Why AI Companies Are Racing to Control Their Supply Chains

Table Of Links

Abstract

Executive summary

Acknowledgments

1. Introduction and Motivation

• Relevant literature
• Limitations

2. Understanding the AI Supply Chain

• Background history
• Inputs necessary for development of frontier AI models
• Steps of the supply chain

3. Overview of the integration landscape

• Working definitions
• Integration in the AI supply chain

4. Antitrust in the AI supply chain

• Lithography and semiconductors
• Cloud and AI
• Policy: sanctions, tensions, and subsidies

5. Potential drivers

• Synergies
• Strategically harden competition
• Governmental action or industry reaction
• Other reasons

6. Closing remarks and open questions

• Selected Research Questions

References

5. Potential drivers

Several potential reasons may make companies in the AI supply chain vertically integrate, create strategic partnerships, or refrain them from doing that. In this section, we provide an overview of what may be driving these patterns.

5.1 Synergies

Vertical integration and partnerships can lead to cost savings and operational efficiencies by consolidating various stages of the supply chain. This integration potentially allows companies to leverage shared resources, infrastructure, and expertise. As the AI industry has high fixed costs, this is probably one of the main drivers. Moreover, vertical integration helps in reducing transaction costs and improving coordination between different stages of production.

\ This is crucial to avoid problems in R&D projects, as mentioned by Acemoglu et al. (2010). In the AI supply chain, this is seen in the interaction between companies that own manufacturing facilities and chip designers. These transaction costs are both contractual and technological. For instance, companies need to communicate and coordinate closely to develop e.g. a new chip design that utilizes EUV technology to the fullest.

\ An example of that is the partnership between ASML, TSMC, NVIDIA, and Synopsis in developing cuLitho, a software tool being developed to use NVIDIA’s GPUs to optimize ASML’s lithography technology. Furthermore, vertical integration and partnerships may also allow companies to acquire talent and technology capabilities by integrating complementary expertise from different stages of the supply chain. This may be an explanation for large technological conglomerates that overreaches a wide range of industries (Chen, Elliot & Koh, 2023) and may also play a role in their recent partnerships and acquisitions of AI labs.

\ Vertical integration serves as a strategic move for securing essential inputs necessary for AI development. Given the inelastic short-term supply of frontier AI accelerators, companies strive to ensure access to these essential inputs. By taking direct control over various steps of the supply chain, companies can mitigate the risks associated with supply chain disruptions, as highlighted in Elliot & Golub (2022). Many companies seek assurance that they will have access to key inputs, such as compute, because they believe the supply is constrained in the medium run.

\ To secure these resources, they either gain more control over the supply chain or establish strategic partnerships with key suppliers. This dynamic may intensify due to the perceived winner-takes-all nature of the AI industry. However, as a company grows bigger, the benefits from increasing its size decrease since it becomes harder to manage. Even though this can be lessened by having different corporate structures where subsidiaries operate independently, it remains a major factor preventing companies from integrating too much, both in the AI industry and across the economy.

\ As companies are not capable of specializing in everything but still seek to accumulate capabilities, they potentially see quasi-vertical integration as a flexible solution for that. This potentially is important to understand the strategy of big tech companies such as Alphabet and Microsoft in the AI industry: they seek to have a stake in major AI labs and integrate their technologies in their diversified portfolios of products, but leaving substantial autonomy to the emergent companies.

\ 5.2 Strategically harden competition

Firms may aim to strategically create entry barriers in a market to maintain dominant positions. By doing so, they deter potential competitors from entering the market, thus preserving their market share and potentially their pricing power. In a similar vein, companies may attempt to foreclose access to essential inputs for other firms, as discussed by Patrick & Tirole (2007).

\ The Federal Trade Commission has blocked the acquisition of Arm Limited by Nvidia mainly under this concern (FTC, 2021). Furthermore, by retaining control over key areas of the business, firms can avoid sharing sensitive data that may otherwise be necessary in more collaborative arrangements. This helps maintain a competitive edge and safeguard proprietary or sensitive information from potential rivals, especially in high-tech sectors, as discussed by Barrera (2019). Firms may also engage in killer acquisitions and capability hoarding to further secure their competitive positioning by either acquiring potential competitors or hoarding critical capabilities to prevent others from accessing them (Cunningham et al, 2019; Boa et al, 2023).

\ 5.3 Governmental action or industry reaction

Governments may offer incentives to encourage vertical integration, especially in strategic industries. These incentives can range from tax benefits and subsidies to preferential treatment in procurement or regulatory advantages. This may be an increasingly important driver as the semiconductor industry is increasingly seen as a matter of national security. Integration may be avoided in the AI industry to mitigate antitrust concerns.

\ Companies involved in the industry may be wary of controlling too much of the supply chain because it could potentially be a concern raised by antitrust authorities. The importance of this is directly impacted by their expectations that this will be an issue for regulators, as discussed by O’Keefe (2021). Compliance with specific regulations, such as data privacy or security requirements, can be facilitated through vertical integration.

\ The integration allows companies to have better control over data flows, risk management, and adherence to regulatory frameworks such as the EU General Data Protection Regulation (Gal & Aviv, 2020; Carugati, 2023), which may have a detrimental effect on smaller firms and competition overall (Campbell et al, 2015)

\ 5.4 Other reasons

As the AI industry is still in a nascent stage, the market for specific services is not that well developed and it is difficult to do major, impersonal transactions since there are no established ways of working. This scenario often drives companies towards vertical integration to secure and streamline operations. Some authors have argued that in early-stage development of general-purpose technologies it is common to see a high degree of vertical integration because of this, followed by vertical separation as markets develop.

\ Additionally, past business decisions, investments, and established relationships significantly influence the inclination to pursue vertical integration. Companies may choose this path to build upon existing capabilities, intellectual property, or market positioning, thereby leveraging established foundations for growth or competitive advantage. Through this lens, both the early stage of the industry and historical path dependencies play crucial roles in shaping the strategic choice towards vertical integration in the AI sector.

\ Finally, the patterns we see in the industry may be driven by a growing belief, especially in AGI labs, that the frontier AI industry will exhibit a winner-takes-all dynamic. For instance, Dario Amodei, CEO of Anthropic, mentioned that the current cost of training leading AI models is around $100 million but is expected to escalate to as much as $5 billion to $10 billion by 2025 or 2026. He attributes this sharp increase to the scaling laws of AI (Decoder, 2024). The continuous escalation of compute clusters will only raise the fixed costs necessary to enter the industry.

:::info Author:

Tomás Aguirre

:::

:::info This paper is available on arxiv under CC by 4.0 Deed (Attribution 4.0 International) license.

:::

\