AI Supply Chains aren't AI Value Chains
The differences matter. Let's talk about why.
We’ve talked about AI supply chains in the past. Today, we’ll consider AI value chains, discussing what distinguishes supply chains from value chains to argue why the choice of terminology matters for both research and policy.
AI development has shifted fundamentally. Before 2020, most AI systems were produced in-house with very little outsourcing. That’s no longer the case. The rising influence of Base (or Foundation) Models and AI-as-a-service (AIaaS) has created an ecosystem where any organization in need of AI has easy access to the services and pre-trained models they desire. As a result, new AI experiences are often the product of a network of outsourced models, data, and tooling.
In late 2022, technologists and policy-makers alike began to pay close attention to these complex networks. The new phenomenon introduces pressing questions about monopolies, responsibility allocation, and system reliability. We discussed it at length in previous posts and called such networks “AI Supply Chains”.
Before this, the only meaningful reference to these networks came from the EU AI Act—the first comprehensive AI legislation proposed by a major regulator. The hundred page document, introduced before AI development networks exploded in complexity, only referenced them in passing and referred to them as “AI Value Chains”. Subsequent amendments have since more explicitly described “AI Value Chains” in an effort to shape regulation. But the use of AI Value Chains to describe the complex modes of deploying modern AI may not be ideal for the current climate (in regulation or in research). Today, we’ll discuss the pros and cons of using AI Supply Chain versus AI Value Chain, distinguishing the bodies of work from which they draw upon to articulate why the choice in language matters1.
What is an “AI Supply Chain”?
A supply chain refers to the full network of entities involved in producing and delivering a product or service to the consumer. Supply chains are primarily concerned with the operational aspects of getting a product from the initial supplier to the end customer efficiently and effectively.
In other words, AI Supply Chains model the life cycle of AI development functionally, focusing on the sequence of operations—the logistical network and processes involved in the sourcing, development, and delivery of AI technologies and services.
What is an “AI Value Chain”?
While supply chains are concerned with logistics and operational efficiency, value chains focus on maximizing value creation and competitive advantage.
Value chains represent the full range of activities that a business completes in order to bring a product or service to consumers. The concept, introduced by Michael Porter in the 1980s, focuses on adding value with each step of the product life cycle within a single firm, from design, production, and marketing all the way through after-sales service. With time, the term was expanded to consider multiple firms contributing to the same product, but the emphasis on a local (not global) view of production remains.
AI Value Chains offer insight into how various processes add monetary value to an AI product, expanding supply chain’s tighter scope to include intangible processes (e.g. design and innovation). The supporting value chain literature views operations through this lens and is not specifically tailored to studying complex logistical processes.
Consequentially, AI Value Chains entered business vernacular long before policy was needed, or before the complex AI development networks of the last year became prominent.
Why the AI Act used value chains
For many years, AI was not developed through complex networks, and language to describe the “chain” of development was not needed on a large scale. In the late 2010’s, major consulting firms and think tanks began to talk about the AI Value Chain. Such articles explored how companies developing AI gained competitive advantage and where the market offered opportunities for growth. From a financial perspective, AI Value Chain was a natural term to use when studying the market.
When the EU AI Act was first proposed in 2021, AI development in industry was just becoming a collaborative, less siloed process. The proposed law refers to the AI Value Chain when describing the obligations and rights of participants in that collaborative process. Precedent and knowledge at the time suited this choice—it was not yet clear that AI development would expand across firms in such a technologically complex way as to warrant a different scope than value-add.
Later work elaborated on the EU AI Act’s AI Value Chain, and several parties began examining and mapping the growing networks of AI development. Because AI has grown so complex, much of this recent work, still referencing AI Value Chains2, was devoted to simply understanding how AI systems are developed—regardless of “value-add”.
Here, the ideas behind the AI Value Chain became disadvantageous, and the term grew overloaded. As researchers work to understand the logistics of AI development networks, the monetary value of individual processes in the market becomes secondary to a technical or functional understanding of how these processes work together. Relative to supply chains, value chain literature is not well-suited to developing such a logistics-forward understanding. As a result, increasing numbers of publications are conflating value chain with supply chain—and reintroducing known concepts from supply chain research.
While the AI Value Chain’s initial adoption within the EU AI Act may have been a natural choice, successive usage has since muddied its definition, making it more difficult to effectively draw from their respective bodies of supporting literature.
Where AI Supply Chains come in
Value chains are most helpful when their user has reasonable insight into the details of each process’s value-add. Value-add is critically affected by how it interacts with related processes. Individual organizations can understand this for their own value chains, and it may have been practical to map it at a large scale a few years ago.
But work mapping how complex AI systems flow through these development networks, how AI products work, and how they are repurposed has become the focus of researchers and policy makers alike. And, fundamentally, this interest in logistics is the bread and butter of supply chains, in contrast to the (narrow) lens of value-add. Given how early we are in the AI boom, this framing provides richer tooling for understanding complex AI systems.
To illustrate this, we’ve previously mapped supply chain literature to best practices and priorities for AI development. For example, modularity in supply chains and redundancy in suppliers or manufacturers are low hanging goals that we believe future regulations should adopt from supply chain literature. And this is only the tip of the iceberg. The existing body of supply chain literature—and its close ties to current regulation—is a well-established interdisciplinary field that extends from business sciences into ethics, operations research, and beyond.34
Looking forward
As we begin to understand, criticize, and eventually tweak the processes of AI development and deployment, the scope of AI Supply Chains and the large body of related supply chain literature should serve technologists and policy makers alike.
AI systems are becoming more sophisticated and integrated across sectors. As a result, researchers are studying unique challenges in accountability attribution, security maintenance, and much more in AI Supply Chains. This work benefits from borrowing abundantly from language and ideas in supply chain literature.
And as the field evolves, value chain literature will inevitably continue to grow more useful in novel ways. It’s critical to take advantage of both terms (without overloading either). We think standardized terminology and modes to translate between terms must be established—both for the sake of collaboration, and the future of AI policy.
Another emerging term is "AI Stack". Unlike the AI Value and Supply Chains, which evoke economic and operational concepts like value-add and logistical efficiency, "AI Stack" is analogous to the software stack and simply refers to the layers of technology and processes contributing to AI systems. By referencing more neutral technical ideas, the term AI Stack avoids economic implications and fails to benefit from the wealth of literature supporting supply chain and value chain usage, which can help capture and shape the complexities of AI development networks. Additionally, alluding to the highly modular software stack may inadvertently obscure unique challenges stemming from AI’s inherent non-modularity.
For work directly related to the EU AI Act, it might have seemed unnatural to introduce different terminology. For other standalone literature, it might have been instinctive to follow an existing precedent.
Ellram, L.M. (1991), "Supply‐Chain Management: The Industrial Organisation Perspective", International Journal of Physical Distribution & Logistics Management, Vol. 21 No. 1, pp. 13-22. https://doi.org/10.1108/09600039110137082
This isn’t to say AI Value Chains aren’t useful—studying value-add in AI deployment networks is appropriate for many contexts. Venture capitalists can use it to understand investment value, while individual organizations might gain competitive advantage by optimizing their individual value chain [7].
Thank you for this work you are doing. It's very important.