to exist.
Such hidden restrictions matter, for if compute access is structurally limited to some, then the regulations placed on AI output risk bolstering the inequalities present in the system. It can be easily inferred that the Global South economies, without the infrastructure to manufacture semiconductors domestically, would remain highly dependent on the terms and policies that advanced manufacturing states set for them. Recognising this emerging divide would reshape how the world is thinking about AI policy. Any robust AI governance policy must thus integrate principles that talk about how one can access, allocate, and monitor compute.
2. Why Compute = Public Infrastructure
Advanced machine learning relies on certain laws that determine its scale by connecting the dots between model performance and increasing quantities of parameters, data, and compute. Thus, the ability for people to innovate, conduct scientific research, and participate in the evolving technological ecosystem is heavily dependent on their access to sufficiently powerful hardware. Without this, research institutions cannot replicate the frontier results or meaningfully evaluate the risks in any field. These emerging barriers to compute access have already had measurable effects, with studies indicating that many universities outside the United States and Europe cannot afford the GPU clusters, leading to high dependence on cloud infrastructure that may not be in their budget or may be blocked by export-control restrictions.
It can be useful to draw certain analogies for the situation with other existing infrastructure. Some of the closest parallels to this situation are the Telecommunications spectrum, electricity grids, and broadband networks, all of which are similar to compute, as they have high fixed costs, exhibit natural monopoly tendencies, and lead to significant public interest benefits with more access. As these demands of training and evaluating large models increase, compute access would for sure determine which actors can contribute to scientific knowledge, safety standards, and technological innovation. Further, as innovation in every field becomes more and more dependent on AI models, it is imperative that people have access to develop sovereign digital resources, lest they become forced to work on the policies of another. Thus, treating compute as public infrastructure is not only a metaphorical claim, but something that can be seen to shape reality and, if ignored, change world orders.
3. The Invisible Gatekeepers and a New Geopolitical Factor
Currently, access to compute is governed by a combination of corporate control, industrial chokepoints, and export-control regimes rather than public law. This can be seen to be shaping reality by factors such as NVIDIA maintaining an overwhelming share of 80-90 per cent of the global market for GPUs capable of training AI, or the Taiwan Semiconductor Manufacturing Company (TSMC) controlling nearly all high-performance chips below 7nm. This gives these private actors a very influential position in which they decide who gets access to these resources, at what price, and at which location. Further, some influence is held by export-control authorities, as seen by the United States Bureau of Industry and Security (BIS) release of export-control rules from 2022-2024, restricting the sale of high-end GPUs and chip-making equipment to Chinese entities. Cloud providers form another layer of informal governance, with their pricing structures, access policies, and queueing mechanisms giving them the power to determine who can access compute.
These actors thus exercise a regulatory power that will influence the global technological capabilities, and yet, is driven by geopolitical negotiations, supply-chain considerations, and corporate policies. Such agendas operate in a regulatory vacuum, bolstering the need for formal compute governance. These are the invisible gatekeepers to compute.
The United States’ blocking of compute to China led Chinese firms to be compelled to redesign models to operate on lower-end chips, and it has largely affected their national AI ambitions. As we look at the effects of an export policy for compute resources, it becomes clear that control over these resources has become a defining feature of geopolitical strategy. This is further supported by the rising concerns in Europe to develop domestic manufacturing and supply of high-end semiconductors, as already pushed for with the new Chips Act. As these high-end chips become instruments used to affect foreign nations, international law must pin down governance of compute, considering it a cross-border resource. The political significance of compute is becoming inseparable from its legal and economic dimensions.
4. Why Law Must Step In
As the compute resource supply operates in a regulatory vacuum, it leads to deterioration of the market structures to the point of market failure. One of its first major effects is the creation of a monopoly or oligopoly. This would lead to the market becoming a product of the corporate whims and fancies, leading to high prices and restriction of access to smaller entities. Continuing this, the lack of accountability as private control prevails over public-interest infrastructure undermines the democratic principles of society. Third, as countries rely on export controls as the primary governance mechanism, it leads to fragmented and politically contingent outcomes.
As we examine this, it becomes clear that this field is quite similar to that of telecommunications spectrum allocation, electricity market regulation, and nuclear material controls, as each of these involves licensing, monitoring, and public-interest obligations to be heavily considered. Compute resources should no longer be treated as any other free-market item and must become a distinct regulated asset, for it is compute that will decide who innovates, who researches, and thus who creates.
5. The India Question
As we discuss the wider problems in the global AI governance regimes and the importance of compute, we must zoom in on India and aim for policies that bolster its digital sovereignty. India has multiple initiatives that talk about compute, these initiatives, like the India AI Mission, Semicon India Mission, Digital India Stack Expansion, and Make in India for electronics manufacturing, all touch upon the development of domestic compute resources, yet they fail to address it directly. These plans fail to take into account the fact that India lacks access to frontier-scale compute, and further, it has no legal or institutional framework to govern compute as a national resource. These concerns, if not addressed, would leave the country at the mercy of foreign compute resources for its scientific and technological advancement. India needs a comprehensive compute governance regulation, and while not exhaustive, the following is a short list of policy suggestions for the country.
- India must establish a National Compute Authority. A statutory body modelled around TRAI or PNGRB, which has the power to license, regulate, and allocate high-performance compute, as well as monitor the price, auditing capacity, and ensure fair access to compute for academia and startups. Compute must become a national resource.
- The country must create and hold a public GPU infrastructure, which is publicly funded and accessible to important research teams. This National Compute Grid must be built across institutions of importance such as IITs, IISc, and select state universities, allowing access to training and development of human as well as technological resources.
- The country must have a mandatory disclosure requirement for any model trained above a set compute threshold (eg, FLOP-based criteria), to disclose training infrastructure and safety measures put forth.
- The country must push forward to mitigate dependence on a few private entities or nations by using semiconductor-linked FTAs, trusted supplier frameworks, and joint GPU-procurement agreements. Further, special efforts must be put in to establish domestic facilities for the creation and maintenance of compute resources.
These policies would help the country to shift from being a passive consumer of computing resources to being a nation with digital sovereignty, which would be essential for any nation to survive in the world order.
6. Conclusion
Concluding, we can say that compute is emerging as the primary determinant of global AI capability. The decisions that revolve around compute access are currently largely shaped by private concern, leading to a gross regulatory vacuum that would shape commercial innovation as well as the distribution of scientific knowledge, public oversight, and geopolitical power. Thus, compute governance has become a practical and urgent challenge that shapes who builds, evaluates, and governs AI systems. Further, India is seen to be lagging in governing compute, with current policies only glancing at the prudent topic. India must undertake certain policy implementations and changes to realise its ambition to become a first-world country. It is high time that the world and India realise the concept of digital sovereignty and recognise compute access as a prudent part of it.