Soham Gupta, GNLU, 1st year
In today’s time, every country in the world is rushing to build its regulations around Artificial Intelligence. Yet, on a closer examination of the AI governance structures, it’s found that all of these laws aim to regulate the downstream product of AI, sticking to the topics of transparency, bias, and content risks. They fail to address one of the most major building blocks of AI models, compute. Compute refers to the high-performance hardware required to build frontier AI models, specifically referring to the semiconductor chips and GPUs. Without these, building AI capabilities for a nation is seemingly impossible. Thus, as global markets converge, NVIDIA has become the near-sole provider of this compute, hinting at glaring competition-law challenges in digital markets. As we look at the markets, legally or traditionally, the formation of a cartel, an association of manufacturers or suppliers to maintain prices at a high level and restrict competition, cannot be seen. But, on analysis, the whole market is seen to be becoming structurally cartelised, made by technological lock-in, supply-chain chokepoints, and geopolitical reinforcement. The existing anti-trust laws, particularly India’s Competition Act, the EU’s Article 102 TFEU, or the U.S. Sherman Act, were not built to handle bottlenecks, which are essential facilities located outside the country’s jurisdiction. It’s thus shown to be imperative that this emerging problem is quickly dealt with, lest the country be excluded from the next technological revolution.
I. Relevant Market Definition: Using EC, U.S., and Indian Standards
As we consider the development of a cartel-like market in the sector, we must first define what constitutes the very market that we are analysing. Market has been defined multiple times in different jurisdictions, be it by the Competition Act, 2002 (Sections 2(r), 2(s), 2(t)), or the EC’s reasoning in United Brands v Commission (Case 27/76), or the U.S. Supreme Court in Brown Shoe Co. v. United States (370 U.S. 294). Each of these emphasised that the market is the sphere in which these products compete, and this is further filtered by “practical indicia” such as industry recognition and product characteristics.
As we look at the said competitors of the high-performance GPUs used for frontier model development, we must rule them out. These are the CPUs (due to a very high performance gap in training AI models), consumer GPUs (due to a lack of essential architecture, like tensor cores, or memory bandwidth), and cloud compute (it being a downstream product of the hardware rather than a competitor). Thus, we form the relevant market to be the high-performance AI accelerators used for training and inference of frontier-scale ML models. In this space, NVIDIA can be seen to be holding a share of more than 80% of this sphere, marking it as dominant under Section 4 of the Competition Act, Article 102 TFEU (dominance by market share), and Sherman Act §2 (“possession of monopoly power”). Further, while this is a global market in terms of production or supply, the EC in Hilti v Commission (Case T-30/89) has shown that the harms of such a market are completely present at a local level, making it a relevant problem for the competition laws of the different jurisdictions. Countries must look into this to avoid higher prices, reduced availability, delayed access, and finally, exclusion from frontier AI development.
II. CUDA as an Essential Facility: Applying Aspen Skiing, Bronner, and Indian Doctrine
Compute Unified Device Architecture, or CUDA, is a parallel computing platform and programming model created by NVIDIA that allows developers to use NVIDIA’s GPUs for general-purpose processing. This structure has served as one of the main drivers in the market, becoming a major barrier for any competitors to be viable, making it an essential facility.
In competition law, the essential facility doctrine can be derived by Aspen Skiing Co. v. Aspen Highlands (472 U.S. 585) (which showed that refusal to cooperate with a competitor can be monopolisation), the EC case Oscar Bronner (Case C-7/97) (which showed that if an indispensable facility is impossible to duplicate, its access may be required), and for India, Schneider Electric v CCI (2016) (which held that denial of access to critical infrastructure can constitute abuse).
As we look at CUDA and its position in the market, it becomes apparent that it’s an essential facility that is blocked off to competitors. First, virtually all ML frameworks, be it PyTorch, TensorFlow, or Jax, have been developed on CUDA kernels. Thus, to develop in these frameworks, the developers have to use NVIDIA GPUs. This makes this facility indispensable, and its exclusivity serves as a point of elimination for competition. Further, while blocking the development of rival ecosystems (ROCm, oneAPI), the CUDA ecosystem is the result of years of development and billions of dollars, making it impossible to duplicate. The denial of open access for competing hardware to this architecture leads to constructive denial of market access under Section 4(2)(c) of the Competition Act, 2002. A close parallel of such a situation can be drawn from the Google Android case in 2018, where CCI held that mandatory pre-installation and exclusionary design prevented rival app stores from gaining market access, regardless of any explicit refusal (self-preferencing). CUDA does the same for GPU competitors.
III. Vertical Foreclosure: Lessons from Microsoft, Google Shopping, and Intel
Time and time again, competition law has stated that vertical integration can become abusive if it shields dominance or forecloses rivals. Vertical integration refers to one company making multiple products required at different stages of a final product’s development or usage cycle, particularly looking to keep their products working seamlessly with each other, and becoming a one-stop solution for that ecosystem’s development. In the past, we have seen the EU Microsoft case, where the tying of Windows Media Player to the Windows operating system became a restriction to rival media player software, or the Google Search case, where self-preferencing was marked to harm competition in vertical markets. Both of these cases are reflected in how NVIDIA’s ecosystem is structured, handling Hardware, Networking, Software, and Clusters together. NVIDIA controls the chips (H100/H200), networking (Infiniband, NVLink), drivers (CUDA libraries, cuDNN, TensorRT), and the orchestration tools and cluster design guidelines. This results in a full-stack closure where new entrants cannot become cluster compatible, hyper-scalers tend to avoid dual-stack optimisation due to integration costs, and developers default to CUDA frameworks due to their deep embedding in the systems. Thus, vertical foreclosure of NVIDIA further blocks the markets and is becoming a leading concern for monopolisation.
IV. Export Controls: Geopolitical Foreclosure as an Antitrust Factor
Another major factor contributing to a global competition law emergency is the lack of any global regulation in this sector. This has led to the whole market operating at the whims and fancies of a private entity, along with some control by export laws. An excellent demonstration of this had been put forth by the United States’ Bureau of Industry and Security, the statutory body controlling export of compute. The BIS, in its recent strategies, released rulings that aim to block access to high-end GPUs (A100s, H100s, H200s), particularly for Chinese entities. While this is a sovereign act, and prima facie cannot be considered in anti-trust violation, the EC case Continental Can (Case 6/72) held that the market effects of such changes need to be considered.
This action prevents the Chinese and Global South competitors from accessing essential inputs. Such a development reinforces NVIDIA’s position in the market, and forces multiple nations to rely on downgraded chips, with the effect particularly being exacerbating scarcity for countries like India. Such effects tie compute access to not just market dynamics, but geopolitical alignment. This case draws parallels with the case of TeliaSonera (C-52/09), where interactions between public regulation and private dominance produced exclusionary outcomes.
V. Structural Cartelisation and India’s Harm
As we analyse the various aspects of the existing GPU market, there exists a blockage to essential facilities, tight and perhaps abusive vertical integration, and even reliance on geopolitical dynamics. Such a market structure lacks agreement or concerted practice, refuting the claims of forming a cartel under Section 3(3). Yet, the structure that has been formed is an exhibit of cartel-like outcomes. The first observation is scarcity pricing. The prices of H100 GPUs have been observed to exceed many multiples of their production cost, falling neatly into a pattern aligning with United Brands’ test for excessive pricing. Further, the supply of these chips remains structurally restricted due to bottlenecks at fabrication and tooling stages, particularly TSMC’s limited advanced-node capacity and ASML’s monopoly over EUV lithography. These constraints, along with NVIDIA’s allocation priorities, create an environment in which artificial scarcity becomes a persistent market condition rather than an accidental shortage. The predictable consequence is that large hyper-scalers secure assured supply, while smaller jurisdictions and emerging markets face prolonged delays and inflated prices.
All of these cement the requirement for urgent addressal of this emerging problem. India must focus on this issue as it becomes a domestic harm, affecting academia, startups, cloud providers, and national AI capability. The effects of this cartel-like market are already prevalent, as excessive pricing becomes evident. The Indian cloud prices for H100 instances are seen to be significantly higher than U.S./EU prices. Further, due to these higher costs and lower availability of compute in India, it has started to limit technical development in violation of Section 4(2)(a) of the Competition Act. Many Indian researchers and academia cannot train frontier models due to the unavailability of resources. This scarcity of the resource, discriminatory allocation patterns, and prohibitive pricing together indicate market denial for India, forming similar patterns as seen in the case of Fast Track Call Cab v ANI.
All of these, if not addressed, hint at a new form of anti-trust problem. They show us the possibility of foreclosure of an entire nation from a new and general-purpose technology.
Conclusion: India Cannot Ignore the GPU Competition Crisis
As we conclude, we have observed that NVIDIA doesn’t form a cartel in the traditional sense. Yet, the current structure of the GPU market shows scarcity, dependency, pricing power, structural exclusion, and geopolitical reinforcement, all behaviours seen in cartels. Further, India stands at a strategic vulnerability, with a long way to go for domestic AI resources to develop. AI is said to be the defining technology of the century, with compute being its essential facility. In the present arrangement, India can be locked out of this facility. It’s time we recognised this issue, and worked to make a global market that treats compute as a distinct regulated asset, and makes sure that it doesn’t become a competition law emergency.