The Little Monopoly in the Digital Clouds Cyber Law Firm

Written By Andrew Dang

I. Introduction

In the rapidly evolving landscape of artificial intelligence, tech giants are forging unprecedented partnerships that challenge the foundations of antitrust law. At the forefront of this paradigm shift is Microsoft, whose strategic alliances with AI powerhouses like OpenAI are reshaping the competitive dynamics of the digital economy. These collaborations, while driving technological innovation, raise critical questions about market power, data monopolies, and the adequacy of current regulatory frameworks.

Microsoft’s and Google’s AI partnerships, particularly their multi-billion-dollar investment in OpenAI and Anthropic, respectively, represent more than mere financial transactions. They embody a new form of market consolidation that transcends traditional notions of mergers and acquisitions. By gaining exclusive access to cutting-edge AI models and vast datasets, these large tech companies are positioning themselves at the nexus of a data-driven ecosystem that smaller competitors may find increasingly difficult to penetrate.

This post argues that Microsoft’s and Google’s AI partnerships challenge traditional antitrust frameworks by creating data monopolies and vertical integrations that current laws struggle to address, requiring a new approach to competition regulation in the era of generative AI. As foundation models become the new currency of the digital age, the concentration of computational power, data, and AI knowledge in the hands of a few tech giants threatens to distort market competition and raise formidable barriers to entry for newcomers.

The implications of these partnerships extend far beyond the tech sector, touching on issues of consumer privacy, market innovation, and the distribution of power in the digital economy. As we will explore, the current antitrust paradigm, with its focus on consumer welfare and price effects, is ill-equipped to grapple with the nuanced challenges posed by AI collaborations that may not immediately result in higher prices but could lead to long-term market distortions.

This analysis will examine the economics of foundation models, the potential for predatory pricing in AI services, and the creation of data monopolies through vertical integration. We will also scrutinize the limitations of current antitrust frameworks, drawing comparisons with other highly regulated industries like pharmaceuticals. Finally, we will consider potential regulatory approaches, including the UK Competition and Markets Authority’s principles of Access, Diversity, Choice, and Fair Dealing, as potential guiding lights for a new era of antitrust regulation in the age of AI.

As the line between competition and collaboration in the tech industry grows increasingly blurred, the need for a nuanced, forward-looking approach to antitrust law has never been more urgent. The decisions made today about how to regulate AI partnerships will shape the landscape of innovation, competition, and consumer choice for decades to come.

II. Background on Microsoft’s and Google’s AI Partnerships

After OpenAI’s success with ChatGPT, Microsoft has invested $13 billion into OpenAI and has also integrated OpenAI’s GPT-4 into their web search and office products. Despite having exclusive access to OpenAI’s top proprietary models, Microsoft remains unsatisfied. exclusive On June 29th, 2023, Inflection AI, another AI startup, announced its $1.3 billion seed funding led by Microsoft. But Microsoft’s and Inflection’s AI partnerships lack transparency. Part of the deal involved retaining co-founders Mustafa Suleyman and Karen Simonyan, along with most of the 70-person team, for a newly created consumer AI unit called Microsoft AI.

Google has also entered GenAI race with their own foundation model, Gemini. Like Microsoft’s large major investment in OpenAI, Google has chosen Anthropic’s Claude to compete with OpenAI’s GPT. And like Microsoft and Inflection AI, Google has hired the founders of Character AI, Noam Shazeer and Daniel De Freitas, and secured an exclusive license to use Character AI models. “If the [Character AI] deal had been a conventional acquisition, a transaction of that size would be Google’s sixth-largest buyout in its history.” The removal of founders from AI start-ups highlights the difficulty of building successful AI companies. AI Partnerships such as Google and Microsoft offer exclusive billion-dollar deals to help FM providers offset high development costs such as model training and inference.

III. The Economics of Foundation Models

A. Cost of Training and Deployment

Foundation models (FMs) are trained on broad data (generally using self-supervision at scale) that can be adapted (e.g., fine-tuned) to a wide range of downstream tasks. The model starts with generic representations learned from vast amounts of text data. After pre-training, organizations fine-tune their models on task-specific datasets for tasks such as text summarization, translation, and question-answering. Yet achieving high performance often comes with the cost of substantial computational resources.

Training foundation models (FM) like large language models requires a significant investment. In 2017, training the Transformer model cost $900. Fast forward to 2019, the expense of training RoBERTa Large escalated to $160,000. Scaling laws, for large models, show that model performance improves predictably with computational power used. In 2023, the price tag for models like OpenAI’s GPT-4 and Google’s Gemini Ultra rocketed to an estimated $78 million and $191. The dramatic increase in training cost is due to the parameter model size. Even after spending billions on training, the cost of servicing foundation models is resource intensive.

Depending on the model size and complexity of the task, servicing a single model can cost from $2,149.20 to $17,193.60 a month. Deployment via an API is a selling point for FM providers, as the expenses associated with training and deploying large language models can be prohibitive for the average company. APIs act as the central nexus linking consumers to AI service providers, rather than paying upfront for training and deploying (which usually includes renting expensive, cutting-edge graphical processing units (GPUs) in the cloud) large language consumers are charged based on token counts when they access these models via third-party APIs. Tokenization is segmenting input text and model outputs into smaller units called tokens that LLMs can process. Under this “token economy,” businesses and developers can integrate GenAI models into their applications, while FM provider firms can monetize their models by charging tokens.

B. Pricing Strategies and Potential Predatory Pricing

Despite the computational cost of training and servicing these models, companies such as OpenAI and Google are consumers at a fraction of the cost. Google and Microsoft’s large share in cloud computing, it offers favorable deals to FM providers like OpenAI. This integration allows for lower pricing of AI services, despite high inference costs. When GPT-4 was introduced, it was priced significantly higher than its predecessors. In 2023, OpenAI GPT-4 was $30.00 per 1M Tokens, however, as competition increase prices goes down. Now, OpenAI’s GPT-4 costs $5.00 per 1M Tokens. To compete with OpenAI’s prices, Google has made its Gemini 1.5 model free, charging nothing per 1M tokens.

Predatory pricing comes into play when a dominant firm deliberately reduces its prices to a loss-making level for a short-term to discipline its existing competitors or foreclose the market to new entrants to strengthen or maintain its market power later by way of the foreclosing effect of this predation. OpenAI’s position as a leader in the generative AI market, backed by Microsoft’s resources and cloud infrastructure, puts the company in a favorable position to recoup losses.

While OpenAI and Google can reduce costs through methods like quantization or AI chip optimizations, these efficiency gains may not fully account for the dramatic price reduction from $30.00 to $5.00 per 1M tokens. The vertically integrated nature of large tech companies in the AI space will create barriers of entry for smaller competitors. To meet market expectations, new players are not only expected to develop models that compete with Google and OpenAI but are also expected to match their competitive pricing. “Big tech can afford to scale without strong economics, which dampens the ability of smaller companies to compete.”

Antitrust Concerns

IV. Antitrust Concerns

A. Data Monopolies and Barriers to Entry

The foundation models nested under Azure Cloud grants Microsoft a data monopoly that harms the competitive process. FMs are representations of their training data, just as AI algorithms are representations of consumer data. While digital mergers typically focus on acquiring proprietary datasets for AI training, FMs uniquely encompass both the proprietary datasets and the algorithms, effectively combining the raw data and the trained product offered to consumers. Here, Microsoft’s partnerships grant Azure Cloud access to exclusive algorithms.

To add on, AI systems exhibit “agentic” behavior; acting autonomously on users’ behalf to achieve set goals while having access to systems that use design patterns such as tool use or planning. These tools include web search, code execution, and more. Agents have a level of autonomy and interaction increases the amount of personal information extracted from consumers during conversations. The data extracted from FMs are more comprehensive than that of cookies and click adds. The collected data is then used to train further models.

If the concern with mergers is the exclusion of firm data being a barrier to entry, then competitor authorities must act against Microsoft’s AI partnerships. Large tech companies with vertical integration dominate the generative AI market.

B. Vertical Integration and Market Distortion

The rapid advancement of Generative Artificial Intelligence GenAI has brought both opportunities and legal challenges to the forefront of the tech industry. Recent years have seen a surge in legal actions against major AI companies, exemplified by high-profile lawsuits from content creators and media giants. These legal battles, coupled with emerging regulations like the EU AI Act, have created a complex landscape that favors established players while posing substantial barriers to new entrants in the AI market.

In addition, a significant part of accessible data is now restricted, either through technical means or legal terms of service. This shift not only limits the raw material for AI training but also threatens to skew the representativeness and effectiveness of future AI models. FM providers unauthorized data practices have resulted in “The Rapid Decline of the AI Data Commons.” Creators’ works are no longer open, but their works are instead locked behind paywalls.

The emergence of FMs coincides with a period where Big Tech has amassed advantages. Big Tech’s influence across various levels of the AI ecosystem. This position lets them shield themselves from AI-induced disruption or leverage it for their specific benefit. Big tech strengthens or cements the dominant positions they gained during the previous major technological shift, potentially hampering future competition.

C. Privacy and Consumer Protection Issues

FMS are in and of themselves proprietary datasets bundled with consumer information. Foundation models raise privacy concerns like those associated with big data algorithms, but with potentially more severe consequences. When firms use consumer data to train models through a singular platform, it becomes easier to create a comprehensive consumer profile. Advanced AI models trained on vast amounts of consumer data can predict behaviors and preferences with increasing accuracy. This predictive power further reduces consumer autonomy in decision-making. The merger of data from different sources can lead to the revelation of personal information. Also, the risk of data breaches is exasperated when data is centralized on a single platform. Data centralization makes it harder to maintain security and compliance.

V. Limitations of Current Antitrust Framework

A. Focus on Mergers vs. Partnerships

Microsoft and other Big Tech corporations evade prosecution because antitrust laws are predicated on mergers. On June 28, 2024, the European Union (EU) ruled out an investigation under the EU’s merger rules. Nevertheless, the EU is investigating Microsoft’s exclusivity clause with OpenAI and whether the clause has a negative impact on competition. Similarly, the United States has taken a “mix-and-match” approach to overcome the mergers requirement in Section 7 of the Clayton Act, using Section 2 monopolization principles to challenge acquisitions of emerging competitors on grounds of maintaining monopolies through acquisition, while also implicating Section 7 as the basis for potentially undoing such transactions.

Section 7 of the Clayton Act combats anti-competitive mergers by enabling oversight of transactions, granting firms control over competitors’ assets. Even still, the nature of Microsoft’s AI partnerships are not mergers. Thus, Microsoft side steps prosecution under Section 7 of the Clayton.

B. Courts’ Hesitation to Intervene in Tech Innovation

In recent months, antitrust cases against major tech platforms in the United States have been mounting, with almost every facet of their conduct under scrutiny or challenge However, U.S. antitrust law traditionally favors the freedom of independent action by firms, including dominant ones, particularly those that excel through innovation. Throughout the courts may invoke the language in Microsoft:

“We may infer causation [of anticompetitive effects] when exclusionary conduct is aimed at producers of nascent competitive technologies as well as when it is aimed at producers of established substitutes . . . . [It]would be inimical to the purpose of the Sherman Act to allow monopolists free rein [sic] to quash nascent, albeit unproven, competitors at will.”

But the Trinko ruling has made courts more likely to defer from intervening in business innovations. The Court expressed concern about the risk of mistakenly condemning pro-competitive or neutral conduct as anticompetitive. Such errors could chill legitimate business conduct and innovation, which antitrust laws are designed to protect. “As a general rule, businesses are free to choose the parties with whom they will deal, as well as the prices, terms, and conditions of that dealing.” As a result, the court’s decision in Trinko has led to a more cautious approach in antitrust enforcement, particularly in cases involving one-sided conduct by monopolists. Courts are now more likely to avoid intervention unless there is clear evidence of anticompetitive behavior.

Ironically, Trinko ruling did have a chilling effect on antitrust law. Section 2 case law hesitates to second guess business decisions and usually sees innovation as best protected by court non-intervention. The Trinko decision enables a high-tech monopolist to dominate its whole product ecosystem, beyond just competitor interactions. For example, in Qwest Corp., the Ninth Circuit held that Qwest’s refusal to deal did not violate Section 2, emphasizing that “a monopolist’s refusal to deal is not inherently anticompetitive unless it lacks a legitimate business justification and is intended to harm competition”.

In Novell, Inc. v. Microsoft Corp., the Tenth Circuit ruled for Microsoft, finding that “antitrust laws do not require a company to assist its competitors,” and that Novell did not show that Microsoft’s conduct had an anticompetitive effect on the market under Section 2. Last, in Honeywell Int’l, Inc., the Ninth Circuit held that “a monopolist’s refusal to deal is not inherently anticompetitive unless it lacks a legitimate business justification and is intended to harm competition,” and found that Honeywell’s actions were driven by legitimate business considerations under Section 2.

The patterns in Section 2 caselaw suggest that courts will probably continue to favor non-intervention in the context of the Microsoft and OpenAI partnership. Courts will underscore legitimate business justifications and the need for clear evidence of anticompetitive intent and effect. This judicial approach aligns with the belief that innovation and business decisions are best protected by allowing companies to operate freely, unless there is substantial proof of harm to competition. And even if the courts intervene, the time to bring and win a case means Microsoft may become irreversibly entrenched before any action could be taken.

Antitrust straddles the clash between protecting consumers and fostering competition. Authorities argue that the core aim of these laws is consumer safety. Yet, predominantly in the U.S., the focus shifts to boosting competition and sparking innovation. This internal conflict within antitrust rules needs resolution to address upcoming challenges in generative AI. Antitrust laws must strike a nuanced balance between consumer rights and competitive practices.

VI. Proposed Solutions

In AI and digital markets, the UK Competition and Markets Authority (CMA) upholds ‘Access’, ‘Diversity’, ‘Choice’, and ‘Fair Dealing’ as pillars of a competitive, innovative environment. ‘Access’ ensures resources like data and technology are available to both new startups and established giants, leveling the playing field and igniting innovation. ‘Diversity’ and ‘Choice’ provide consumers with a wealth of options, allowing them to shape the market. ‘Fair Dealing’ preserves integrity, making sure competition is based on merit rather than sheer power.

Another possibility is to a new antitrust standard to this complex knowledge economy. Antitrust law engages biopharmaceutical industries with unique considerations due to the industry’s reliance on patents and regulatory frameworks. Pharmaceuticals rely on patents to recoup the investments required for drug research and development. Patents grant exclusive rights to the holder, which can seem to conflict with antitrust principles aimed at promoting competition. Similarly, FM providers have proprietary datasets for a competitive advantage in developing superior FMs. But when abused, these patents become anti-competitive and limit market entry for generic drugs and keep prices high for consumers. Similar, in the AI industry, exclusive access to proprietary datasets gives companies a competitive edge in developing competing models. While exclusivity is essential for driving innovation, the lack of transparency becomes anti-competitive if misused, such as by restricting data access or engaging in practices that unfairly stifle competition, ultimately harming consumer choice and market fairness.

Conclusion

Addressing the inherent tension within antitrust laws is important as we navigate the complexities of generative AI and digital markets. The UK CMA’s focus on ‘Access,’ ‘Diversity,’ ‘Choice,’ and ‘Fair Dealing’ provides a robust framework to balance consumer protection with competitive innovation. Still, new antitrust standards are appropriate to handle industry-specific challenges, such as the reliance on patents in biopharmaceuticals or the exclusive datasets in AI. Ensuring these competitive advantages are not abused will help maintain market fairness and protect consumer interests, fostering an innovative and equitable marketplace.

Citation

1- Microsoft, Apple walk away from OpenAI board, Australian Financial Review (2024), https://www.afr.com/technology/microsoft-apple-walk-away-from-openai-board-20240710-p5jsm8 (last visited Aug 26, 2024).

2- Google buys out Character.ai founding team, stoking antitrust concerns – The Washington Post, https://www.washingtonpost.com/technology/2024/08/02/google-character-ai-noam-shazeer/ (last visited Aug 26, 2024).

3- AI Index Report 2024 – Artificial Intelligence Index, https://aiindex.stanford.edu/report/ (last visited Aug 11, 2024); Jordan Hoffmann et al., Training Compute-Optimal Large Language Models, (2022), http://arxiv.org/abs/2203.15556 (last visited Aug 11, 2024).

4- Lambda GPU Cloud | 1-Click Clusters, https://lambdalabs.com/service/gpu-cloud/1-click-clusters (last visited Aug 26, 2024).

5- What are tokens and how to count them? | OpenAI Help Center, https://help.openai.com/en/articles/4936856-what-are-tokens-and-how-to-count-them (last visited Nov 27, 2023).

6- How much does GPT-4 cost? | OpenAI Help Center, https://help.openai.com/en/articles/7127956-how-much-does-gpt-4-cost (last visited Jul 31, 2024); Pricing, https://openai.com/api/pricing/ (last visited Jul 31, 2024).

7- How FlashAttention-2 Accelerates LLMs on NVIDIA H100 and A100 GPUs, (2023), https://lambdalabs.com/blog/flashattention-2-lambda-cloud-h100-vs-a100 (last visited Nov 28, 2023); Tim Dettmers et al., QLoRA: Efficient Finetuning of Quantized LLMs, arXiv.org (2023), https://arxiv.org/abs/2305.14314v1 (last visited Oct 9, 2023); Scaling AI, Center for Security and Emerging Technology, https://cset.georgetown.edu/publication/scaling-ai/ (last visited Jul 31, 2024).

8- Google buys out Character.ai founding team, stoking antitrust concerns – The Washington Post, supra note 3.

9- Sayash Kapoor et al., AI Agents That Matter, (2024), http://arxiv.org/abs/2407.01502 (last visited Jul 10, 2024).

10- Junyou Li et al., More Agents Is All You Need, (2024), http://arxiv.org/abs/2402.05120 (last visited Jul 10, 2024).

11- Shayne Longpre et al., Consent in Crisis: The Rapid Decline of the AI Data Commons.

12- U.S. v. Microsoft Corp. 253 F.3d 34 (D.C. Cir. 2001).

13- Verizon Communs., Inc. v. Law Offices of Curtis V. Trinko, LLP, supra note 1.

14- MetroNet Servs. Corp. v. Qwest Corp., 383 F.3d 1124 (2004).

15- Aerotec Int’l, Inc. v. Honeywell Int’l, Inc., 836 F.3d 1171 (2016).

Graphs are credited to https://www.linkedin.com/in/peter-gostev/

I. Introduction