AI’s Hidden Monopoly: The Rise of Infrastructure-Level Dominance

[Alok is a student at Dr Ram Manohar Lohiya National Law University.]

The competitive landscape of artificial intelligence (AI) is increasingly extending at the infrastructure level in the form of specialized chips, cloud platforms, and APIs that underlie every model.  In India, the Competition Commission of India’s (CCI) 2025 Market Study on Artificial Intelligence and Competition highlights that concentration across the AI stack presents a challenge to fair competition. The study highlights that a few global firms control AI resources such as GPUs and cloud capacity, creating bargaining power imbalances in contracts between Indian startups and hyperscalers. As per the data of the CCI reports, barriers to entry for domestic AI developers arise from limited data availability (68%), high cloud service costs (61%), shortage of skilled talent (61%), prohibitive compute costs (59%), and restricted funding access (56%). To tackle these issues, the Indian government launched the IndiaAI Mission, worth INR 10,300 crores, to democratize computing. The government's policy objectives, such as the recent India AI Governance Guidelines, also demonstrate its intent for infrastructure and capacity building.

Nvidia’s GPUs and the major cloud providers (AWS, Google, Microsoft) act as invisible gatekeepers, controlling essential inputs for AI. Yet, much of the public competition discourse still emphasizes algorithms and models, overlooking how vertical integration of hardware and cloud services can lock out rivals. Nvidia’s $100 billion investment in OpenAI and its expansion from a chipmaker to an AI ecosystem illustrate the trend of vertical integration in AI. This issue is highlighted by the U.K. Competition and Markets Authority, the US Department of Justice and the US Federal Trade Commission and by the members of G7.

This could be termed a “hidden monopoly” that could stifle innovation, especially for startups lacking resources. Regulators are beginning to take note of this latent factor. For example, the EU has highlighted the risks of a few firms controlling GPUs and cloud capacity. In India, where local startups rely heavily on external computing, the dominance of NVIDIA, AWS, and others could translate into structural barriers, prompting calls for remedies like a national computing grid and open-access rules.

The AI Stack and Its Gatekeepers

AI systems rest on a multilayered stack of resources, which is also highlighted by the CCI’s stack typology. The upstream layers comprise data (its collection, preparation, and labelling), computing/infrastructure (which consists of servers, specialised chips, cloud/edge computing), development (algorithms, programming frameworks, and neural networks), and foundation models (large-scale models, such as LLMs). Following the upstream layer, models are fine-tuned for specific domains, deployed via cloud or on-premises infrastructure, integrated into decision-making systems, and governed through compliance mechanisms. 

The CCI's primary research itself reveals India's structural dependence on upstream layers. According to the study, approximately 67% of surveyed Indian AI startups operate in the AI application layer, building end-use applications on top of existing models and infrastructure. Only 3% develop foundation models, and roughly 10% focus on compute infrastructure; this number is very less considering the ambition of India to be a global leader in “Safe and Trusted AI”. 

GPU Monopoly

Even at the hardware level, the concentration is severe. According to the CCI’s study (Figure 5), Nvidia dominates the market with an overwhelming number in AI-grade GPU sales. By some counts, Nvidia now accounts for over 80-95% of AI-grade GPU sales. This dominance spans both revenue share (Nvidia captured 87% of its own revenue from data centre solutions in 2024) and workload concentration (the majority of AI training and inference workloads globally run on Nvidia GPUs). New entrants or startups cannot easily switch to alternative chips when Nvidia’s supply is scarce or priced at a premium.

As a Time magazine analysis notes, Nvidia today is not just a chipmaker but “the most valuable company in the world” that effectively monopolizes AI hardware. Other chip vendors might struggle to displace Nvidia’s incumbency; even cloud giants often turn to Nvidia GPUs for top-performance AI clusters.

Cloud Platform’s and API’s Oligopoly

Cloud platforms layer consisting of massive data centres and cloud services forms an oligopoly. Amazon Web Services, Microsoft Azure and Google Cloud account for roughly two-thirds of global cloud capacity by revenue in Q2 2025, with AWS holding 30%, Azure 20%, and Google Cloud 13% respectively, thereby determining priority access to limited GPU slots.

The cloud layer creates its own barriers. Because training large models requires enormous data-centre horsepower, most startups cannot build their own on-premise clusters and instead rent GPUs on hyperscale clouds. However, these services can be expensive and are often provisioned unequally. According to anecdotal evidence, many AI startups faced delays or price spikes as hyperscalers rationed GPU time. The CCI reports that many AI developers “depend on [big tech] firms for critical cloud and computational infrastructure, often through restrictive contractual arrangements,” yielding a “bargaining power imbalance”. This can mean that start-ups are forced to sign lock-in contracts or pay premium rates to access the necessary GPUs.

In addition to this, proprietary models / APIs are also significant players, as many cutting-edge foundation models are only accessible via proprietary APIs. Thus, controlling a popular model’s API becomes another gatekeeping lever: its provider can throttle, price or tie that API usage. Together, these layers form a vertically integrated chain, and at each step, a few big companies can exert control, resulting in a dominant position.

Problems of Concentration and Vertical Integration

This high concentration has real competitive effects and becomes adverse if Vertical integration occurs. Vertical integration happens when a firm performs an activity itself that it could otherwise procure on the market. This can generate efficiencies in the form of system optimisation, reduced transaction costs, and faster innovation.

Hyperscalers determine who receives priority on the limited GPU slots, often guided by their vertical integration with cloud providers. Since major cloud vendors also often invest in or partner with certain AI companies. For example, Amazon and Google have stakes in Anthropic, Microsoft has a stake in OpenAI, blurring the lines between supplier and customer. 

This raises concerns about fairness. The EU competition authorities have flagged the issue of preferential arrangements that give some players privileged access to “key components of generative AI (e.g., GPUs, supercomputing power, cloud capacity),” risking the foreclosure of rivals.  They note such deals “may increase the degree of market concentration” and “make access to critical inputs more difficult”. It implies that cloud dependence means a few hyperscalers can effectively curate who gets to build AI and on what terms. Similarly, the US Federal Trade Commission found indications based on internal documents and industry evidence that some cloud providers may have prioritised scarce GPUs for companies they had invested in (rather than independent challengers). Therefore, it can be said that hyperscalers can “pick winners and losers among their customers” because they operate like unregulated utilities. The main issue arises when dominant infrastructure providers compete downstream or preferentially allocate resources to invested companies, intensifying foreclosure risks.

One competition scholar observes that an oligopoly in GPUs and cloud means “oligopolists in these layers can leverage their power downstream through self-preferencing, tying, or vertical integration”. For example, an incumbent could bundle chip sales with exclusive cloud discounts or lock out rivals from using their GPUs. Indeed, the hardware layer “was [historically] a monopoly structure…with NVIDIA the predominant provider of processing hardware”, and evidence suggests Nvidia is already moving to offer its own cloud services, hinting at further integration.  Emerging cloud-scale GPU startups (like CoreWeave or Lambda Labs) may challenge Nvidia’s direct chip sales, but even they depend on Nvidia’s designs.

Interestingly, vertical integration reaches to other tech titans as well. Google and Amazon now design custom AI chips (Tensor Processing Units, Trainium, Inferentia) and build datacenters around them. Microsoft similarly rolled out Azure-exclusive GPUs and invested heavily in OpenAI.  These moves blur the lines between “chip, cloud and service,” reinforcing incumbents’ control over the stack. 

The question is not whether integration can be efficient, but whether current concentration levels (given high barriers in GPU manufacturing and cloud infrastructure) foreclose competition that would yield greater innovation. At this scale, anticompetitive effects may outweigh efficiency gains.

Barriers for Startups: Indian Context

For Indian AI startups, these global trends pose significant challenges. Few domestic firms have billions to invest in dedicated data centres, so most are forced to run on external clouds and GPUs. This exposes them to the whims of foreign providers. The CCI’s AI market study explicitly notes that “control of a few large firms across the AI stack may create barriers to entry for smaller players”. The CCI's survey quantifies the depth of this dependency, noting that about 52% of user-industry respondents report that larger firms' superior data access confers a significant competitive advantage in AI deployment, establishing data concentration as a structural barrier. Moreover, 78% of AI user firms rely on third-party AI solutions rather than building proprietary systems, creating downstream dependency risks where startups and enterprises alike become locked into suppliers' ecosystems with limited exit options.

From a competition law perspective, the CCI's framework enables scrutiny of such concentrations. Exclusivity clauses in cloud supply agreements or restrictive IP licensing practices can be assessed under Sections 5 and 6 of the Competition Act 2002. These sections govern combinations (mergers and acquisitions) and abuse of dominant position.

Notably, the deal value threshold was introduced in 2024, which permits the CCI to review transactions exceeding INR 2,000 crore, even when the target enterprise falls below traditional asset or turnover thresholds. This applies if the target has "significant business operations" in India. This reform directly addresses the "killer acquisition" phenomenon, enabling the CCI to assess whether partnerships involving hyperscalers and infrastructure providers might foreclose competition or entrench dominance, particularly when cloud access or GPU capacity becomes a contested resource.

A CCI’s survey confirms the pain points. Indian AI startups frequently report unequal access  to computing resources as a key concern in the competition. The chart given below shows the perception of the stakeholders regarding competition issues arising out of the adoption of AI.

In addition to the given data, the CCI's user-industry survey indicates that 90% of AI-adopting firms leverage AI to monitor and track customer behaviour, 27% deploy AI for monitoring supply chain efficiency, while 24% use AI for forecasting pricing trends and 21% for predicting inventory requirements. These adoption details reveal that most firms deploying AI are fundamentally dependent on external compute capacity. The CCI warns that AI could reshape market structure if left unchecked, as currently, the firms cannot function competitively without reliable, affordable access to the cloud infrastructure and GPU capacity controlled by a handful of hyperscalers.

Policy Remedies and Future Outlook

Indeed, startups often resort to open-source models and cheaper hardware to survive (about two-thirds build only AI applications with open-source tech for cost reasons), but this only partially offsets the need for high-end GPUs on demand. Another issue is that these open-source solutions can’t fully substitute for specialised compute infra. This structural imbalance means most Indian startups are downstream consumers rather than upstream providers of core AI resources, exposing them to the pricing, allocation policies, and strategic priorities of a few global gatekeepers.

The Indian government has recognized the crunch. Through the IndiaAI Mission, it is building a “common compute” infrastructure to democratize access to AI. Practically, India has pooled thousands of GPUs through public-private partnerships. By mid-2025, national compute capacity is expected to surpass 34,000 GPUs, including modern chips from multiple vendors, thereby reducing vendor lock-in. The trend extends beyond government projects.  Indian conglomerates are also building AI data centres: Nvidia’s collaborations with Reliance and Tata will bring cutting-edge supercomputing to India. All these efforts aim to counter foreign hyperscale dominance by creating robust domestic computing options.

One approach is to treat cloud and GPU services as essential facilities, requiring leading providers to offer fair, reasonable and non-discriminatory access to their resources. The EU's Digital Markets Act 2022 lists cloud computing as a potential core platform service, though no providers have yet been designated as gatekeepers.[1] [AxH2]  Competition authorities could scrutinize exclusive GPU / cloud partnerships similarly to mergers. The EU Commission's Article 22 review of Nvidia's acquisition of Run:ai, although ultimately cleared in December 2024, demonstrates that even below-threshold infrastructure transactions face regulatory examination, signaling heightened scrutiny of hardware consolidation.

Another remedy is structural separation. This is similar to past antitrust fixes where railroads were forced to relinquish coal mines; in telecommunications, structural competition issues were partly addressed through regulatory “antitrust fixes” such as mandated network interconnection under the Telecommunications Act 1996; similarly, an AI Compute Grid could function as a neutral backbone where no single provider has control. In crux, it is argued that key layers should be separated: “chips must be independent from clouds, and clouds must be independent from AI models”  to preserve open competition.

In fact, India’s “common compute” initiative is a prototype of this idea.  India effectively creates a national computing grid by pooling GPUs under government coordination. Startups and researchers can draw on this shared infrastructure equally, limiting single-supplier bottlenecks

Finally, transparency and effective competition monitoring are crucial. The CCI recommends audits of AI systems and workshops on competition compliance to prevent self-preferencing and collusion. The study urges all enterprises deploying AI with market impact to adopt a six-pillar self-audit framework comprising governance, algorithm design, testing, monitoring, transparency, and compliance integration.

Firms are expected to disclose the purpose of AI in decision-making, ensuring that sensitive or proprietary details remain protected. Internal algorithm audits should check for signs of inadvertent collusion or price discrimination, with human-in-the-loop review. The CCI explicitly calls for greater transparency on GPU allocation criteria and cloud pricing. It requires providers to lower information asymmetries and enable effective regulatory oversight of resource distribution, thus guarding against market abuses while preserving commercial confidentiality for innovative algorithms. Ultimately, steps are taken to achieve the goal of ensuring that hardware gatekeepers do not hinder AI innovation, while also preserving efficiencies from integration where these truly outweigh competitive harms. It can be done by striking a balance between preventing hardware gatekeepers from hindering rivals and allowing integrated models to deliver proven technical or economic benefits.

Conclusion

AI's infrastructure monopoly shapes competition through control of chips, clouds, and APIs, rather than relying solely on algorithms. Without intervention, these layers risk becoming unregulated utilities controlled by a few firms, determining who participates in the AI economy. Policy responses should prioritise transparency in GPU allocation and cloud pricing, expand public compute infrastructure along the lines of India's model, and scrutinise vertical partnerships. India's compute grid demonstrates that open, shared infrastructure can level the playing field. The question is whether other jurisdictions will act before concentration becomes irreversible.