Antitrust & AI supply chains

Introduction

The buzz these days is AI, specifically generative artificial intelligence and foundation models, which are “trained on broad data (generally using self-supervision at scale) that can be adapted to a wide range of downstream tasks.” ^[1] In February 2024, AI chipmaker Nvidia joined the multi-trillion-dollar market valuation club, joining existing members Microsoft, which has partnered with OpenAI and is incorporating OpenAI’s foundation model across its ecosystem, Google, which is developing its own foundation model, and Apple, which in 2024 shifted resources from driverless cars to AI. As Nvidia’s CFO noted, “Building and deploying AI solutions has reached virtually every industry.” ^[2] New technological developments promise to further catapult generative AI technology, with significant implications for citizen autonomy, social stability, and democracy.

When surveyed on AI’s likely future impact on society, AI experts have expressed concerns about personal privacy, human rights, democracy, and civility. ^[3] Tech titans Elon Musk and Larry Page have famously debated whether artificial intelligence would ultimately elevate or destroy humanity. The risks that generative AI may present to humanity have also been linked to the departure of Ilya Sutskever, chief scientist and cofounder of OpenAI, from the company. ^[4] His newly founded company, “Safe Superintelligence,” has attracted much interest, with a valuation of over $5 billion in its first months of operation. ^[5] The lightspeed evolution of the technology and the continuing investment in AI-related initiatives capture an increasing portion of our social and economic activity ^[6] and fuel the debate about the technology’s impact.

Alongside these significant developments, a separate discussion has emerged, looking at how AI markets will evolve and who will control the technology. In this context, policymakers and competition officials are exploring foundation AI models and this technology’s likely impact on innovation and competition. The critical question is whether these foundation models will herald disruption that would challenge existing gatekeepers and tech barons and bring about new dynamism to digital markets, or will AI further entrench market power. This question lies at the heart of this Article.

Assessing the impact of foundation models on competition is nuanced. There might be more innovation, but the type of innovation (sustaining versus disruptive) and its value (positive, extractive, or negative) might vary. In some ecosystems, we might see many more sustaining innovations that primarily extract value from consumers or other producers. We may see more toxic innovations that reduce our and our children’s wellbeing.

The importance of identifying the trajectory of generative AI markets and ensuring their contestability cannot be overstated. As many competition officials have candidly admitted to us, their agencies missed or underappreciated the digital market dynamics that led to “a winner-take-most-or-all” that led to the present Big Tech Barons, namely Google, Apple, Meta, Amazon, and Microsoft. Currently, enforcers around the globe are scrutinizing these dominant ecosystems, which are heavily investing in generative AI. But there is no consensus over how to prevent, among other things, the distortion of competition and innovation. The concern is that feeble or delayed action will miss the potential tipping of these emerging AI markets. As the experience with the Big Tech Barons reflects, antitrust enforcement, if too little (e.g., primarily monetary fines) or too late (cases that take years to develop and litigate), will not restore competition and innovation levels. It is far easier to keep markets contestable than to try to restore contestability.

Because the AI landscape is nuanced, any intervention needs to appreciate the complexity of the technology stack, the competition dynamics at stake, and the existing dominance of a few ecosystems. Ultimately, two fundamental inquiries emerge:

• First, the digital economy has several factors and characteristics that can lead to concentrated markets. Are there similar factors in the emerging AI foundation model supply chain that will lead to a “winner-take-most-or-all” situation?

• Second, could AI herald new business models and innovations that disrupt the dominant ecosystems? Or will the current ecosystems dominate these AI-driven foundation models and applications?

Within this context, this Article considers whether this generative AI technology forms a disruptive force that could destabilize the current powerhouses, foster competition, and ensure future disruptive innovation that improves our wellbeing, or might the technology herald the next step in the ongoing entrenchment of critical ecosystems? This inquiry requires careful consideration of the emerging foundation model supply chain and the extent to which current strategies, market characteristics, and technologies could entrench a few powerful firms.

To explore this issue, Part I briefly elaborates on foundation models and their significance. Part II outlines the current dynamics in many digital markets that lead to a winner-take-most-or-all situation. Part III next examines the emerging AI foundation model supply chain and considers several potential antitrust risks that emerge should specific layers become concentrated and firms extend their power across layers. Part IV raises several countervailing factors that might lessen or prevent these antitrust risks. This Article concludes with suggestions for the policy agenda to promote both healthy competition and innovation in the AI supply chain.

I What Are Foundation Models?

Narrow non-foundation AI models are “trained on specific data for a specific task and context” and are “not designed for reuse in new contexts.” ^[7] For example, a bank may train a narrow AI model to predict the risk of a customer defaulting on a loan. However, that model could not do many other things, like chat with the customer to assist them with a transaction. ^[8] In contrast, foundation models are “AI models designed to produce a wide and general variety of outputs. They are capable of a range of possible tasks and applications, such as text, image or audio generation. They can be standalone systems or can be used as a ‘base’ for many other applications.” ^[9] Their flexibility and capacity enable these models to serve as the foundation for many other applications—from creating a poem to acing a biology test to creating a podcast about this Article. ^[10]

Let us begin with the data to see how this works and how foundation AI models differ from non-foundation models. A foundation model is trained on vast amounts of data. In collecting a lot of data from different sources, the developers can train the foundation model to “infer patterns, recognize relationships, discern context and generalize their knowledge.” ^[11]

Next is the modality. The types of foundation models vary, including “generative adversarial networks (GANs), variational auto-encoders (VAEs), transformer-based large language models (LLMs), and multi-modal models.” ^[12] As the latter reflects, foundation models can be uni- or multi-modal, with reference to the type of data used to train the model. Uni-modal models are trained to handle one type of data (such as generating text outputs from text inputs). ^[13] Google’s Gemini model, in contrast, is multi-modal in that the foundation model is trained using different types of data, including words, images, videos, and audio. ^[14] Prominent among the foundation models are “large language models” and “image generation models” that enable creative text, video, and audio and, as such, differ from narrower AI technologies used to analyze existing data.

So, one could ask a multi-modal foundation model to draw an image or create a movie from a written text prompt. Below is one prompt that is part of a U.S. copyright infringement action against the AI model’s developer. Here, the multi-modal model takes a text prompt (show me a painting of the Brooklyn Bridge in winter in the style of this particular artist) and produces the following image: ^[15]

Text prompt: “Painting of the Brooklyn Bridge in winter in the style of Roy Lichtenstein.”

Source: Center for Art Law, DreamUp generated output image

A third differentiator is the model architecture. Many foundation models rely on a deep learning architecture, which “uses multilayered neural networks to mimic the human brain’s decision-making process.” ^[16] Typically, this deep learning architecture helps the foundation model to undertake a “wide range of tasks and operations.” ^[17] After training and evaluating the foundation model using certain benchmarks, one can fine-tune the pre-trained foundation model for further particular tasks. ^[18]

Foundation models have already been integrated into many facets of our lives and businesses to facilitate search, improve software, communications, text completion, text-to-speech conversion, content creation, entertainment, financial services, legal services, healthcare, education, pharma, fashion, marketing, insurance, and more. ^[19] These technologies promise to shake up the status quo in many industries as they drive new business ventures, creativity, efficiency, and innovation. Not surprisingly, the market has reacted with excitement, resulting in high valuations of the leading players and expectations for exponential growth in the coming years. ^[20]

Earlier versions of this generative AI have given way to more advanced performances. Consider Bill Gates. In June 2023, Microsoft’s founder was unimpressed with ChatGPT’s performance: “I was like, ‘Yeah, it’s kind of an idiot savant. I don’t think it’s practical.’” ^[21] So, he told the OpenAI CEO to get back to him when the foundation model could achieve the highest score on the college advanced placement test for Biology. Gates thought, “OK, that’ll give me three years to work on HIV and malaria.” Two months later, OpenAI’s developers returned, and Gates watched ChatGPT achieve a 5 on the AP Biology test. “I’m still, personally, in a state of shock at ‘Wow, it is so good,’” Gates said. (Only 14.3% of high school students who took the exam in 2023 obtained that top score. ^[22])

Its GPT-4 model, according to OpenAI, has done even better, exhibiting “human-level performance on various professional and academic benchmarks.” ^[23] These benchmarks include GPT-4 racking the highest scores on many other AP subjects (including history, statistics, psychology, and micro- and macroeconomics), scoring in the 90th percentile on the Uniform Bar Exam for lawyers (the GPT-3.5 model scored only in the 10th percentile), in the 99th percentile on the graduate school GRE Verbal Assessment, and even a 77% on the Advanced Sommelier examination. ^[24]

The performance of these AI models is constantly improving, and key players are now racing toward the ultimate goal: developing Artificial General Intelligence (AGI)—an AI system with a human level of cognition and self-teaching capacity to undertake various tasks autonomously. ^[25] In 2024, OpenAI released its o1 model with “reasoning” abilities, bringing us closer to human-like artificial intelligence. ^[26] The new generative AI models increasingly excel at complex problem-solving, math, and coding, and bring us closer to the ideal of AGI.

Add to the mix the increasing investment in AI. Nearly one-third of investments by venture capital firms in 2024 have been in AI companies. ^[27] As the Wall Street Journal remarked in late 2024, “Generative artificial intelligence has sparked one of the biggest spending booms in modern American history, as companies and investors bet hundreds of billions of dollars that the technology will revolutionize the global economy and one day lead to massive profits.” ^[28] The belief is that this AI will transform many different markets, which vary in contestability and other competitive dynamics.

Not surprisingly, besides new value opportunities and efficiencies, the latest AI foundation models bring notable risks of misuse and concerns over human autonomy and social cohesion. ^[29] President Joe Biden’s executive order on AI, for example, warned that “irresponsible use could exacerbate societal harms such as fraud, discrimination, bias, and disinformation; displace and disempower workers; stifle competition; and pose risks to national security.” ^[30] These risks include non-experts, such as terrorists, using these foundation models to “design, synthesize, acquire, or use chemical, biological, radiological, or nuclear (CBRN) weapons” or enable “powerful” cyber-attacks. ^[31] Thus, the risks from these models extend far beyond antitrust.

But before delving into whether this AI will disrupt the existing Big Tech Barons, let us first review the competitive dynamics of the digital economy, which led the Big Tech Barons to grow their powerful ecosystems.

II The Digital Market Dynamics that Led to the Big Tech Barons

Most U.S. companies live short lives. One study found that the average half-life of U.S. publicly traded companies is close to 10.5 years, meaning that half of all companies that began trading in any given year have disappeared in 10.5 years. ^[32] Nor are platforms guaranteed long lives. One study calculated that 209 platforms had failed and died over the past 20 years. Most of them (85 percent) were transaction platforms, which had shorter lives (on average 4.6 years) than the innovation platforms (5 years) or hybrid platforms (7.4 years) in the survey. ^[33] So, suppose many companies and platforms die within ten years of their birth. In that case, why have the Big Tech Barons successfully dominated multiple markets for years, and why do they seem poised to continue their domination over the next decade?

In the Western world, despite the billions invested by Microsoft in its search engine Bing, Google, as of 2024, still dominates general search services and search text advertising. ^[34] Despite the rise of TikTok, Meta still dominates personal social networks ^[35] and social media ads. ^[36] Meta, Google, and Amazon dominate digital advertising generally. ^[37] Amazon remains the dominant online shopping platform. ^[38] Google and Apple remain the leading operating systems for smartphones. ^[39] Amazon, Microsoft, and Google control the leading cloud computing platforms. ^[40]

As this Part explores, the digital economy generally exhibits several characteristics that support a winner-take-all, or winner-take-most, outcome, with only a few companies dominating markets. Noteworthy elements include scale economies, big data, network effects, big analytics, the ability to channel digital innovation, and an array of anticompetitive measures, including acquire-copy-or-kill. ^[41] As we, among many others, have explored these themes elsewhere, ^[42] let us just summarize here a few:

Big data, which drives much of the innovation and offering in the digital economy, has risen to become a key asset in many digital markets. Data may be structured or unstructured, ^[43] gathered voluntarily, through observations, or by inferring new information using existing data. ^[44] When considering personal data’s role in the digital platform economy, it is helpful to take note of the “four Vs of Big Data” – Volume, Variety, Velocity, and Value. ^[45] Alongside the many improvements Big Data offers, it is noteworthy that limited access to relevant and timely personal data may inhibit entry, expansion, and innovation. ^[46] Data has been referred to (somewhat controversially) as the “new oil” which powers our modern society. ^[47] While personal data is often “non-rivalrous” in the sense that it may be duplicated and used by multiple parties without depleting its value (and as such differs from oil or traditional assets), ^[48] its control, especially where the four Vs are present, offers a distinct advantage in the digital economy. A data advantage over rivals can enable leading players to achieve critical economies of scale, which could tilt the data—and competitive balance—in their favor. A positive feedback loop, in which control over data reinforces market power, which in turn increases the four Vs, helps the strong become stronger as the weak get weaker. ^[49]

Network effects occur when a good or service becomes more valuable as others use it. ^[50] Each of the Tech Barons’ ecosystems exhibits multiple network effects. They include direct, indirect, and data-driven effects that create a feedback loop that attracts users, sellers, service providers, developers, and advertisers to a network. ^[51] Alongside the unparalleled efficiencies and economies of scale offered by network effects, they may support the rise of significant barriers to entry and limit the competitive pressure on the incumbent. ^[52] Indeed, the combination of network effects, access to data, and analytics may tip the market in favor of a leading provider, which may become inefficiently entrenched. ^[53] Data-driven indirect network effects may, in particular, reinforce the feedback loop and lead markets to tip in favor of the leading platform.

Big Analytics offer the power to optimize the use of data, identify patterns, improve the understanding of market dynamics, and open the door to accelerated innovation. Advanced analytics have also been central in identifying consumers’ needs and wants. Recent advancements in analytics and AI may enable further optimization even with relatively limited or unstructured data. ^[54] Here, alongside the benefits, one can identify worrying trends. Data mining, data trade, online marketing, pattern recognition, demand estimation, and price optimization have been used to approximate reservation prices, identify biases, and power exploitative practices. Big Data and Big Analytics often reinforce each other. ^[55]

Ability to channel digital innovation. On the one hand, the leading digital players (Google, Apple, Meta, Amazon, and Microsoft) invest heavily in research and development. ^[56] Their investments are often associated with new technologies and disruptions of third parties’ value chains. ^[57] However, their control over their respective ecosystems also enables them to affect the nature and composition of innovation that reaches the market. The key here is how powerful ecosystems can distort the supply and demand for innovation, suppress disruption that threatens their value chains, and promote sustaining innovation that supports their business models. ^[58]

What is noteworthy is that despite the ongoing antitrust litigation, ^[59] heightened scrutiny in Europe under its new regulatory schemes, such as the Digital Services Act (DSA) and Digital Markets Act (DMA), ^[60] and the threat of further regulations in the EU and elsewhere, these digital ecosystems still dominate. In fact, despite ongoing enforcement and increased regulation, the Tech Barons’ power remains entrenched.

Consider Meta. It faces greater scrutiny in Europe under the DMA ^[61] and DSA. ^[62] The German Federal Cartel Office has also challenged Meta’s terms and policies on data sharing across its apps and collection from third-party websites via its business tools as violating German competition law. ^[63] In the U.S., the FTC is challenging Meta’s acquisitions of Instagram and WhatsApp. ^[64] A bipartisan coalition of states is challenging Meta’s exploitation of teenagers. ^[65] New Mexico has sued Meta for putting “profits ahead of children’s safety” in having its platforms becoming “prime locations for predators to trade child pornography and solicit minors for sex.” ^[66] The U.S. Senate Judiciary Committee excoriated Meta’s CEO (as well as TikTok’s), promising in January 2024 further regulations. ^[67] But the day after the Senate hearings, Meta’s stock jumped over 20%. ^[68] While such a price increase might be seen with a penny stock, one had never seen a nearly $200 billion increase in market capitalization in one day: Meta set a new stock market record. ^[69]

Wall Street clearly does not see Meta’s dominance as being under threat. Why? Meta controls an ecosystem, not markets, which provides greater power than a dominant platform. ^[70] Google, Apple, Meta, Amazon, and Microsoft each dominate several interlocking, multi-sided markets, which connect individuals, service and product providers, and advertisers. With positive feedback loops and ongoing expansion, their digital ecosystems benefit from unparalleled access to data and other critical inputs of the digital economy that can be used to improve products, services, technology, and algorithms and give them a significant advantage over others.

As our most recent book explores, ecosystems can be viewed as autonomies that control the competition dynamics, entry conditions, fees, and data flow within their respective frameworks. ^[71] Positive feedback loops increase the power asymmetry, resulting in a few key players influencing competition parameters, barriers to entry and expansion, the nature and flow of innovation, and user behavior and choice. ^[72]

With that background, let us consider whether AI will disrupt or protect the Big Tech Barons’ ecosystems. The answer lies in the emerging foundation model supply chain, which the next Part examines.

III The Foundation Model Supply Chain and Potential Antitrust Risks

In assessing whether AI will disrupt or protect the Big Tech Barons’ ecosystems, one must explore the supply chains at the heart of generative AI and consider their likely impact on competition and innovation. The Ada Lovelace Institute, an independent UK research institute with a mission to ensure that data and AI work for people and society, illustrates the emerging AI foundation supply chain: ^[73]

Source: Ada Lovelace Institute

At the upstream level, as of 2024, processing power is predominantly provided by Nvidia, which develops graphics processing units (GPUs) and AI accelerators. ^[74] Also central are cloud service providers, such as Amazon Web Services (AWS), Google Cloud Platform (GCP), and Microsoft Azure, that offer computing resources. Also vital at this infrastructure level are data-gathering companies that harvest public data or companies that own proprietary data. ^[75]

Below these operators, one may find the development layer where leading operators (such as Google, OpenAI, Meta, Anthropic, and Mistral AI) develop proprietary AI models and “open access models,” which we elaborate on later.

These pretrained models are deployed downstream via download or API access through applications or dedicated programs.

In a competitive market, each layer in the foundation model supply chain would have multiple rivals, and no rival would operate on multiple layers. For example, many semiconductor firms would provide the accelerator computer chips with the computing power needed to train the foundation model. Another set of companies would provide the data required to train the foundation model. A third set of companies would offer computational power and data centers. A fourth set would develop an array of AI foundational models. Next are the hosts who make these models available to developers who use the foundation model for their apps, products, or services. Businesses and customers could switch among the foundation models and apps, which are interoperable with each other.

However, as we explore below, once we relax these assumptions, antitrust risks emerge when these stacks become concentrated and powerful firms vertically integrate along the supply chain. Due to economies of scale and scope, feedback loops, high entry barriers, and the need for significant investments, critical segments of the supply chain can become highly concentrated. While the foundation model supply chain has multiple levels, its evolution has already led to significant vertical integration, which raises numerous concerns, including self-preferencing. ^[76]

As the G7 has noted, “[a]n inability to access these key inputs may inhibit competition to develop AI and AI applications, reducing innovation and harming consumers.” ^[77] As the FTC has similarly observed,

If a single company or a handful of firms control one or several of these essential inputs, they may be able to leverage their control to dampen or distort competition in generative AI markets. And if generative AI itself becomes an increasingly critical tool, then those who control its essential inputs could wield outsized influence over a significant swath of economic activity.” ^[78]

Let us consider these antitrust risks by exploring several critical layers in the stack.

B Concentration of Foundation Model Developer Level

Training AI foundation models, as of 2024, generally requires the following inputs:

• Significant volume and variety of data to train the generative AI foundation model, fine-tune it, and provide up-to-date responses; ^[102]

• Significant large-scale computational resources, including cloud computing resources, with specialized chips; ^[103]

• Significant amounts of energy; ^[104] and

• Human capital, including “a skilled engineering and research workforce with expertise in areas such as machine learning and natural language processing.” ^[105]

Thus, the early advantage in developing foundation models will go to firms that already have a significant volume and variety of data (e.g., hundreds or thousands of gigabytes of data ^[106] across different modes) to train these models, the large-scale computational resources (either internally or committed cloud computing resources), and the human capital, including the human feedback needed to fine-tune the model’s output (such as preventing biased, false or harmful outputs). ^[107]

So, who has these advantages? The Big Tech Barons, who in the first half of 2024 spent collectively over $100 billion in capital improvements related mainly to AI. ^[108] Overall, the Tech Barons controlled multiple foundation models: by 2023, Google had 38 models (28 of which were closed); OpenAI had 18 models; Microsoft had an additional ten models (six of which were closed); and Meta had ten models (two of which were closed). ^[109] As of 2023, the higher-performing models were closed-source. ^[110] The significant investment needed to develop the technology supports ongoing concentration, ^[111] underscoring the risk of further entrenchment of the Big Tech Barons. We are already seeing OpenAI (through its partnership with Microsoft), Google, and Meta develop their own foundation models, from which other firms may fine-tune their AI applications, products, and services. As the company reports, Google’s Gemini foundation model was the first “to outperform human experts on MMLU (Massive Multitask Language Understanding), one of the most popular methods to test the knowledge and problem-solving abilities of AI models.” ^[112] Gemini also powers Google Bard. ^[113]

The Big Tech Barons are investing not only in their own foundation models and AI infrastructure but also in other foundation models. Microsoft has a partnership with OpenAI’s ChatGPT ^[114] and DALL-E ^[115] and is coordinating with Meta on its Llama-2 foundation model. ^[116] Microsoft reportedly invested over $13 billion in OpenAI by 2023. ^[117] Google and Amazon have invested over $2 billion and $4 billion, respectively, in Anthropic. ^[118] Microsoft, Nvidia, and others have invested over $1.3 billion in Inflection AI, ^[119] and Elon Musk’s plans to raise $1 billion in investment in the startup “xAI.” ^[120]

The UK’s competition authority has illustrated the Big Tech Barons’ myriad investments and strategic partnerships in the other foundation models: ^[121]

Source: CMA AI Foundation Models Update Paper 2024

Consequently, the Big Tech Barons have strategically collaborated with the leading independent foundation models.

Also significant is how the Tech Barons are vertically integrating across the supply chain, leveraging their relative advantages over access to data, cloud servers, and foundation models to control the downstream consumer and business interfaces.

This privileged position is unique to the Big Tech Barons. It distinguishes them from other companies, even successful ones, that, while competing with their models, also heavily rely on the Tech Barons for financial support, computing power, or data. Here a frenemy relationship may evolve, ^[122] where the Tech Barons both cooperate with, and compete against, smaller operators that develop foundation models and apps and services that rely on these models. These dynamics may entrench the Tech Barons’ power, widening the moat that insulates them from future disruption.

Consequently, the Big Tech Barons, as the following illustration from the CMA reflects, are vertically integrating: first, they are vertically integrating their foundation models throughout their ecosystem of existing digital products and services (such as their foundation models integrated into their search engines, personal computer operating systems, and social media). ^[123] Second, the Tech Barons are integrating more broadly across the foundation model supply chain—developing their own accelerator and inference chips, ^[124] expanding their data centers, ^[125] and entering into strategic partnerships with other foundation models: ^[126]

Source: CMA AI Foundation Models Update Paper 2024

As the Tech Barons are integrating across the foundation model supply chain and integrating their own foundation models into their downstream products and services, let us next assess the potential impact this vertical integration can have on other downstream competitors and market participants.

IV Countervailing Factors

Supply chains can change (consider the supply chain changes after Covid). So too, the AI foundation model supply chain can change, depending on several factors. One factor will be the continued importance of the inputs for the foundation model. Research by OpenAI, among others, has found power-scaling laws that “increasing model parameters, dataset size and compute with transformers and cross-entropy loss increases the model performance.” ^[172] But to reach optimal scaling in the future, one might not require, for example, the breadth and volume of data. ^[173] Moreover, the availability of computational resources may increase while computing costs may decline. With some predictions that open-source models will continue to improve, ^[174] the cost of computing and the cost of doing business will likely decrease. This openness can lower costs and barriers to expansion and entry and, in doing so, level the playing field. ^[175]

Another factor will be the proliferation and scale of other foundation models advanced worldwide. Notable in the West ^[176] will be other key AI platforms, such as Stable Diffusion, which describes itself as the “world’s leading open source generative AI company” that delivers “breakthrough, open-access AI models with minimal resource requirements in imaging, language, code and audio.” ^[177] Europe’s Mistral AI (which includes Microsoft, Amazon, and Google as investors) is seeking to develop a quasi-open-source foundation model. ^[178] According to the company, its foundation model cost about $22 million to train and performed similarly to OpenAI’s and Google’s models, which cost far more to train (between $50 and $100 million for OpenAI). ^[179] The extent to which developers can easily switch among foundation model platforms can potentially dampen the anticompetitive effects.

Third is the openness (even if limited at times) and accessibility of the foundation models and the ability to use these models to disrupt both within and outside the foundation model’s ecosystem. A notable aspect of several commercial providers of generative AI has been their effort to make the technology accessible to others. AI systems, cloud computing, and open-source repositories have offered opportunities for growth for those who would otherwise have no, or limited, access to the technology.

Fourth is the potential commoditization of foundation models. As of 2024, the emphasis is on building and training the foundation models. Thereafter, developers using their data can fine-tune the foundation model for their particular applications, and for inferences, when the model is asked to process new information and respond. ^[180] From 2023 to early 2024, 40 percent of Nvidia’s data center business was for deploying AI systems, not training. ^[181] But going forward, in the fine-tuning stage, developers may not require Nvidia’s accelerator chips and might be able to rely on less powerful, cheaper chips. ^[182] So, one issue will be the relative importance of the foundation model versus the fine-tuning and inference stages.

Fifth, and related to the above, is the democratization effect of the technology. Foundation models and generative AI can lower barriers to entry and enable new entities to enter markets and provide services where they could not have done so before. The technology promises to reduce costs associated with generating and editing content. It can enable reliance on less data to execute tasks, lower data collection costs, lower investment costs, reduce the need for human input, and support further automation. ^[183] It can also assist in faster and more efficient software development by improving code quality and generating new codes, and in doing so, opening the software development market to new players. ^[184]

Citing these competition and innovation dynamics, some argue against antitrust intervention and call for limited regulations on AI systems. ^[185] After all, in a fast-moving innovation market, intervention could easily miss the mark and chill, rather than promote, competition. Competition dynamics, it is argued, could offer a superior outcome and ensure the effective evolution of the technology and markets.

V Refining the Antitrust Playbook

Over the past two decades, we have seen many digital markets tip to one or two powerful firms. We have also seen how these powerful firms expanded their ecosystems. Positive feedback loops, control over a broad user base, extraction of personal data, and anticompetitive practices have enabled a few corporations to entrench their power and affect the dynamics of competition and innovation. Digital markets, while dynamic in many ways, have proven to not easily self-correct once markets tip in favor of these Tech Barons. Overall, experience shows that the digital economy gravitates toward concentrated economic power, which often translates to political power.

The emerging AI foundation model supply chain can disrupt these dominant ecosystems. But, as with the digital economy generally, a few firms could use the positive feedback loops to ensure that more developers rely on their foundation models, thereby strengthening their ecosystems. Indeed, in controlling the critical inputs and leading foundation models, the next wave of Tech Barons may have even more power: not only will they control the leading AI app stores, but they will also control the underlying AI technology, data inputs, data storage, and hardware design. The risks are immensely greater: biases, flaws, and vulnerabilities in these foundation models can ripple across society.

So, while the Biden administration has stepped up antitrust enforcement considerably against the Big Tech Barons, the antitrust enforcement has not appreciably increased contestability in many digital markets. Antitrust enforcement in the U.S. remains unwieldy, with costly, time-consuming litigation under the Supreme Court’s rule of reason legal standard. To put it bluntly, antitrust has failed to curb the excesses of the Big Tech Barons, and we are paying for the result.

We are at another critical juncture: the AI foundation model supply chain can potentially tip to where a few powerful firms can again chill competition and innovation. But that is not preordained. Again, the issue is whether antitrust enforcement can prevent this result.

In recent years, several jurisdictions have upgraded their antitrust and regulatory tools for the digital economy. Notable is the European Union with its Digital Market Act, Digital Services Act, Data Act, and forthcoming AI Act.

So, what needs to be done? Along with many others, we are exploring avenues to enhance the effectiveness of competition law and related regulations to align the market participants’ incentives with broader societal goals. Broadly speaking, multiple tools are necessary to ensure a healthy market trajectory (including competition, privacy, consumer protection, and IP laws).

At the competition law level, for antitrust to be more effective, the agencies and courts must better understand the power of ecosystems (as opposed to the power of particular platforms), including their power to distort innovation paths. Ecosystem power differs from traditional notions of market power. ^[186] Consequently, agencies and courts must focus on the possible abuse of ecosystem power aimed at entrenchment through leveraging, tying, and exclusionary practices (even when the market is not yet concentrated). Structural separation may need to be explored more often in the current state of AI supply chains.

Consideration of the scope of competition law and the use of external regulatory benchmarks to inform competition decision-making would be of value. ^[187] This is particularly so since intervention will undoubtedly dictate tradeoffs between efficiency and economies of scale and scope, and the need to safeguard the evolution of AI markets and ensure their contestability.

At the regulatory level, the current design and scope of the EU’s Digital Markets Act and Digital Services Act offer some, yet incomplete, relief regarding the concerns raised by foundation models. The forthcoming AI Act could supplement the regulatory structure when dealing with the risk associated with foundation models but does not directly address the problem of entrenchment and increased concentration. Changes may be required to improve the relevance of regulatory tools to address the AI supply chain. In that respect, ex-ante regulation could be important in aligning private corporate behavior with public interests, even though entrenched interests will likely challenge any intervention. When considering the dense EU regulatory landscape, it is important to note that the wider EU regulatory framework concerning data, privacy, and political stability may generate inconsistent effects for the AI supply chain. ^[188]

Possible changes to regulatory tools would need to consider the overall burden on undertakings, the possible externalities between jurisdictions, the interconnections between trade policies and competition law, and the potential impact on foreign direct investments.

Enforcement capacity at both the regulatory and competition levels must increase to match the level of business activity. Global cooperation and alignment are integral and necessary to ensure effective outcomes. No jurisdiction can shoulder the regulatory framework.

Support policies should be used to instill growth and innovation in AI. The EU has a role in creating the conditions and infrastructure for EU innovation and disruption. Financial grants, tax breaks, loans, guarantees, and capital investments can play an important role in increasing the competitiveness of the AI supply chain. Policies should facilitate technology transfers, commercialize innovations, facilitate access to input, and strengthen available human capital. These efforts should be guided by the “Value, Incentives, and Diversity” principles we have explored elsewhere. ^[189]

Finally, unregulated competition is not an elixir. We have already seen companies racing to release foundation models, even though the models suffer from biases, hallucinations, and other quality problems. Competition can pressure firms to release faster foundation models that can undertake even more services without mitigating the models’ potential risks. Deep-pocketed companies are racing to release new technologies and disregarding many emerging risks and concerns in their attempts to dominate the new space. Calls for regulation, or slowing down of some developments, are being met with warnings about the competitive disadvantages to Western countries who regulate and the risk that other less regulated jurisdictions may take the technological lead. Encouraging this toxic competition can cause even greater harm to society. Consequently, the antitrust enforcement challenges raised by generative AI are further amplified by competition of the wrong sort—namely, a race to the bottom that emerges between companies. Enforcement, through competition law or regulation, should be mindful of the nature and quality of rivalry and the innovation it fosters: namely, we want to promote a race to the top and not to a potentially horrific bottom.

Conclusion

The AI landscape has changed in the short period we drafted this Article and will continue to change over the next decade. Ultimately, two potential antitrust risks emerge: first, as we have seen, several factors and market characteristics have enabled a few Big Tech Barons to expand their ecosystems, maintain their dominance, and alter innovation paths to marginalize potential disruptive threats. Likewise, similar factors in the emerging AI foundation model supply chain can lead to a “winner-take-most-or-all” situation. Second, while AI may herald significant disruption, the risk is that it will not disrupt the dominant ecosystems. Instead, several of the current dominant ecosystems, in seeking to capture essential parts of the AI foundation model supply chain, can minimize the risk of such disruption. As consumers, workers, and market participants, we will pay the price.

Antitrust enforcers are rightfully alert. Failure to move swiftly at the substantive and enforcement levels may leave us with a repeat of past policy shortcomings. AI will only magnify the current legislative and policy deficits. And the results, as the AI experts and the public predict, can harm our wellbeing, autonomy, and democracy.