What are the real sovereignty and competition concerns in the generative AI value chain, and can law remedy it?

The Generative AI value chain

As per data, each day, about ten billion searches are conducted worldwide, of which about half, that is five billion searches do not return the best results. Furthermore 94.9% of these searches are conducted on the Google Search Engine, while Microsoft Bing receives only 6% of the remaining searches. Considering that over half the search results are unsatisfactory, Google first started to offer ‘AI Overviews’, which was soon followed by its ‘AI Mode Experiment in Search’, an ‘AI powered response’. Microsoft too integrated its Bing search and Edge browser with AI Copilot and Chat. These rapid developments in the search market, and changes even by an incumbent monopolist such as Google to its super dominant search functionality also caught the attention of the world’s leading competition authorities. To understand the concerns following the rise of generative AI, and how it functions, and what competition concerns may potentially emerge, the French Competition Authority FCA, identified three layers in its value chain.

1) The upstream AI infrastructure layer: Computing power, Data, Skilled Workforce

The AI infrastructure layer comprises of computing power, data, skilled workforce, and access to capital, such as venture capital (VC) or corporate venture capital (CVC). Computing power requires substantial resources such as chips, and cloud services. In the market for chips, though Nvidia is the undisputed market leader, particularly with its high-powered and disruptive graphic processing units (GPUs), it also faces healthy competition from Amazon, Microsoft and Google, together also known as the hyperscalers in cloud computing. Cloud services are not only highly concentrated markets, but they are also a key input to train the generative AI general purpose models. In fact, Open AI could successfully lead the AI market, as Microsoft dedicated ‘a specialised infrastructure’ for Open AI to train its GPT models.

In addition to the cloud infrastructure, one also needs access to high quality training data to train these models. Google, Amazon, Microsoft, Meta and Apple (GAMMA) again have a lead here, as they benefit from the data feedback loop, economies of scale and scope and learning effects from user-generated data.

Thus, one sees that winners in the digital platform economy have a competitive edge in the race for generative AI, and in case it is a new entrant, it is one (Open AI) that richly benefits from infrastructure and capital offered by the incumbent digital player (Microsoft). In addition to cloud infrastructure, data, availability of capital, industry, and the large digital players in particular, thanks to their deep pockets, can hire top technical and managerial talent, either through recruitment or through acqui-hire, issues that also raise novel concerns for merger control law.

2) The AI modelling: Training the model, developing the mode, and fine-tuning models for vertical specialization

Following the infrastructure layer, comes the training layer. AI models may be general purpose, also known as General Purpose AI Models (GPAI), or these may be forked to create vertically specialised models. GPAI are fine-tuned to perform specific tasks in a particular domain. For example, whereas LLama-2 is a GPAI offered by Meta, AstroLLama is vertically specialised model that has been trained for astronomy. These models may be open sourced or closed sourced. As open source is seen as key to innovation, and technological sovereignty, in June 2026, the European Commission announced targeted measures as part of its open source strategy, and includes measures to ensure open source alternatives for workplace tools, secure email, and an open internet architecture. This strategy is part of a larger technological sovereignty package.

3) The AI deployment: Model deployment for use by end users

The deployment layer is the user facing layer, and includes applications such as Apple’s Apple Intelligence, that is integrated into Apple devices and services, and Meta’s AI-based answer engines on Facebook, Instagram and WhatsApp. A look at these applications suggests that AI is being increasingly integrated into available software ecosystems. Functionalities such as AI-powered search, be seen as embedded features in the gatekeeper’s currently designated core platform service, search. Such a constructive view is also in alignment with the goals of the DMA.

Fairness and contestability (as explicit), and sovereignty (as implicit) goals of the DMA

It is vital to understand the policy rationale of a legal instrument to facilitate effective implementation. Fairness and contestability are widely understood as the explicit policy goals of the DMA. The DMA aspires to ensure that citizens and companies in the internal market can participate in the market economy, by creating a ‘society of equals by setting obligations on digital platforms’. 

To elaborate the connection between the DMA and its implicit policy goal of sovereignty, one must look at the infrastructure layer in the generative AI value chain, which is the  “infrastructural gateway”, which creates both infrastructural and innovation dependencies. As the DMA offers cloud computing as a core platform service (CPS), the Commission in November 2025 initiated market investigation on the sector. Even though Amazon Web Services (AWS), and Microsoft Azure do not quite meet the quantitative thresholds, the Commission took a preliminary view on designation, and pointed out that their ‘significant turnover, operational capacity and investments’, may after all qualify them as a CPS under the DMA.

Self-preferencing in generative AI

Self-preferencing is a situation of vertical foreclosure by an integrated form to discriminate in favour of its affiliates, and to disadvantage competitors. As generative AI firms are oftentimes vertically integrated (Google, Amazon, Meta, Apple), or have benefitted from a large digital vertically integrated big tech (OpenAI/Microsoft), competition authorities expressed concern whether firms with control over critical inputs in the infrastructure layer may foreclose downstream competition. Should this happen, and to the extent these concerns in generative AI follow from the gatekeepers’ CPS, in other words – in case the AI is integrated to enhance the functionality of the CPS (such as AI-powered search), these embedded services can and must be subject to the DMA. Additionally, the Commission, following a market investigation, must ascertain how a new list of AI-related CPS be added to the DMA.

Summary

While the AI market may look dynamic and fast evolving, two factors attract immediate attention to the market. First, competition concerns along the generative AI value chain are not mere contestability concerns, they are also sovereignty concerns, as dependence on the hyperscalers can create long-term innovation and infrastructural dependencies. Second, the large digital platforms also seem to have a competitive advantage in the AI race, and may be able to leverage these advantages to generative AI. 

For a timely and meaningful intervention, one must look at the available legal tools, and understand their underlying policy rationale for stronger and effective enforcement. The DMA may be peppered with fairness and contestability, but one must also take note of its subtle flavour of sovereignty. To savour the flavour, cloud computing, following a qualitative route be designated as a core platform service, as the hyperscalers are also the gatekeepers.