AI: From Fierce Competition to Potential Monopolies

Introduction

As this weekend turned out to be quite grey and cold, I used the opportunity to read a very recent National Bureau of Economic Research paper from Anton Korinek and Jai Vipra titled "Concentrating Intelligence: Scaling and Market Structure in Artificial Intelligence" (NBER Working Paper 33139). It′s really worth also a second read... The paper highlights a fascinating paradox: While the AI industry currently experiences intense competition, powerful economic forces might soon steer it toward significant market concentration. This shift has profound implications for businesses, policymakers, and society at large.

The world of artificial intelligence (AI), particularly in foundation models like large language models (LLMs), is evolving at an astonishing pace. It's reminiscent of the early days of the internet, where countless startups emerged overnight, each hoping to become the next big thing. Today, a multitude of companies - amongst them of course my emplyoee 微软 -are racing to develop advanced AI models, pushing the boundaries of what's possible and transforming the technological landscape.

The Current Landscape: A Frenzy of Innovation

Right now, the AI market is buzzing with activity. New models and updates emerge almost daily, much like the surge of smartphone releases in the late 2000s. Companies like OpenAI, Google DeepMind, xAI, and Anthropic are neck and neck in developing LLMs.

This fierce competition benefits users through lower prices and rapid improvements, similar to how competition among airlines can lead to better services and cheaper flights for travelers. For now, the competitive dynamics keep innovation thriving and ensure a diversity of approaches, but storm clouds are beginning loom on the horizon.

The Gathering Clouds: Potential for Market Concentration

Despite the vibrant competition, several factors could lead to the AI industry being dominated by a few key players.

1. Economies of Scale and Scope

Developing and training foundation models require enormous upfront investments, akin to building a massive factory or launching a satellite into space. Once the initial model is developed, the cost of running it for additional users is relatively low. This setup favors larger companies (economies of scale) that can afford the initial expense and then spread the costs over a large user base, much like a big-box retailer offering lower prices by buying in bulk.

Moreover, these AI models are incredibly versatile. Just as a smartphone replaces multiple devices—a camera, GPS, and music player—a single AI model can perform a wide array of tasks, from drafting emails to providing health advice. This versatility, or economies of scope, gives an advantage to companies that can deploy their models across various industries.

2. The Intelligence Feedback Loop

An intriguing aspect of AI development is how advanced AI systems contribute to their own improvement. As AI becomes more sophisticated, it assists in developing even better versions of itself. Imagine a robot that can build increasingly advanced iterations of itself—this self-reinforcing cycle could accelerate AI advancement rapidly, widening the gap between leading companies and others.

3. Some graphical insights on the market taken from Korinek and Vipra (2024):

Valuation of leading LLM providers:

Principal players in the market:

Market share of leading vendors in the GPU market:

Key Resources: Compute, Data, and Talent

Three critical resources are central to AI development: computational power (compute), data, and talent.

Compute

The computational resources needed to train advanced AI models are immense and growing exponentially. It’s comparable to the energy requirements of running a large industrial plant. As models become more complex, they demand more computational "fuel," creating a significant barrier to entry. Established companies with access to large-scale computing infrastructure are at a considerable advantage, much like how access to vast tracts of land benefits large-scale agriculture.

Data

High-quality data is the lifeblood of AI models. Training these models requires vast amounts of information, similar to how a chef needs a wide variety of ingredients to create a gourmet meal. However, access to such data is becoming increasingly restricted and proprietary. Large tech companies, with their extensive user bases and data collection capabilities, hold a significant advantage, much like how a few conglomerates control the majority of media outlets.

Talent

Developing cutting-edge AI systems requires highly specialized expertise. The competition for skilled researchers and engineers is intense, leading to soaring salaries. Smaller companies and startups may struggle to attract and retain top talent, akin to how small towns find it challenging to keep their best and brightest from moving to big cities with more opportunities.

Vertical Integration: Pros and Cons

Many AI companies are moving toward vertical integration, controlling multiple stages of the AI value chain. For instance, some firms develop their own hardware (like specialized chips), create the AI models, and integrate these into end-user applications. This approach is similar to a car manufacturer that not only assembles vehicles but also produces its own parts and sells directly to consumers.

While vertical integration can lead to efficiencies and a seamless user experience, it also raises concerns about competition. If dominant companies control both the resources and the distribution channels, it could stifle innovation and make it difficult for new entrants to compete—much like how owning both the railways and the goods being transported can lead to monopolistic control in the logistics industry.

Safety Considerations: A Critical Balancing Act

As AI systems become more powerful, ensuring their safe and ethical use becomes increasingly important. However, the race to develop and deploy new AI technologies can create pressure to cut corners on safety, similar to how a company might overlook environmental regulations to expedite construction. Policymakers and industry leaders face the challenge of promoting innovation while also safeguarding against potential risks.

Looking Ahead: Navigating the Future of AI

The rapid evolution of AI presents both incredible opportunities and significant challenges. Key questions we need to address include:

• Balancing Competition and Safety: How can we encourage innovation without compromising safety and ethical standards? Is there a way to incentivize companies to prioritize both?
• The Role of Open-Source AI: Can open-source models democratize AI development, similar to how open-source software like Linux provided alternatives to proprietary systems?
• Equitable Distribution of AI Benefits: How do we ensure that the advantages of AI are shared broadly across society, rather than concentrated among a few powerful entities? Could policies similar to those promoting universal education or healthcare be applicable here?

Conlsuion: Embracing Collaboration for a Better Future

The rapid rise of foundation models marks a transformative era for artificial intelligence, carrying profound economic and societal implications. These models, capable of reshaping industries, exhibit a strong tendency toward market concentration due to economies of scale and scope, high pre-training costs, and bottlenecks in critical inputs like data and talent. This raises a pivotal question: will AI’s future be dominated by a single provider, or will it evolve into a diverse and competitive ecosystem?

Shaping the future of AI requires a collaborative approach involving businesses, policymakers, technologists, and the public. By fostering open dialogue, encouraging responsible innovation, and implementing thoughtful regulations, we can harness the transformative potential of AI while mitigating risks. It's a collective effort, much like addressing climate change or public health crises—we need everyone working together to navigate these challenges successfully.

In the short term, policymakers face the urgent task of promoting competition while mitigating risks. Strategies like mandating data sharing, enforcing interoperability standards, and fostering open research can help counteract concentration. Antitrust scrutiny of vertical integration within the AI value chain may also prevent undue market power. Additionally, as AI becomes central to the economy, treating foundation models as public utilities with fair access rules could ensure equitable usage.

Balancing these efforts with the need for AI safety is a complex medium-term challenge. As AI systems grow more powerful, their misuse or malfunction could have far-reaching consequences. Policymakers must strike a delicate balance between fostering innovation, ensuring safety, and preserving a competitive landscape.

Looking ahead, the governance of AI will require unprecedented collaboration among technologists, policymakers, and ethicists. The focus must extend beyond economic gains to ensure equitable access to AI-enabled tools and opportunities, reshaping participation in society.

The decisions made at this pivotal moment will define AI’s role in our world. With thoughtful regulation, proactive competition policies, and a commitment to broad societal benefit, we can harness AI’s transformative potential while steering it toward an equitable and prosperous future.