Executive summary
European Union policymakers want to close the artificial intelligence innovation gap with the United States, as a way to accelerate lagging productivity growth. The EU focus is on expanding an existing supercomputer network with more AI hardware and computing infrastructure, with taxpayer support. However, this computing infrastructure is not adapted to AI modelling. The cost of catching up with leading big tech AI computing centres is already prohibitive for EU budgets, and is set to become even more so.
The hardware focus overlooks missing EU markets for complementary services that are required to set up a successful AI business: large-scale business outlets for frontier generative AI models to generate sufficient revenues to cover huge fixed model training costs, hyperscale cloud-computing infrastructure and private equity financing for AI start-ups. In the absence of (or with insufficient) complementary services markets in the EU, start-ups are forced to collaborate with US big tech firms. Injecting taxpayer subsidies to make up for these missing markets may further distort EU markets. Regulatory compliance costs, including uncertainty about the implementation arrangements for the EU Artificial Intelligence Act, add to market problems.
The EU should address a wider range of market failures in its policy initiatives. It should strive to increase productivity growth below the AI technology frontier, by facilitating investment and applications of AI-driven services produced by derived and specialised generative AI models, or AI-applications that build on top of existing generative AI models. Building these below-frontier AI applications requires far less computing capacity and less heavy investment costs. Promoting the uptake of AI application services across a wide range of industries can substantially stimulate productivity growth.
That requires a razor-sharp focus on pro-innovation guidelines, standards and implementation provisions for the EU AI Act, shortening the Act’s regulatory uncertainty horizon as much as possible, and facilitating collaborations between EU AI startups and big tech companies. Widening and deepening the EU private equity and venture capital market would also be very helpful.
1 Introduction
In the first half of 2024 alone, more than $35 billion was invested globally into artificial intelligence startups
. The European Union attracted only 6 percent of that. The EU is doing better in AI patents and in training AI researchers, but the outputs from this tend not to stay in the EU, but rather to flow to the United States (Renda, 2024). Unsurprisingly, this situation has triggered considerable debate in EU policy circles about what can be done at EU level so the EU can catch up with the US and China on AI, in particular by developing its own AI models, fostering more AI startups, accelerating the uptake of AI-based services in the EU economy.
In this context, the European Commission in January 2024 published a package of proposals, decisions and plans to support AI startups
. This seeks to capitalise on the European High-Performance Computing (EuroHPC) network of supercomputers – very large, high-performing computers – used primarily for scientific research. The Commission proposed an amendment to the network’s governance rules to facilitate collaboration with the private sector – that amendment has since been adopted (Regulation (EU) 2024/1732). The plan is that EuroHPC should be the core of a network of ‘AI factories’ for the development by EU startups of large-scale general purpose AI models and applications.
This approach recognises that these supercomputers need to be upgraded to AI capabilities, to be financed equally by the EU and the computer-hosting EU countries
. But the EuroHPC budget of €7 billion for 2021-2027 remains for now unchanged
.
The AI computing infrastructure budget could be increased very substantially if the Commission and EU countries listen to former Italian prime minister and European Central Bank governor Mario Draghi. His September 2024 report on the future of European competitiveness, produced to steer EU policy in the next five years (Draghi, 2024), attributed the EU’s weak productivity growth to insufficient investment and uptake of digital technologies, including AI.
His proposed remedies include private and public investment in EU-developed general and sectoral AI models, upgrading EuroHPC, creating an AI incubator similar to that of the CERN nuclear and particle physics laboratory, creating EU-wide large data pools for AI model training, facilitating consolidation among EU cloud providers to create hyperscale computing infrastructure and more financial resources for quantum computing. Draghi (2024) also recognised that the EuroHPC computers cannot compete with US-based hyperscale AI firms and proposed to allocate €100 billion for AI infrastructure.
All this suggests a consensus in EU policy circles that catching up on AI requires public sector involvement and subsidies. There has been less analysis, however, of why the EU AI value chain and business ecosystem have ended up falling behind the US and China in terms of AI model development
and uptake in services industries, and why this should justify public sector involvement and subsidies. There is even less debate on how these problems could be addressed through structural reforms that could incentivise more private investment in EU AI industries.
This Policy Brief explores why EU AI investment has fallen behind the US and the types of market failure that may have led to that situation. We ask how the EU should position itself in the competition over AI and discuss two possible responses. Should the EU try to catch up with the US, reach the AI technology frontier and develop its own AI capacities, independent of US big tech firms? Or can the EU prosper below the AI technology frontier, in derived AI products and services markets? We also look at the geopolitical context and the risks of EU dependence on US big tech.
2 Building AI models on existing EU supercomputers?
It is clear that the EU is running behind the US in digital technology investment and uptake in general, and in AI specifically. It is less clear whether the response should be to invest taxpayer money in physical infrastructures for AI, as advocated by the Commission policy initiatives and Draghi (2024). It might be possible to resolve some market and regulatory failures in AI-related markets with public money, but many other problems cannot be resolved this way. In this section, we discuss the main considerations that should be factored in to the EU approach to AI.
2.1 Computing hardware issues
The EuroHPC network of nine supercomputers is not up to the task of delivering a state-of-the-art AI computing infrastructure for commercial use. These computers were designed for scientific research, not for training of general-purpose AI models or generative AI models like ChatGPT
. Their hardware architecture is not suitable for that purpose. They have no more than a few thousand Nvidia graphics processing units (GPUs) that play a central role in AI model training. This is a tiny capacity compared to Meta’s most advanced AI computing centre, which reportedly contains 600,000 Nvidia AI chips
.
Hobbhahn et al (2023) explained how AI hardware differs from classic computing architectures that revolves around central processing units (CPUs). Handling the massive amounts of data in GenAI model training requires GPUs. Nvidia became successful in AI hardware because of its original specialisation in GPUs for gaming applications. Handling data traffic between many thousands of GPUs requires extensive communication bandwidth between GPUs and memory storage, though one way to reduce computational requirements can be to reduce the number of digits behind the decimal point in calculations
. AI developers are increasingly designing their own dedicated hardware, including for specific applications such as inference, meaning the making of predictions based on newly supplied data after the model has been trained.
2.2 AI infrastructure costs
Nvidia AI chips each cost more than $30,000. For Meta’s most advanced computing centre with 600,000 of these chips, this amounts to $18 billion for the dedicated AI chips alone, excluding other hardware needs. In other words, the cost of chips for a single computing centre is more than twice the current EuroHPC budget.
Moreover, technological progress in AI chips is so fast that the latest generation of AI chips will be outdated and written off in less than a year (Hobbhahn et al, 2023). Spending $162 billion per year (ie nine EuroHPC supercomputers x $18 billion/year) is simply beyond the financial resources of the EU. Even if the EuroHPC network were to be upgraded to train state-of-the-art AI models, it would still have a hard time running these models on a daily basis to respond to user queries because that requires additional investment in a different type of inference accelerator chip, such as NVIDIA’s Jetson processors, to reduce the cost of responding to user queries.
The costs of training state-of-the-art generative AI models (ie those that can produce new images, video, audio or text based on prompts) are exploding, running into hundreds of millions of euros (Martens, 2024a). Cottier et al (2024) estimated that GenAI model training costs are increasing exponentially by a factor 2.4 to 2.6 per year, or around 240 percent per year from 2016 to 2023. Extrapolating the costs of the largest frontier models now to 2030 leads to an estimated training cost for a single GenAI model of $60 billion.
New frontier GenAI models are coming out every week. Cottier et al (2024) also estimated the cost of AI computing infrastructure at ten times the cost of model training. That infrastructure can be used to train several models but the hardware amortisation rate is estimated at 140 percent per year, or 100 percent depreciation in 8.5 months. By that time, a new generation of AI computing chips will have arrived with superior performance. Infrastructure costs for GPT4 by the end of 2023 may have been as high as $800 million. Extrapolation could push that figure up to $500 billion by 2030. This is beyond the financial reach of EU public and private budgets. Even the largest US big tech firms will have a hard time financing this, and may be forced to collaborate.
2.3 Integration of AI into business models
To succeed, AI startups require not only computing infrastructure but also an important complementary asset: a business model to generate revenue that pays for these costs. Most AI start-ups have close collaboration agreements with US big tech firms, to access hyperscale computing capacity and because they can directly plug their AI models into big tech’s established business models to generate revenue. Microsoft uses AI in its business software, Meta uses it in advertising and Google in search, advertising and many other services.
AI startups can also try to launch their own business models from scratch to generate sufficient revenue to finance AI model development. But this is very hard. Even successful start-ups such as OpenAI have a hard time generating sufficient revenue, despite running a successful business model
. An upgraded EuroHPC network may have the hardware capabilities but offers no commercial outlet channels. EU startups would have to move their AI models to incumbent big tech firms to generate revenue to finance the fixed training costs.
2.4 Missing markets for EU AI startups
In the absence of home-grown big tech platforms, EU AI startups have to turn to US companies with global business models that have sufficient market scale to amortise the huge fixed costs of training generative AI models. In the absence of sufficient domestic private equity and venture capital in the EU, US markets and big tech firms can provide financial resources for EU startups. EU public funds might perhaps replace private equity but cannot replace business outlets for AI models. Alternatively, EU AI startups could focus on specific AI model applications, derived from big generative AI models. This avoids very high model training costs and makes it easier to plug derived models into existing services markets, where there is demand for these specific applications.
The absence of business model considerations in the Commission’s January 2024 ‘AI factories’ initiative, and in the AI recommendations in Draghi (2024), is problematic
. But the exclusive hardware focus of these plans is not surprising. Eckert (2024) presented an insightful historic overview of EU digital policies over the past 40 years. A recurrent pattern has been the emphasis on telecoms infrastructure and hardware in general, and the almost total absence of digital services markets and business model considerations. Draghi (2024, Part B, Figure 4) showed how the value of telecoms services has become negligible compared to digital services markets. Nevertheless, his recommendations focused on telecoms, cloud and AI hardware, and do not mention digital or AI services markets. More than anything, four decades of path dependency in EU digital policies may have contributed to an ever wider yawning gap between EU and US digital performance – which still continues today.
Draghi (2024) pointed out that the EU should do more to create its own hyperscale cloud computing infrastructure in support of generative AI model development, and reduce dependence on the US big tech firms that currently dominate the cloud services market in the EU
. There may be competition failures in EU cloud computing services, another important complementary input for AI. A few big tech players can leverage their positions in cloud software- and platforms-as-a-service, rather than just offering basic infrastructure-as-a-service
. This increases entry barriers for smaller EU cloud service providers, leaving them unable to expand their computing infrastructure, which would be suitable for AI model training (Ennis and Evans, 2024; Biglaiser et al, 2024). Throwing taxpayer money at this problem is unlikely to be a good solution, however. Draghi (2024) recommended consolidation among smaller EU cloud players. That does not solve the problem of lack of complementary software and platform services.
2.5 Derived AI model markets are very competitive
There is no indication of a market failure that would require public policy intervention, let alone taxpayer subsidies, in derived and special applications of AI models. Draghi (2024) recommended that EU AI funds could support European AI startups to develop specific industry or company application models. That market is already very competitive (Martens, 2024b). While more than a dozen new state-of-the-art GenAI models are released every month, more than a dozen derived models are released per hour
. Just as app stores for mobile phones contain millions of special-purpose apps, there are now also millions of industry-, sector- or company-specific applications of the ChatGPT model in the OpenAI store.
For example, there are ChatGPT applications that help consumers with their shopping questions or financial decisions. Developers of these applications make them widely available to anyone who can use them. A derived model is created when a company uploads its own proprietary data into ChatGPT for specific marketing, logistics or industrial process applications within the company. Since they run on proprietary data, these models are of course not made widely available.
2.6 Risk of regulatory failure
Policy intervention may create new market distortions in AI services markets. The amended EuroHPC regulation (Regulation (EU) 2024/1732) now allows collaboration between public and private computing and cloud services providers. Commercial firms can access publicly-owned computers.
This raises the question of how scarce computing capacity will be allocated between users, at what price and under what conditions. The amended EuroHPC regulation does not explain this. Will authorities use auctions for commercial applications and sell capacity at market prices? Or will there be a subsidy component in pricing, thereby opening the door to unfair competition with private providers? How will capacity be allocated between paid commercial and presumably unpaid non-commercial use, for instance for scientific projects, which the EuroHPC network was involved in from the start? More importantly for AI start-ups, what happens after the training of their AI model has been completed? Will they have guaranteed access for inference, daily running of their models? Can they easily scale up capacity when their startup rapidly expands? Computers may be provided by the public sector but they are not non-rival non-excludable public goods. They are rival and easily excludable.
The European Commission and Draghi (2024) claim that these AI policy initiatives can capitalise on EU regulation, including the EU AI Act (Regulation (EU) 2024/1689; see Box 1), the general data protection regulation (GDPR, Regulation (EU) 2016/679) and other data regulations. The claim is that these EU regulations attract investment because they give users confidence and regulatory certainty.
The available empirical evidence, however, does not support that view. There is considerable evidence that the GDPR has reduced investment in consumer-oriented online services in the EU (Demirer et al, 2024; Goldberg et al, 2023; Jia et al, 2023; Peukert et al, 2024). Consumers may be better off without some of these privacy-infringing services, though that may not be the case for all.
There is also evidence that EU regulation is limiting EU access to AI services. At the request of the Irish Data Protection Commission, Meta held back the roll-out of its most advanced AI models in the EU
. European data regulators have doubts that the legitimate interest clause in the GDPR (Article 6(1)(f)) constitutes a sufficient legal basis for Meta to use publicly posted messages on its Facebook and Instagram social media platforms as inputs for AI model training. Other US AI developers, including Apple, Google and OpenAI, face similar EU uncertainty about the use of personal data for model training
. The Irish Data Protection Commission launched an enquiry into Google’s AI services
. Social media text has become an important source of AI model training data when other sources are insufficient to meet the training requirements of very large AI models, especially in relation to less-widely spoken languages, for which the available volume of human text data is limited. Regulatory uncertainty about this alternative source is holding back innovative AI services from entering the European market.
Longpre et al (2024) showed that, following the release of ChatGPT in 2023, copyright holders are making more active use of their right to exercise an opt-out for content from use for AI model training, granted to them under the EU AI Act and the EU Copyright Directive (Directive (EU) 2019/790). This has led to a 20 percent to 30 percent reduction in the availability of AI training data.
Strict enforcement of data privacy consent rules could have a similar negative effect on the availability of AI model training data. AI model training is already running into data shortages (Martens, 2024b). This could especially affect small language communities in the EU that already suffer from insufficient language training data. Moreover, the AI Act (Box 1) generates not only high compliance costs for model developers and deployers, but also considerable regulatory uncertainty regarding the specific implementation rules for copyright and privacy protection. The finalisation of the AI Act in mid-2024 was only the start of a regulatory process that will take several years to complete dozens of implementation guidelines and enforcement standards, including on copyright and data privacy.
