Summary Under the proposed Cloud and AI Development Act (CADA), the distinction between a "frontier AI model" and a "frontier AI system" is structural, not functional. Article 2(4) defines "frontier AI" broadly as "AI models or AI systems built upon such models that can perform a wide variety of tasks and that approach, reach or exceed the current state of the art." This definition explicitly captures two distinct entities: the foundational model itself, and the operational system constructed upon it. Both are eligible subjects for recognition as "frontier AI priority projects" under Grand Challenge 3, ensuring that EU support covers the entire innovation value chain from foundational research to advanced application deployment.

Detail

To understand the strategic scope of the proposed Cloud and AI Development Act (CADA), one must examine how the proposal defines its core technological assets. The proposal aims to strengthen Europe's cloud and AI ecosystem by reducing dependencies on third-country providers and boosting domestic capabilities. A central pillar of this effort is the support for cutting-edge technologies, specifically "frontier AI." The legal framework distinguishes between the underlying algorithmic engine and the deployable application, yet treats both as critical strategic assets.

The Legal Definition: Article 2(4)

The foundation for this distinction lies in Article 2(4) of the CADA proposal. It defines "frontier AI" as:

"AI models or AI systems built upon such models that can perform a wide variety of tasks and that approach, reach or exceed the current state of the art."

This definition is intentionally inclusive. It does not restrict the term "frontier AI" solely to the foundational models trained on massive datasets. Instead, it explicitly captures two distinct categories of innovation:

  1. Frontier AI Models: These are the foundational, large-scale models (often general-purpose) that possess high-impact capabilities. They represent the "engine" of the AI stack.
  2. Frontier AI Systems: These are the specific applications or systems constructed using these frontier models as their core component. They represent the "vehicle" that delivers value to end-users.

The critical connector in the definition is the phrase "built upon such models." This indicates that a system is considered "frontier" in the context of CADA if its capability to perform a wide variety of tasks at a state-of-the-art level is derived from the underlying frontier model. The definition acknowledges that the state-of-the-art capability can reside in the model itself or be realized through the sophisticated integration of that model into a system.

Models vs. Systems: The Structural Distinction

For technical leaders, architects, and project proposers, the difference is architectural and operational:

  • The Model (The Engine): A frontier AI model is the trained mathematical structure, including its weights, parameters, and architecture. It is the asset that is trained on vast datasets to achieve general reasoning or generation capabilities. In the CADA context, the model is the source of the "state of the art" capability. It is often agnostic to the final use case until it is integrated into a specific workflow.
  • The System (The Vehicle): A frontier AI system is the deployable product. It integrates the model with software interfaces, data pipelines, safety guardrails, specific application logic, and user interaction layers. It is what the end-user interacts with.

While the EU AI Act (Regulation (EU) 2024/1689) also distinguishes between models and systems for regulatory purposes (e.g., obligations for general-purpose AI models versus high-risk AI systems), CADA focuses on this distinction to determine eligibility for strategic support and capacity allocation. The CADA recognizes that innovation happens at both levels: developing the foundational model and building the sophisticated systems that unlock its value for specific industrial or public sector needs.

Grand Challenge 3: Support for Both

The practical implication of this dual definition is found in Article 4 and Annex I of the proposal, which establish the "Cloud and AI Leadership Initiatives." These initiatives are operationalized through "Grand Challenges" designed to address major technological and industrial challenges.

Grand Challenge 3 is explicitly titled "Frontier AI." Its objective, as stated in Annex I, is:

"Developing the next generation of multimodal frontier AI models and systems and pioneering novel capabilities."

The text of Annex I further elaborates that Grand Challenge 3 focuses on:

  • The architectural design and development of next-generation multimodal models.
  • Pushing boundaries in advanced reasoning, cross-modal understanding, and agentic capabilities.
  • Investigating novel approaches to model efficiency, cognitive modelling, and alternative computational structures.

Crucially, Article 8 of the CADA allows the Commission to recognize projects as "frontier AI priority projects" if they support Grand Challenge 3. This means that funding, strategic recognition, and access to computing resources are not limited to companies training base models. Projects that build complex, state-of-the-art systems using these modelsβ€”provided they meet the criteria of being pioneering, involving broad participation, and contributing to the strategic objectivesβ€”are also eligible subjects for support.

Why the Distinction Matters for Support

The CADA is designed to foster an entire ecosystem, not just a few model providers. By defining frontier AI to include both models and systems, the proposal ensures a balanced approach to ecosystem development:

  1. Model Developers receive support for the massive computational and research resources needed to train state-of-the-art models, addressing the "supply-side" gap.
  2. System Integrators and SMEs can access support to build the applications that utilize these models, ensuring that European industry can deploy these technologies effectively and avoid a "demand-side" bottleneck.

This aligns with the broader CADA goal of increasing the adoption of AI across the public and private sectors, as outlined in Article 3(1)(c). It prevents a scenario where Europe might develop the models but fail to build the competitive systems required to leverage them in strategic sectors like healthcare, transport, and manufacturing.

What this means for you

For CTOs, architects, SMEs, and research consortia evaluating the practical impact of CADA, this distinction clarifies your eligibility for support and your strategic positioning.

1. Eligibility for Priority Projects If your organization is building a sophisticated application that relies on a frontier model, you are not excluded from the "frontier" label simply because you did not train the base model. If your system performs a wide variety of tasks at a state-of-the-art level by virtue of being built on a frontier model, it falls under the definition in Article 2(4). You should monitor open calls for expressions of interest related to Grand Challenge 3, as your project may qualify as a frontier AI priority project under Article 8.

2. Strategic Partnerships The definition encourages collaboration across the stack. Model providers and system builders are part of the same defined ecosystem. If you are a system builder, partnering with model providers who are engaged in CADA-supported initiatives could strengthen your application for support. Conversely, if you are a model provider, demonstrating that your models are being used to build recognized frontier AI systems can validate their strategic importance and help meet the "broad participation" criteria required for priority project status.

3. Computing Resource Allocation Article 9 of the CADA mandates that the Union and Member States allocate sufficient AI computing resources to support frontier AI priority projects. This means that if your project (whether model-focused or system-focused) is recognized as a priority, you may gain access to European high-performance computing (EuroHPC) capacity. This is a critical resource for both training large models and running complex system-level evaluations and validations.

4. Compliance and Sovereignty While the CADA focuses on development and deployment support, remember that the systems you build must still comply with other EU laws, such as the AI Act and the GDPR. The CADA's sovereignty framework (Article 16) will also apply if you are providing cloud services to public sector bodies. Understanding whether you are acting as a model provider or a system deployer will help you navigate these overlapping obligations, as the CADA support mechanisms are distinct from the regulatory compliance requirements of the AI Act.

Common misconceptions

Misconception 1: Only base model trainers are "frontier AI" under CADA. Correction: Article 2(4) explicitly includes "AI systems built upon such models." If your system leverages a frontier model to achieve state-of-the-art performance across a wide variety of tasks, it is covered by the definition. The proposal recognizes that the value of frontier AI is realized in the system as much as in the model.

Misconception 2: The CADA replaces the AI Act's definitions. Correction: The CADA and the AI Act are complementary instruments. The AI Act regulates the safety, fundamental rights, and risk classification of AI systems. The CADA focuses on strengthening the ecosystem, capacity, and sovereignty. You must comply with both. The CADA's definition of frontier AI is tailored specifically for the purpose of identifying projects for support under the Leadership Initiatives, not for regulatory risk classification under the AI Act.

Misconception 3: Any system using a large model is a "frontier AI system." Correction: The definition in Article 2(4) sets a high bar. The system must "approach[es], reach[es] or exceed[s] the current state of the art" and must "perform a wide variety of tasks." A simple wrapper around a model that does not add significant architectural innovation, novel capabilities, or state-of-the-art performance may not qualify as a frontier AI system for the purposes of Grand Challenge 3 support.

Misconception 4: CADA support is automatic for all frontier AI. Correction: Support is channeled through specific mechanisms. Projects must be selected through open calls for expression of interest and recognized as "frontier AI priority projects" by the Commission under Article 8. They must also meet specific criteria, such as being pioneering projects and involving the participation of at least three Member States.

Official sources

Related

This is general information about a draft EU regulation, not legal advice.