Shifting Grounds? On the Need to Rethink Europe’s AI Ambitions and Competitiveness

CSDS POLICY BRIEF • 26/2025

By Antonio Calcara and Riccardo Bosticco

25.9.2025

Key issues

• The Artificial Intelligence (AI) market is undergoing rapid change. Lower costs, wider usage uptake and increased investments mean the AI market is shifting from infrastructure development to end-user applications;
• It is increasingly difficult for Europeans to compete in creating the full AI stack, so European efforts should increasingly be directed to where the most value can be made;
• While creating a “European AI stack” is desirable in principle, this risks underestimating the resource constraints hanging over the EU and the challenge of prioritising parts of the AI stack.

Introduction

‘Too often, I hear that Europe is late to the [AI] race’, thundered President Ursula von der Leyen at the AI Action Summit in February 2025 – ‘I disagree. Because the AI race is far from over. Truth is, we are only at the beginning’. The assessment of the European Commission President is correct. While AI is increasingly embedded in everyday life, AI performance on benchmarks keeps improving, investments are booming and models are becoming increasingly affordable. However, this does not mean that the European recipe to keep up in the AI race is consistent with the technology and its market. In the European approach to AI, there is a risk of a fundamental misunderstanding of what is required to achieve and maintain competitiveness, given the current conditions of the European Union’s (EU) economy and resource constraints.

Influential policy and intellectual discussions in Brussels have primarily focused on making the EU “sovereign” in the digital sphere. The idea is for Europe to gain control over the digital infrastructure and technologies it uses – 80% of which are currently imported – while fostering intra-European interconnectedness and improving economic performance. According to this perspective, the EU should replicate an indigenous AI stack – to integrate AI infrastructure, from hardware to software, as cohesively as possible – that can compete with, and not rely on, those of the United States (US) and China. Estimates underpinning this vision indicate that it would take nearly €300 billion and 10 years to achieve it.

This CSDS Policy Brief contributes to this discussion by highlighting the overlooked importance of two interrelated trends that are shifting the ground of the AI market. First, while different types of companies are investing in AI, cloud providers are already successfully monetising it by incorporating AI products in their cloud offerings. This makes it increasingly difficult for Europeans to compete in creating the full stack, as they are already heavily dependent on American cloud providers. Amazon, Microsoft and Google alone occupy 70% of Europe’s cloud computing market. Breaking this dominance would require enormous resources despite the overlapping challenges the EU faces, including rearmament and climate resilience. Furthermore, catching up in the cloud computing sector would have distributional implications within Europe that would make the process politically challenging. Second, as the technology matures, it may become less necessary to replicate the entire AI stack. In fact, ever-increasing global investment in AI infrastructure will lead to lower computing costs and the emergence of more AI models. This will allow access to more efficient, less expensive AI computing and models, contributing to a shift in focus in the AI race from infrastructure creation to end-user applications.

Taking stock of these two trends, we argue that Europe should focus its efforts precisely on end-user applications. We therefore propose that it adopts a pragmatic approach to AI infrastructure, encouraging, first and foremost, European public and private actors to test and scale AI applications across the wider economy. Only by promoting the use of AI within the European economy will it be possible to take advantage of the AI market and technology trends and build competitiveness. To develop this claim, this CSDS Policy Brief reviews recent European initiatives and debates on this subject, describes the two trends identified in the AI market, and, finally, proposes an alternative trajectory for Europe’s AI policy.

The EU’s approach to AI: overview and context

There has certainly been no shortage of appreciation in Europe regarding the importance of AI for productivity gains, economic growth and European military power. At the same time, however, the “Draghi Report” confirmed what many had long suspected: ‘Europe largely missed out on the digital revolution […] and the productivity gains it brought’, with direct consequences on the competition for AI today. Previous studies have situated the root causes of Europe’s shortcomings in market fragmentation, an inability to scale digital companies, a lack of a single capital market, relatively high energy costs and a certain naivety in protecting and stimulating investment in critical technologies.

Recognising the challenge, European policymakers have come up with strategies, initiatives and action plans. The goal is to build the computing infrastructure that is needed to ‘support [Europe’s] innovative start-ups to develop, train, and deploy their next-generation AI models’. The latest of such initiatives is the AI Continent Action Plan, whose objective is to strengthen the European AI ecosystem and boost the competitiveness of the domestic industry. Building on previous initiatives, the plan doubles down on AI factories – sites equipped with supercomputers and data centres where developers and start-ups can grow and test AI models. Given that Europe currently accounts for only 4% of the world’s AI computing power, the AI factories are intended to replicate the platforms provided by major non-European tech companies, providing access to computing power, data and talent, particularly for AI start-ups and SMEs. Moreover, the plan foresees a €20 billion investment in the purchase and deployment of 100,000 AI processors. In fact, the EU AI Continent Action Plan places particular importance on developing European advanced semiconductors (in the context of the EU Chips Act) and cloud computing infrastructure. This aligns with ongoing work on the Cloud and AI Development Act – a commitment to ‘at least tripling the EU’s data centre capacity’ within the next five to seven years – and on a Data Union Strategy to enhance and streamline access to high-quality data across industries and research domains.

These initiatives indicate that the EU is set to aim for a “European AI”. While the idea of making Europe’s digital sovereignty traces back to at least President von der Leyen’s first mandate, echoing calls for greater strategic autonomy, today’s forces in this direction are accelerated by the increasingly confrontational geopolitical context within and around the EU. With Russia’s persistent threat to the European security architecture, China’s support for it and the growing volatility of transatlantic relations, it feels rational for the EU to cut weaponisable ties with each great power and reduce its vulnerability to economic and digital shocks. A group of leading experts has recently contributed significant work to grounding Europe’s digital avenues in more sovereign terrain. The outcome of this work – the Eurostack report – defines the way forward for Europe to establish a stronger presence throughout the AI stack, from critical raw materials to specialised, high-precision chip manufacturing, data centres and advanced model training.

Other experts have called for Europe to champion a “third AI stack”. Amid Sino-American competition, Europe should engage with a constellation of other countries, such as Japan and South Korea, to diversify and improve market competition beyond the US and China and boost the practice of value-based innovation – providing in this way a global alternative to current AI developments. Yet, while all these policy proposals are desirable in principle, they risk underestimating the resource constraints hanging over the EU and the challenges that choosing which parts of the AI stack to prioritise would entail. Instead, a more realistic way forward can be derived from current and emerging trends in the AI market.

Getting it right: two market trends across the AI stack

To help tailor the EU’s approach to AI to current technological and market realities, we have identified two foreseeable trends in the global AI market and technology.

The first trend is that cloud providers – primarily Amazon Web Services (AWS), Microsoft Azure and Google Cloud – are already monetising AI by incorporating large language models (LLMs) in their cloud service suites and making them available to the millions of enterprise customers who already rely on their services. This is because, as Microsoft CEO Satya Nadella stated, AI’s ‘real demand’ is in ‘the enterprise space’. 71% of organisations now use generative AI services, and by 2030, data centres are projected to require $6.7 trillion in spending worldwide. Sensitive to different types of corporate demands, their services allow customers to fine-tune and manage the lifecycle of models on top of the existing cloud infrastructure. By doing so, Microsoft generates about $13 billion in AI revenue largely by giving customers access to foundational models such as OpenAI’s ChatGPT. This enables them to fine-tune the models and just supply the infrastructure to run inference workloads at scale. The ability to monetise AI enables cloud providers to invest more heavily in AI infrastructure and strengthen their competitive advantage. In addition to offering AI models directly, cloud providers are also indispensable for all other AI companies. Market leaders, such as OpenAI, also rely on the cloud services of major digital platforms to keep their products affordable. In fact, when smaller companies and start-ups transition from the pilot to the production stage, they need to find partnerships with cloud providers to scale up their operations.

The second key trend in AI is derived from the first, but it could lead in a different direction. Increasing investment in AI infrastructure is leading to lower computing costs and the emergence of more AI models and more accessible ones. Economic theory and economic history converge in suggesting that greater investment in a technology field leads to lower prices and increased competition in the long term. Competition is intensifying, and investments are surging across the entire AI stack.  For example, NVIDIA’s partnerships with CoreWeave and Lambda Labs support smaller cloud players in competing against AWS, Azure and Google Cloud. Moreover, major cloud firms are developing proprietary chips, and model developers such as xAI are building data centres to reduce their dependence on large cloud providers. Additionally, cross-layer moves usually spur competitive responses. OpenAI’s release of ChatGPT, for example, forced Google to accelerate the commercialisation of its generative AI capabilities, leading to the launch of products such as Gemini, Duet AI and AI-enhanced search.

Meanwhile, the cost of AI is dropping. Mary Meeker and her team have observed that the cost of AI inference is falling rapidly. Driven by increasingly capable models, the inference cost for a system performing at the level of GPT-3.5 has fallen by over 280-fold between November 2022 and October 2024. Meanwhile, hardware is improving too: NVIDIA’s 2024 Blackwell GPU uses 105,000 times less energy per token> than its Kepler GPU from 2014. Hardware costs have declined by 30% annually, while energy efficiency has improved by 40% each year. This is good news for users and developers, as they face increasingly lower unit costs to access AI infrastructure. However, as AI infrastructure becomes more affordable, added value will shift from infrastructure building to end-user applications. This also reflects Clay Christensen’s law of conservation of attractive profits: when one layer of a system becomes more widely available, profits shift to adjacent layers. In the case of AI, these are end-users’ applications.

The implications for the EU’s AI policy

The market concentration of cloud-based AI offerings, coupled with increased competition and potential cost reductions for accessing AI infrastructure, presents both constraints and opportunities for Europe’s approach to AI.

The fact that cloud providers are leading the way in AI offerings is not good news for Europe. Europeans are struggling to catch up in the cloud market because cloud first movers and incumbents enjoy economies of scale and substantial cost advantages, which benefit their customers. In fact, despite the rhetoric on digital sovereignty in Europe, all member states and major European companies rely on US cloud providers to some extent. Moreover, switching to any alternatives would be particularly expensive not only in economic but also in political terms. In Europe, the vast resources that cloud computing requires – computing power, land, water and energy – are not equally distributed across all 27 member states. Coordinating these resources would require significant political effort and redistribution of profits, which is challenging to achieve in a fragmented market. In fact, this is a recurring problem when it comes to European integration in strategic sectors. Economic logic would favour greater intra-European consolidation, but this concerns smaller countries, who fear that such consolidation could benefit larger countries (e.g. France and Germany) disproportionately, thereby undermining the level playing field within the EU.

The distributional implications within Europe are often overcome by framing European initiatives as a collection of national projects to create win-win situations. For example, the newly formed AI factories mainly cluster around national supercomputers. European policymakers are proud of the fact that Europe has some of the best supercomputers, and the EuroHPC infrastructure that supports AI factories is indeed a positive achievement. However, its computing power is not comparable to that of leading American providers. The EU’s total public AI computing capacity of around 57,000 accelerators is an order of magnitude smaller than a single US tech giant’s infrastructure. By way of comparison, Meta intends to deploy infrastructure equivalent to almost 600,000 NVIDIA H100 GPUs – more than ten times the size of the entire EuroHPC fleet. Moreover, combining all the different AI factories does not follow normal arithmetic, because they include combinations of proprietary hardware produced by companies such as NVIDIA and AMD. This variety creates interoperability problems and makes models created in one factory not transferable to another.

While the first trend entails mainly negative consequences, the second trend could have positive implications for the EU. Growing investment in AI infrastructure will lead to lower computing costs and an increasing number of affordable AI models. This indicates a shift in the competition from infrastructure creation to end-user applications. The European priority, therefore, should shift from creating EU-owned AI infrastructure to nurturing the demand for AI applications – an issue that is too often underestimated. While everyone is keen to emphasise the need for European AI, in fact, few consider who would purchase or utilise European AI models. According to a McKinsey survey, Europe lags behind North America in the adoption of generative AI (“Gen AI”), with 40% of North American companies surveyed saying they have adopted Gen AI in at least one business function, compared to around 30% of European companies. Only around 20% of companies successfully use AI in daily operations, while nearly 80% remain stuck in the pilot stage or have yet to begin implementation. Certainly, this is a structural issue, as large companies are more likely to adopt AI than small companies, and yet small companies dominate the European landscape.

Still, if the focus of AI competition shifts from infrastructure to end-users, and if Europe wants to remain competitive, it makes sense to prioritise end-user applications. But how could this be achieved?

Looking ahead: what Europe can and should do

The AI race is not over, but the locus of competition is shifting from infrastructure to end-user applications. Moving ahead, Europe should take stock of global trends in the AI market and adapt accordingly. We set out three policy recommendations to help Europe exploit these trends to its advantage.

Firstly, we propose that the European institutions and member states adopt a pragmatic approach to AI infrastructure. Replicating AI infrastructure within Europe poses economic, technological and political challenges. European efforts must focus on the application of AI. This does not mean abandoning Europe’s AI infrastructure plans but rather developing complementary infrastructure that runs in parallel with existing, more competitive foreign ones. The faster, more efficient and cheaper the computing power Europe can access, the better.

That this approach will lead to a deepening of dependencies on foreign suppliers is probably unavoidable in the short term, but it can be mitigated through portfolio diversification in the medium-to-longer term. Europe should implement open standards to ensure that European companies’ AI solutions are portable among suppliers. As suggested by Luis Garicano, this could be achieved by encouraging the use of standardised application programming interfaces (APIs) and data portability rules, enforcing open standards that allow European companies to easily switch AI providers, retain ownership of their innovations and keep the profits they generate. The establishment of data commons tailored to specific industries is certainly a good idea. In manufacturing, this could mean expanding projects like Germany’s Manufacturing-X across the continent, so that anonymised production data is aggregated into large-scale datasets for training industrial AI. In healthcare, the EU could leverage the strength of its universal healthcare systems by pooling anonymised patient records and genomic data. Doing so would provide pharmaceutical companies with unparalleled datasets for advancing drug discovery. Furthermore, our suggestion to increase investment in AI applications would indirectly improve Europe’s position vis-à-vis foreign AI infrastructure providers, as they would be more incentivised to ensure interoperability to avoid losing access to a large, profitable and innovative market.

Secondly, the private sector must embrace the innovative potential unlocked by AI advancements and focus on how these can improve their day-to-day operations and final products. In the past, European companies such as Spotify and SAP have developed rewarding applications relying on operating systems owned by foreign platforms. They avoided commoditisation by differentiating their products. Thus, Europe’s goal must be to use AI in ways that are hard to replicate and that make it lucrative even for the platforms they rely on.

European companies at the technological frontier have already realised this. ASML’s €1.3 billion investment in Mistral is an example, as it aims to improve internal production processes through AI to strengthen its competitive advantage in the global semiconductor industry. Given announced and expected major investments, the defence industry holds opportunities. AI could be used for experimentation during the production phase (e.g. digital twins in the production of the European sixth-generation fighters) or for data analysis (e.g., regarding air defence systems), as it is already used to optimise rockets and launch vehicles in the space sector. Similarly, AI is essential if our car manufacturers are to remain competitive in the era of software-defined vehicles. Volkswagen’s recent announcement to invest around €1 billion in AI by 2030 to support AI-driven advancements in vehicle development and high-performance IT infrastructure is a positive development. Shifting the focus to how AI can optimise existing business activities can temper the often-exaggerated hype surrounding the idea that AI can transform the entire economy.

Finally, European institutions and member states must incentivise European companies to adopt AI. The forthcoming Apply AI Strategy provides an opportunity to ensure that incentives, such as tax credits, are provided to encourage companies to adopt and test AI applications in their daily operations. These could reimburse the costs of purchasing tangible (hardware) and intangible (software) capital goods related to digital and AI transformation. Moreover, to attract talent, tax breaks could be offered to European companies that hire staff involved in digitalisation and the transition towards AI. European institutions and member states should incentivise companies, particularly large ones, to join AI factories and acquire a stake in this computational structure. At present, there is rightly a great deal of interest in AI factories and gigafactories, but they are mainly intended for AI developers or start-ups. Large companies, which currently work mainly with non-European cloud providers, should instead be encouraged to join forces in these initiatives. Businesses are, after all, the ones that will run large AI workloads (e.g. drug discovery models, autonomous driving simulations) on the shared infrastructure. To achieve this, they require assurances that their data is secure and that their efforts are interoperable with those in other AI infrastructures. Ultimately, what matters is the final product – the value that the application will generate for the company’s internal production processes or the service or product that will be offered to the market.

In sum, we urge the European public and private sectors to assess the shifting grounds of the AI market, adopting a pragmatic approach to AI infrastructure while simultaneously demonstrating ambition and forward-thinking regarding AI end-user applications.

__________