【国际商会ICC】2024生成式AI与虚拟世界的竞争报告_17页_305kb

报告摘要

Summary of ICC's Response to the European Commission's Call for Contributions on Generative AI and Virtual Worlds

Core Content

The International Chamber of Commerce (ICC), representing over 45 million companies in more than 170 countries, has responded to the European Commission's call for contributions on the competition and innovation aspects of generative AI and virtual worlds. The response emphasizes the importance of a balanced and evidence-based approach to regulation, ensuring that innovation and competition in these fields are not hindered by premature or disproportionate interventions.

Main Components of Generative AI Systems

ICC identifies the following as the main components necessary to build, train, deploy, and distribute generative AI systems, including Foundation Models (FMs):

Data: Essential for both pre-training and fine-tuning. Publicly available datasets are widely used, and proprietary data is also a key resource.
Compute Capacity: High-performance GPUs and other AI accelerators are crucial for training large models. NVIDIA is currently the leading supplier.
Capital: Necessary to scale FM development and leverage other resources.
Technical Expertise: Includes data scientists, machine learning engineers, and domain-specific knowledge.

Barriers to Entry and Expansion

Upstream (GPU and compute infrastructure): Financial resources and specific technical know-how can be barriers, though competition is growing.
Downstream (FM applications): While the market is expanding, access to essential components remains important. ICC highlights the need for non-discriminatory and affordable access to all components, especially for SMEs.

There are no material barriers to sourcing the main components, but access to data and computing power is more concentrated.

Drivers of Competition

Competition in the generative AI space is driven by multiple factors across the AI technology stack:

AI accelerator chips: Companies like NVIDIA, Google (TPUs), and others are developing or supporting custom chips for AI.
Compute infrastructure and energy: Availability of compute resources and affordable energy is a key enabler, with a growing number of specialized AI cloud providers.
Data: Scale and quality of data are critical, especially for training and fine-tuning models.
FM development and applications: Customers prioritize model performance over brand or parameters, leading to a diverse ecosystem of models and applications.

Potential Competition Issues

ICC believes that existing competition law tools are flexible enough to address potential issues in the generative AI space. Possible concerns include:

Access to key inputs: If dominant firms restrict access to data or compute resources, it could stifle competition.
Market power: There is a risk that companies with significant market power could abuse their position, for example, by degrading access conditions for competitors.

However, ICC supports the European Commission's consultation process as a means to gather market insights and respond quickly with existing tools.

Monetisation of Generative AI

Monetisation occurs across different layers of the AI stack:

AI accelerator chips: Sold to cloud providers and FM developers.
Compute resources: Provided as a service to FM developers.
Data: Licensed or freely available, depending on the dataset.
FM and applications: Accessible via APIs or open model registries, with monetisation through direct sales, service offerings, or integration into products.

Open-Source vs Proprietary AI

Open-source generative AI systems and components can effectively compete with proprietary systems when adapted and maintained. They enable faster development and innovation, especially when combined with access to data and computing power. Examples include models from Meta, OpenAI, and Hugging Face. While downstream services use similar inputs, competition occurs at the application level.

Role of Data and Interoperability

Data: Critical for training and fine-tuning models. Quality, scale, and availability determine the effectiveness of generative AI systems.
Interoperability: Facilitates scalability and switching between platforms, reducing barriers for new entrants. Open standards and APIs are key enablers.

Vertical Integration and Competition

Vertical integration can provide competitive advantages, especially when companies control multiple layers of the AI stack. However, ICC notes that no single company currently controls all components. Competition authorities should monitor such integration to ensure fair market practices and prevent abuse of dominance.

Investments and Acquisitions in Generative AI

Large companies often invest in or acquire smaller AI providers to extend their offerings or create synergies. These investments generally promote competition and innovation. However, existing merger control rules, like Article 22 of the EU Merger Regulation, may create legal uncertainty. Clearer rules and enforcement would be beneficial.

Adaptation of EU Antitrust Concepts

ICC believes that the current EU antitrust framework, including Articles 101 and 102 TFEU, the EU Merger Regulation, and Regulation 1/2003, is sufficient to address competition concerns in the generative AI space. New theories of harm can be developed as needed without requiring new legislation.

Adaptation of Antitrust Investigation Tools

Competition authorities should update their tools to better understand and investigate AI-related issues. The use of AI-powered tools can enhance their ability to detect harmful behavior. Technical knowledge is essential for effective enforcement in this evolving field.

Virtual Worlds Overview

Virtual worlds are immersive digital environments with applications in entertainment, education, business, and social networking. They can be categorized as B2C (consumer-focused) or B2B (industrial-focused), with distinct competitive dynamics.

Entry Barriers in Virtual Worlds

Entry cost: High infrastructure, development, and marketing expenses.
Expertise: Requires knowledge in 3D graphics, physics simulations, etc., which can be outsourced.
Intellectual Property (IP): Protection and monetization of IP are important but can create conflicts with third-party developers.
Connectivity: Reliable and low-latency connectivity is essential, especially in industrial applications.

These barriers vary based on the maturity of the market, but many are common across both emerging and established sectors.

Drivers of Competition in Virtual Worlds

Key drivers of competition in virtual worlds include:

Access to data: Crucial for enhancing user experiences and content creation.
Own hardware/infrastructure: Quality and scalability of infrastructure affect performance.
Open APIs: Enable integration of third-party assets and services.
Domain expertise: Helps in interpreting and utilizing data effectively.
Vertical integration: Offers competitive advantages but must be monitored to prevent abuse.

These drivers are likely to remain important, with a growing emphasis on data privacy and security in the B2C segment.

展开完整摘要