I. Introduction

Graphics Processing Units (GPUs) enable seamless parallel processing of complex workloads and have today become the essential engine for everything from 3D rendering to deep learning. GPU Computing is a service that provides powerful GPU resources in the cloud. This allows businesses and developers to rent GPU power on demand instead of purchasing or managing expensive physical GPU units. The GPU compute market has now become the backbone of the global tech economy. It is the “digital oil” that determines which nations will lead the next industrial revolution. 

India’s ability to become a key stakeholder in the global Artificial Intelligence (AI) economy is determined not by algorithms, talent, or data, but by its ability to compute. Without reliable and affordable GPU capacity, India cannot train competitive models, support its rapidly expanding AI ecosystem, or realise the ambitions embedded in the IndiaAI Mission. This paper, therefore, addresses a foundational question for India’s technological future. Does the concentrated control of GPU infrastructure by a handful of global firms amount to a competition bottleneck serious enough to warrant regulatory intervention under the Essential Facilities Doctrine (EFD)?

The urgency of this inquiry lies in the widening gap between India’s AI aspirations and its ‘compute’ reality. Even as the country boasts world-class engineering talent and a vibrant startup ecosystem, its ability to train modern AI systems is constrained by the severe shortage and concentrated control of GPUs and related accelerators. Although AI models are software, their development depends on scarce, capital-intensive hardware. NVIDIA now dominates the upstream chip market, while practical access to these chips is controlled by three hyperscalers, i.e., Amazon Web Services, Microsoft Azure, and Google Cloud. The allocation decisions determine which firms can innovate, scale, or even enter the market. For Indian startups and research institutions, access is often prohibitively expensive or simply unavailable, creating a structural dependency that no amount of software talent can overcome.

This supply scarcity has evolved into a competition problem. If compute is the “oil” of the AI age, GPU clusters and cloud infrastructure are the “pipelines” controlled by a small number of foreign firms with the power to decide who gets to build the next generation of AI systems. This raises a critical question for the Competition Commission of India (CCI - Should high-performance GPU compute be treated as an essential facility whose restricted access amounts to a denial of market access under Section 4 of the Competition Act?

This paper argues that India should treat EFD as a diagnostic tool but rely on the Digital Competition Bill’s ex-ante duties as the primary mechanism to regulate GPU access. This paper will, firstly, outline the legal foundations and evolution of the Essential Facilities Doctrine and its implicit adoption in Indian competition law; secondly,  apply the four-prong EFD test to the contemporary GPU compute ecosystem, drawing on market structure and access dynamics; and thirdly, evaluate normatively whether EFD is the appropriate regulatory tool in this sector, contrasting it with the emerging ex-ante framework under the Digital Competition Bill. Finally, it will propose a hybrid regulatory model and conclude with implications for India’s long-term AI competitiveness.

II. Essential Facilities Doctrine: Legal Framework

To determine whether India can compel dominant technology firms to provide fair access to GPU infrastructure, it is necessary to examine the legal boundaries of the Essential Facilities Doctrine (EFD). This principle originated from early U.S. antitrust jurisprudence and has since influenced global competition regulations. Articulated in cases involving railroads, it has since developed to state that a monopolist in control of a facility essential to their competitors, which cannot reasonably be duplicated, must provide access to that facility on reasonable terms. Although traditionally applied to physical infrastructure, the logic of the EFD has increasingly been expanded  to include digital bottlenecks whose control determines participation in downstream markets.

Although Indian competition law does not explicitly codify the Essential Facilities Doctrine, the Competition Commission of India (CCI) has gradually absorbed its animating principles into Section 4 of the Competition Act. Articulated particularly in Section 4(2)(c), it prohibits dominant enterprises from engaging in conduct that results in a “denial of market access in any manner.” Indian jurisprudence illustrates this evolution. In Competition Commission of India v. Fast Way Transmission Pvt. Ltd., the Supreme Court recognised that control over cable transmission infrastructure placed a “special responsibility” on the dominant undertaking. Denying access to this indispensable infrastructure impaired the ability of downstream broadcasters to compete. This thereby mirrors the essential facilities concern of a dominant intermediary foreclosing market access by controlling a non-replicable facility. Similarly, in Matrimony.com Ltd. v. Google LLC, CCI held that Google’s dominance in search created a critical gateway for businesses seeking visibility on the internet. By leveraging its upstream position to impose restrictive conditions and distort downstream markets, Google effectively undermined competitive neutrality. Though neither case formally invokes the EFD, it recognises that exclusion from a critical platform can itself constitute abuse, even absent explicit refusal to deal.

A further nuance to the doctrine emerges from the U.S. Supreme Court’s judgment in Associated Press v. United States (1945), which held that when a consortium of competitors jointly controls an essential input, their coordinated restrictions on access may constitute an unlawful restraint of trade. The Court reasoned that collective control over an indispensable resource can produce exclusionary effects equivalent to unilateral dominance, particularly where membership or access decisions determine whether new entrants can effectively compete. In the contemporary GPU compute market, while NVIDIA (with 88% of the global market), dominates the upstream supply of AI-grade chips, the practical point of access for Indian startups lies with the cloud hyperscalers Amazon Web Services, Microsoft Azure, and Google Cloud. These firms not only compete horizontally in cloud services but also collectively shape global availability, prioritisation, and pricing of GPU clusters. Being the primary providers of infrastructure, they create limited options for startups, hence allowing them to control prices. By making users dependent on their interfaces while hoarding scarce compute for themselves and their elite partners, hyperscalers create a structural exclusion. This makes it easier for hyperscalers to gatekeep technology and produce systems that resemble collective control. Although Indian competition law does not formally recognise “collective dominance,” the CCI has repeatedly acknowledged in its market studies on telecom, e-commerce, and, more recently, AI that concentrated platform markets can produce de facto joint control over essential inputs, particularly where switching costs and technological lock-ins limit effective alternatives.

Taken together, these doctrinal strands suggest that Indian competition law, while not explicitly adopting the Essential Facilities Doctrine, already embodies its central commitments. It prevents enterprises from controlling non-duplicable infrastructure, whether individually or through structurally coordinated ecosystems, or distorting downstream competition by restricting access. This framework sets the stage for assessing whether the GPU compute stack i.e., specialised chips, CUDA-dependent software, and hyperscaler-controlled cloud clusters, functions as an essential facility under Section 4’s prohibition on denial of market access. 

III. The Four-Prong Test and Digital Infrastructure

A crucial step in determining whether the above infrastructure may be treated as an essential facility in India is the application of the four-prong test articulated in MCI Communications Corp. v AT&T, the most influential formulation of the Essential Facilities Doctrine. Under this test, an authority must determine: (1) control of the essential facility by a monopolist; (2) a competitor's inability practically or reasonably to duplicate the essential facility; (3) the denial of the use of the facility to a competitor; and (4) the feasibility of providing the facility.

In the GPU context, control may lie with NVIDIA, which dominates not only the supply of high-performance chips but also the proprietary CUDA software stack that makes these chips uniquely functional, or with cloud hyperscalers such as AWS, Azure, and Google Cloud, who exercise practical control through their allocation of scarce GPU clusters. Duplication of such a facility is largely unrealistic. Although data centres or alternative accelerators can theoretically be built, Indian startups cannot replicate the necessary capital expenditure, supply-chain access, or software compatibility, hence making them dependent on hyperscalers. Contrary to the misconception that digital infrastructure is easily replicable, the deep physical, organisational, and ecosystem lock-ins which range from fabrication capacity to middleware integration make meaningful duplication implausible. The third element, denial of access, increasingly takes the form of constructive rather than explicit refusal, as access may be priced prohibitively, provided on restrictive terms, or made unreliable through prioritisation of internal research teams or strategic partners, effectively excluding smaller rivals despite nominal availability. Finally, feasibility of providing access is rarely in genuine dispute. While firms may justify restrictions by citing supply shortages or commercial considerations, the core question for competition law is whether withholding or rationing access to GPU compute distorts downstream innovation, investment, and competitive conditions in markets that depend on such infrastructure.

Comparative jurisprudence offers two contrasting paths for India. The U.S. approach, exemplified by Trinko, warns against imposing sharing obligations on firms whose investments created the infrastructure, implying that NVIDIA’s dominance or hyperscalers’ preferential allocation of scarce GPUs may be commercially rational rather than exclusionary. The EU’s more effects-based framework, reflected in Oscar Bronner and the Microsoft interoperability cases, treats refusal to supply as abusive when it forecloses viable competition. This logic can be mapped closely onto India’s AI ecosystem, where CUDA lock-in, supply-chain constraints, and hyperscaler prioritisation may impede domestic model development. India’s policy direction, including debates around the Digital Competition Bill, signals movement toward this European stance. On such a reading, the Competition Act already gives the CCI sufficient flexibility to treat GPU compute as an essential facility where scarcity, vertical integration, and technical lock-in materially impair downstream innovation.

IV. Market Analysis: GPU Monopolies and Barriers

To assess whether GPU compute meets the economic conditions of an essential facility, it is necessary to examine the structure of the market that supplies and allocates this resource. The facility in question is not merely a semiconductor chip but a vertically integrated stack comprising specialised hardware, proprietary software, and hyperscaler-controlled cloud infrastructure. At the upstream level, NVIDIA exercises overwhelming market power. It controls close to 88% of the data centre GPU market. Yet its dominance stems not only from hardware performance but from the CUDA software ecosystem, which has become the de facto programming standard for modern AI. CUDA is the USP of NVIDIA, developed in-house. Its optimisation across AI libraries, developer tools, and academic workflows creates high switching costs. Moving to AMD or Intel accelerators requires substantial code rewriting, performance compromises, and loss of community support. This software-based moat renders nominal alternatives commercially impractical, reinforcing the non-duplicability prong of the EFD analysis.

At the downstream delivery level, access to GPU compute is mediated almost entirely through three cloud hyperscalers: Amazon Web Services, Microsoft Azure, and Google Cloud. These firms possess preferential access to NVIDIA’s latest chips and determine how scarce GPU clusters are allocated. Their dual role as both suppliers of compute and developers of proprietary frontier AI models produces the risk of foreclosure: they may reserve low-latency, high-capacity clusters for internal teams or strategic partners, leaving startups with inferior “spot” access or substantially higher prices. Denial in this market often takes the form of constructive rather than explicit refusal. Egress fees further entrench lock-in by making data migration prohibitively expensive, effectively trapping firms within a given cloud ecosystem and limiting competitive switching. The result is a bottleneck wherein a handful of global platforms control not only the availability but the competitive conditions under which Indian AI firms can operate.

The Government’s IndiaAI Mission provides additional evidence that GPU compute has acquired characteristics of an essential facility. The Mission’s plan to build a sovereign compute infrastructure of roughly 10,000 GPUs reflects an acknowledgement that the market, left to itself, has failed to provide affordable or timely access. Yet this intervention, while significant, cannot match the scale or technological integration of hyperscaler clusters. Public investment therefore reduces but does not eliminate reliance on private monopolies, underscoring the persistent structural vulnerability. Together, these market realities demonstrate that GPU access satisfies many of the economic conditions associated with essential facilities, particularly non-duplicability, dependence, and the potential for exclusionary allocation practices.

IV. Normative Evaluation: Should India Regulate?

Whether India should regulate GPU compute as an essential facility depends on whether the remedies available under EFDs can address the competitive risks emerging in AI infrastructure markets. Although the doctrine helps explain how control over a non-duplicable resource can harm downstream rivals, its rigid requirements and narrow remedies do not map well onto today’s GPU markets, which evolve quickly, face chronic shortages, and are dominated by vertically integrated firms. The first difficulty lies in the legal requirement of non-duplicability. Courts applying EFD, especially after Bronner and subsequent jurisprudence, demand a demonstration that duplication is not merely expensive but practically impossible.  Yet the AI compute market features emerging, though presently inadequate, alternatives such as CoreWeave and Lambda Labs, and the gradual entry of AMD and Intel in the accelerator space. These developments allow dominant firms to argue that duplication, though burdensome, is technically feasible, weakening the indispensability claim. Yet for Indian startups, such alternatives remain commercially impractical, even if their mere existence undermines the strict “essentiality” threshold required under EFD jurisprudence.

A second structural obstacle is the nature of denial in digital infrastructure markets. Traditional refusal-to-deal turns on explicit exclusion. Today’s compute markets more often involve constructive denial. Rather than refusing access, hyperscalers may price GPUs beyond the reach of startups, reserve top clusters for internal use, or relegate outsiders to unreliable spot capacity. Such conduct can be exclusionary yet fall outside classical doctrine, and courts tend to view high pricing or prioritisation in scarcity as commercially justified rather than anticompetitive. As Trinko illustrates, the judiciary is wary of converting competition law into a mechanism for price regulation, especially in capital-intensive and technologically complex markets. This casts doubt on whether constructive denial of GPU access would meet the third prong of the EFD test.

Timing and remedial feasibility further weaken reliance on EFD. The doctrine is inherently ex-post. The harm must materialise, be investigated, litigated, and adjudicated before a remedy can issue. Yet AI compute markets evolve on 12–18-month cycles, while antitrust cases often take years, meaning any remedy would likely arrive after the market has already shifted. This renders the intervention largely symbolic. Moreover, even if the CCI were to find an abuse, crafting a workable remedy poses formidable challenges. Setting fair GPU prices, supervising allocation across thousands of users, and policing self-preferencing are functions competition authorities are not institutionally equipped to perform. An EFD ruling risks becoming doctrinally elegant but practically inert.

Yet the risks of under-regulation are equally stark. The combined control of hardware (NVIDIA), the software layer (CUDA), and cloud delivery (hyperscalers) creates a vertically integrated bottleneck through which all meaningful AI innovation in India must pass. Technical lock-ins, opaque allocation rules, and costly egress fees generate a structural dependency that may prevent domestic firms from building competitive models, regardless of talent or demand. Allowing such a configuration to go unregulated would risk a severe Type II error (under-enforcement), i.e., allowing exclusionary structures to harden long before overt anticompetitive conduct appears. The normative question, then, is not whether to regulate but how to design a framework that avoids both the rigidity of EFD and the innovation risks of coercive access mandates.

This is where the emerging framework under the Digital Competition Bill offers a more appropriate regulatory pathway. Unlike EFD, which requires proof of indispensability, denial, and feasibility, the DCB recognises that digital markets characterised by network effects, vertical integration, require ex-ante obligations to preserve contestability. Designating major cloud providers as Systemically Important Digital Enterprises the Bill would allow the imposition of non-discrimination duties that prevent preferential GPU access for internal model teams, transparency requirements over allocation criteria and access terms , and interoperability and data-portability mandates, aimed at reducing CUDA-driven lock-in and egress-fee dependency. These obligations do not compel firms to transfer proprietary IP or act as involuntary wholesalers. Instead, they target the mechanisms through which competitive harm actually arises: opacity, prioritisation, and ecosystem lock-ins.   Crucially, ex-ante rules operate prospectively, avoiding the timing mismatch inherent in EFD by deterring harmful allocation practices before they crystallise into long-term structural exclusion.

Hence, hybrid model is normatively preferable. The DCB’s ex-ante duties should provide the main guardrails for fair access, with EFD available as a limited backstop in cases of clear exclusion. This lets India use EFD’s conceptual clarity without relying on it as the primary remedy. Combined with continued public investment through the IndiaAI Mission, this structure enhances compute sovereignty and limits dependence on private hyperscalers. The result is a regulatory framework that maintains innovation incentives while preventing long-term structural foreclosure.

V. Conclusion

India’s AI ambitions ultimately depend on securing fair access to high-performance compute, a resource now constrained by the concentrated control of GPUs, CUDA-dependent software, and hyperscaler cloud infrastructure. While the Essential Facilities Doctrine helps describe the competitive risks created by this configuration, its strict requirements, especially proof of non-duplicability and explicit refusal make it ill-suited to fast, supply-constrained AI markets. Relying on EFD alone therefore risks under-enforcement, particularly where exclusion occurs through pricing, prioritisation, and technical lock-ins rather than overt denial.

A more suitable approach lies in a hybrid model. The Digital Competition Bill’s ex-ante obligations transparency, non-discrimination, interoperability, and limits on self-preferencing are better equipped to address structural bottlenecks, especially when combined with public investment under the IndiaAI Mission.

Ultimately, ensuring fair compute access is a matter of technological sovereignty. India should use EFD as a conceptual backstop but rely primarily on ex-ante regulation to build an AI ecosystem that remains open, contestable, and resilient.

About the Author

Managya Sinha is a second-year law student at National Law School of India University, Bengaluru.