TLDR: This paper analyzes the growing trend of ‘openness’ in AI, where leading firms release models and tools under various conditions. Using a global value chain approach, it explores the motivations behind this ‘strategic market openness’—ranging from market development and standardization to cross-subsidies. The research identifies five distinct governance models (hierarchical, captive externalized, open platform, controlled openness, and full openness) that illustrate how foundational AI firms maintain control and power even while offering open resources, ultimately shaping downstream capabilities and opportunities in the AI ecosystem.
The rapid ascent of Artificial Intelligence (AI) has not only reshaped investment landscapes and promised widespread economic disruption but has also sparked a crucial debate about “openness” in the AI sector. While AI has often been associated with the concentration of power and value among a few “big tech” giants, a parallel trend of AI firms claiming openness, offering “open source” models, and providing free access is emerging. This development raises questions about whether these open resources can truly foster technological transfer and enable smaller players to catch up against the backdrop of dominant AI industry power.
A recent working paper, “Openness in AI and downstream governance: A global value chain approach,” authored by Christopher Foster from the University of Manchester, UK, delves into these complex dynamics. The paper aims to bring conceptual clarity to these discussions by viewing openness in AI as a unique type of relationship between firms, making it suitable for analysis through a global value chain (GVC) lens. This approach helps understand the capitalist forces driving foundational AI firms to “outsource” key components and the resulting governance and control mechanisms that emerge as AI is adopted across various industries.
Understanding Openness in AI
The paper defines “openness” broadly, encompassing claims made by AI firms about open source and free provision. Historically, open resources like PyTorch and TensorFlow libraries and datasets like Common Crawl have existed. However, the debate intensified with the opening of general-purpose AI models. Notable examples include Meta’s LLaMA model, released under a permissive open license allowing commercial reuse and fine-tuning, and Chinese models like Deepseek-R1 and Qwen (Alibaba), which offer near open-source conditions. Advocates suggest this fosters technological transfer and learning, potentially enabling firms to “catch up.” Critics, however, view it as “open washing,” a tactic that reinforces the power of leading AI firms by appearing ethical without true transparency or equitable access.
AI Through a Value Chain Lens
To understand the broader implications of AI, the paper adopts a relational approach, conceptualizing AI production and use within “AI value chains” or “AI supply chains.” This framework helps map the fragmented processes, from data and infrastructure harnessing to model production and application across industries. It highlights the distributed nature of AI development and the need to consider the division of labor, value distribution, and ethical responsibilities across the chain. The paper extends existing AI value chain mappings to critically examine the trade-offs between the power of leading firms and the opportunities for AI users in this “open” era.
Adapting Global Value Chain Frameworks for AI
The research adapts the Global Value Chain (GVC) framework to analyze AI and openness, focusing on three key areas:
- AI Assets and Mediating Relationships: Unlike traditional GVCs focused on goods, AI value chains are mediated predominantly by software, services, and integrative logic. Relationships can be transactional (contracts, payments) or involve free software/services, licensing terms, indirect payments through data, or even “institutional” power embedded in algorithms. This “generative” structure means AI solutions can be bundled and repurposed in unexpected ways.
- Capitalist Dynamics of Foundational Leaders: The paper introduces “strategic market openness” to explain why foundational AI firms open up high-value components. This goes beyond simply outsourcing “low-value” activities. Motivations include market development, fostering competition to reduce costs, setting industry standards (e.g., Google’s TensorFlow driving use of its TPU hardware), and even geopolitical competitiveness. For some, it’s a “freemium” strategy, offering lightweight models to attract users to proprietary platforms.
- Openness and Governance: The traditional GVC governance taxonomies, based on transaction cost economics, are less suitable for AI’s open nature. Instead, the paper draws on concepts from modular production networks, platform governance, and open innovation to understand how foundational firms govern downstream networks.
LLMs as a Case Study for Openness
The paper uses Large Language Models (LLMs) as a prime example to illustrate the dynamics of openness. While many foundational LLMs are now released with open weights (e.g., Meta’s LLaMA 3, DeepSeek), allowing users to download and adapt them, significant constraints remain. These include varying licensing conditions (e.g., user limits, restrictions on training other LLMs), delayed or lightweight releases from major players like OpenAI and Google, and, crucially, the lack of full training data and detailed model creation information. This opacity can lead to implementers becoming dependent on foundational firms, giving rise to “open washing” critiques.
The paper highlights a “gradient framework for model release,” ranging from fully proprietary to fully open models, demonstrating the diverse strategies employed by firms regarding AI openness.
Diverse Governance Models in AI Value Chains
Based on its analysis, the paper proposes five types of governance models in AI value chains, reflecting different balances of control and capability building:
- Hierarchical: Foundational firms internalize AI competencies, offering it as a service or software. Users are often pulled into captive relationships, with most innovation remaining in-house (e.g., Siemens, IBM).
- Captive Externalized Governance: Firms provide AI access via clouds, APIs, or platforms. This offers flexibility but comes with constraints on intellectual property, usage, and costs, leading to potential “platform lock-in” (e.g., OpenAI GPT, Google Gemini).
- Open Platform: Infrastructure providers (like Amazon Bedrock or Hugging Face) offer a range of open models (even from competitors) within their optimized infrastructure. While offering customizability, these providers control key “choke points” like compute power, maintaining a degree of captive relationship.
- Controlled Openness: Leading foundational AI firms release near-cutting-edge models under relatively permissive licenses (e.g., Meta LLaMA, DeepSeek). These offer flexibility but still have policy limits and inherent constraints (like lack of full data/compute) that prevent true independence for implementers. The goal is often to promote adoption and shape standardization.
- Openness: This aligns most closely with formal open-source AI, where full models, toolchains, and infrastructure are in the public domain under highly flexible licenses (e.g., BLOOM, some university-driven initiatives). Here, motivations extend beyond purely economic goals to include academic, geopolitical, or sovereignty objectives.
Also Read:
- Designing the Future: Navigating the Emergence of AI Agent Economies
- Charting the Path to Self-Driving Science with AI Agents
The Future of Open AI
The paper concludes that while openness in AI is a significant and potentially enduring trend, foundational AI firms are unlikely to relinquish their lead positions. Even in more open scenarios, mechanisms like platformization, cross-subsidies, control of “choke points” (like infrastructure), and standardization efforts allow leading firms to maintain considerable power and control. Therefore, claims of openness are not contradictory to the powerful positions of leading AI firms.
Despite these control dynamics, the paper suggests that openness can still be a crucial step in expanding AI use and capabilities. As computational costs decrease, fine-tuning becomes more efficient, and new applications emerge, open models and tools are likely to become even more viable. This dynamic holds important implications for capability building, specialization, and upgrading opportunities within global value chains, offering new avenues for local development and innovation.
