Generative Artificial Intelligence: Unravelling the Competition Conundrums

Introduction

Artificial Intelligence (AI) refers to machines that process and interpret input data to produce outputs in the form of decisions, suggestions, and other responses.¹ While the concept of AI seems like a recent innovation, it has existed since the advent of digital technology. Over time, what has evolved is its capabilities and scale. A notable example of this evolution is the development of generative AI, which, although often used interchangeably with AI, is a subset of the latter.

Generative AI is capable of simulating human intelligence by employing parameters for the statistical analysis of large datasets to produce outputs such as images, videos, audio, text, and computer code.² This capacity has enhanced user interaction by providing more accurate and tailored outputs. However, these advancements have also raised competition concerns in the sector.

Building blocks: Generative Artificial Intelligence

The elements involved in the development of generative AI must be understood before addressing the competition concerns associated with generative AI. This understanding provides clarity and context for analysing the competition concerns.

Technically, the foundation model of generative AI is built upon various interconnected components that enable it to process input and generate output. These include algorithms, computing power, data, neural networks, parameters, tokens, and human expertise. However, for the purpose of competition concerns in generative AI, we will adhere to three pivotal elements, i.e. data, computing power, and human expertise. These inputs are critical for the creation and effective functioning of the foundation model of generative AI.

Firstly, data is an essential element in the development and performance of the foundation model. By feeding an enormous amount of data, the foundational model learns patterns from the data and generates outputs.³ To ensure diverse outcomes, a variety of data needs to be fed to the foundation model, such as text, images, videos, audio, etc. Secondly, computing power is another essential component, as data will not automatically turn into outputs; hence, computing power is required, often facilitated by the graphical processing units (GPUs), to enable the rendering of large datasets at high speed and turn inputs into meaningful outputs.⁴ Thirdly, no matter how much data and computing power an entity possesses, the development of generative AI is impossible without human expertise, which is required for optimising and refining the foundational model throughout time.⁵

A well-established principle of competition law states that harm to competition leads to a concentration of power, thus undermining consumer interest and market dynamics.⁶ Ensuring competition in the AI sector is imperative, given its pervasive influence across sectors, especially the economy.⁷ Promoting a competitive AI ecosystem ensures that the benefits of this technology are widely distributed, rather than controlled by a few major players. The competition issues associated with the generative AI sector are analysed under the following headings.

Data: The fuel driving AI dominance

In the contemporary digital economy, data is regarded as the new currency, a notion that is also reinforced by generative AI, which relies heavily on vast datasets to train its foundational models. Generally, firms use two kinds of data for this purpose, (i) public data and (ii) proprietary data.

Public data does not pose major competition concerns since it is available to all market participants. However, large digital platforms benefit from the lower acquisition costs for such public data due to their well-established presence in adjacent markets and robust supply channels for the data.⁸ The primary competition concerns that arise about generative AI pertain to proprietary data, which is more unique and specific. The Competition Authorities are likely to face five key dilemmas with regard to the proprietary data:

Firstly, it is stated that a significant portion of the world’s data exists as licensed or proprietary data, which cannot be replicated by other firms; for example, Google has access to the video transcripts from YouTube,⁹ Microsoft leverages professional data from LinkedIn,¹⁰ and Amazon has access to a database of personalised customers’ shopping activities.¹¹ This proprietary data enables large digital platforms to train their models in a tailored manner, enabling the generation of more specific and rational outputs in response to user queries.

Secondly, while developers of the foundational model of generative AI may not have direct access to proprietary data, they can capitalise on their first-mover advantage to secure exclusive licensing agreements with proprietary data holders, such as publishers and data scraping firms. For instance, Google has entered a strategic partnership with Stack Overflow, a platform for programmers to address programming-related queries, to enhance its generative AI development.¹² Similarly, OpenAI has partnered with Axel Springer and Financial Times to use its content for training foundational models. These arrangements provide developers access to unique data for training their models.¹³ Furthermore, by formalising such a partnership through an official exclusive agreement, these developers can ensure that they will not come within the purview of the infringement of intellectual property rights, i.e. copyrights.

Thirdly, developers can gain access to proprietary data through arrangements with a service user. In such cases, the data is provided by the user entity in exchange for the generative AI services. For instance, a securities advisory firm requires a personalised chat service for client queries and may grant access to its personal datasets as terms of agreement in return for providing the services. This arrangement allows developers to fine-tune the foundational model using personalised data, tailoring it for specific purposes.

Fourthly, at the deployment stage, developers benefit from a personalised model created with the licensed data and may gain a significant competitive advantage. These refined models are likely to attract more users, enabling developers to collect the personal data submitted by the users through queries, and this data can be used to further fine-tune and enhance the model, creating a self-reinforcing cycle.¹⁴ This will further attract a large number of users, causing the network effects, which the rival would not replicate.

Fifthly, once developers acquire proprietary data, either through agreements or user interactions, they may employ tactics to self-reinforce their market position and limit users’ switching. Customers may find it difficult to switch to alternative AI services if their data is locked with the developer. Further, users who became accustomed to the interface and outputs of a particular generative AI may be reluctant to switch to other service providers. This situation creates barriers to entry for other firms, hindering fair and equitable competition in the market.

Such scenarios underscore how firms with secured access to personalised data derive substantial competitive benefits, which are further amplified in the case of the first-mover advantage. This early lead allows them to rapidly scale and refine their models, ensuring their control in the generative AI sector. Moreover, the generative AI sector is prone to economies of scale, where the ability to train models on vast datasets enhances model performance and reduces marginal costs. Firms that attain such scale are in a position to monopolise the market by creating barriers to entry for new entrants, leading to the consolidation of market power.

Furthermore, major digital platforms in the digital market, such as Google, Microsoft, Amazon, etc., possess extensive repositories of proprietary data. Their position enables them to exert vertical control over the generative AI value chain. By restricting access to data for new entrants that are dependent on them, these firms can enhance their market power and hinder effective competition.

GPUs: The powerhouse of AI

The foundation model of generative AI operates by analysing extensive datasets, necessitating substantial computing power for effective training of the model. This method entails performing statistical analysis on enormous datasets to provide accurate outputs, which necessitates the use of specialised hardware because such operations cannot be performed on conventional computers, i.e., CPUs.¹⁵ As a result, developers rely on graphics processing units (GPUs), which are capable of processing large volumes of data at a fast pace.¹⁶ In such a scenario, developers have two alternatives for meeting their computing needs: (i) purchasing GPUs and creating their computing infrastructure, or (ii) signing a service agreement with firms that provide cloud computing services. In both cases, competition law concerns may arise.

In the first alternative, the Competition Authorities may confront two major dilemmas: Firstly, a few businesses dominate the worldwide supply and production of GPUs. For instance, one of the biggest suppliers of GPUs is Nvidia.¹⁷ Other industry players, particularly Google, also develop their processors for generative AI services, i.e., Tensor Processing Unit (TPU).¹⁸ The increasing interest in generative AI has accelerated the demand for GPUs, leading to shortages and rising prices of GPUs.¹⁹ Furthermore, training, fine-tuning, and hosting generative AI models demand massive computer capacity, necessitating a considerable capital commitment. For instance, it was stated that the GPT-4 model training cost more than $100 million.²⁰ This creates an ecosystem where major firms, with their well-established supply chains and deep pockets, can survive since these players can afford to purchase GPUs at elevated prices and bear the high computational cost. This, in turn, entrenches their market positions and creates barriers to entry for new entrants.

Secondly, competition concerns also arise from the fact that many chip manufacturers are also developers of foundational generative AI models. For instance, Nvidia has developed its own model, Nvidia NeMo, capable of generating text and images.²¹ Similarly, Google has its generative AI model, Gemini.²² As these entities operate in vertical markets and also control key inputs, i.e., computing power, they may restrict access to these resources for new entrants developing competitive foundational models. This could enable them to entrench the market power in the generative AI sector and hinder fair competition.

Since most new entrants lack access to substantial capital and face the risk of an incumbent firm’s potential restrictions on key inputs, they are likely to rely on cloud service providers to meet the need for computational power needs. However, this shift introduces additional competition concerns. In this scenario, the Competition Authorities would encounter two crucial dilemmas concerning cloud computing:

Firstly, the supply of cloud computing services is highly concentrated with a few majors firms, such as Microsoft (Azure AI), Amazon (Amazon Web Services), and Google (Google Cloud).²³ These entities are often referred to as hyperscalers, as they operate colossal data centres worldwide and are the primary providers of cloud computing services to other entities. A significant competition concern arises from their dual role as both gatekeepers of key inputs of generative AI and providers of their own AI services to users. This structural overlap creates strong incentives for discriminatory conduct, where access to cloud services may be selectively denied to new competitors who threaten their market positions or who don’t follow the terms and conditions of cloud computing services, in order to preserve their market share. For instance, reports suggest that Microsoft has restricted access to certain datasets for firms that are newly entering the generative AI sector.²⁴

Secondly, it has been observed that new entrants in the generative AI sector are increasingly pursuing vertical integration by signing agreements with major players in the digital market to secure access to cloud computing services. For instance, Microsoft provides exclusive cloud services to OpenAI for its generative AI model, ChatGPT;²⁵ Google offers cloud services to Anthropic;²⁶ and Amazon partners with Hugging Face for cloud services.²⁷

The competition concerns arising from such arrangements include the use of cloud service credits, which serve as incentives to tie clients to a specific provider’s ecosystem. Additionally, the issue of data interoperability arises because cloud service providers impose egress fees on users for transferring data from one cloud service provider to another. This limits consumers’ freedom to select the cloud service providers of their choice, resulting in a lock-in effect. Furthermore, since these major digital market players also offer their AI services, they may take advantage of their position by lowering the calibre of computing platforms offered to competitors. By taking these steps, they may be able to reshape the generative AI sector to suit their needs and preserve their market leadership

Monopoly on minds: Backbone of AI

The efficiency of the foundational model of generative AI hinges upon large datasets, which necessitate training by human experts, making the process inherently labour-intensive. For instance, foundational models require large datasets for the parameter training and to ensure the system remains free from prejudices, human experts must scrutinise and eliminate the biased data. Moreover, the role of experts extends beyond the pre-training stage to the operational stage, where they assess the feedback of queries and fine-tune the parameters of the foundational model to enhance accuracy and appropriateness in its response. This process reflects that human expertise is a linchpin for the functioning of foundational models.²⁸ However, these requirements have enhanced the demand for experts in AI, resulting in a scarcity of experts and driving up wages²⁹.

In this context, competition concerns arise particularly when entities engage in practices such as predatory hiring and poaching, aiming at harming competitors by hiring their top-skilled workforce.³⁰ Since employees are valuable assets to any entity, such actions can intensify competition issues. These hiring practices, amid the scarcity of the skilled workforce, drive up wages to levels that may be unsustainable to new entrants. This creates a barrier to entry for the new entrants to enter the market because they will struggle to afford the human resources essential for developing the foundation model of generative AI, ultimately forcing them out of the market, and establishing the dominance of established entities.

Even when an entity manages to pay higher wages due to the practice of major players engaging in predatory hiring by offering inflated salaries to employees, increasing the overall cost of the end products, which consumers ultimately bear, however, the major players can mitigate this impact by leveraging economies of scale and temporarily offsetting prices to exclude competitors from the market. This process further strengthens the position of the major entity in the market. For instance, OpenAI has enticed top AI talent from Google by offering compensation packages of $10 million³¹ and Google has spent $2.7 billion to rehire former employees, such as Noam Shazeer, as Vice President.³² Similarly, Microsoft has brought on Mustafa Suleyman, CEO of Inflection AI, to lead its consumer AI business.³³ Furthermore, Amazon has entered into an agreement with Adept AI, where the CEO and key AI personnel will now work for Amazon.³⁴ Also, Huawei has been poaching key employees from Zeiss SMT, known for developing powerful semiconductors for supercomputers.³⁵ These examples underscore the predatory hiring practices employed by firms to disrupt the labour market and weaken potential competitors in the market.

Conclusion

While the development of generative AI is still in its early stages, the exact competition concerns that may arise are uncertain. Therefore, it would be prudent for Competition Authorities to conduct comprehensive market studies on the AI sector to gain a better understanding of its dynamics and potential antitrust challenges. These studies would provide valuable insights into the structural and behavioural dynamics of the market, helping to identify emerging antitrust risks. Additionally, these market studies would help to promote awareness among stakeholders (policymakers, developers, and users) about generative AI’s potential impact on the competition.

To effectively implement the aspect of market studies, Competition Authorities across jurisdictions should consider entering into international cooperation agreements for the exchange of information about the AI sector. Collaborative efforts, such as joint market studies, would enhance regulatory preparedness in light of the rapid advancements in the AI sector. At this preliminary stage, offering further regulatory suggestions beyond those mentioned above would be premature and lacking the necessary empirical foundation to ensure a practical and reasonable outcome.