European Union lawmakers are intensifying efforts to position the bloc as a major player in the generative AI landscape, unveiling a comprehensive package of measures designed to bolster domestic AI startups and scale-ups. The focus is on artificial intelligence technologies, particularly generative AI models that leverage extensive datasets to generate text, images, and audio on demand. Recognizing the strategic significance of AI for the EU's competitiveness, officials admit they have been somewhat unprepared for the necessary compute infrastructure to train these cutting-edge systems.
The swift emergence of generative AI tools, like OpenAI’s ChatGPT, took many by surprise last year. While the EU has an impressive array of high-performance supercomputers, such as MareNostrum 5, these resources historically cater to scientific applications and have not been fine-tuned for the types of training required by disruptive generative AI models. This oversight is a critical area that EU lawmakers are eager to address.
Dubbed the "AI Innovation Package," this policy framework aims to strengthen Europe's AI ecosystem through support measures spanning multiple areas—most notably enhancing high-performance computing infrastructure for model training and ensuring access to the essential skills and talent that enable the creation of groundbreaking AI technologies.
In a statement, Margrethe Vestager, the EU's digital chief, emphasized the importance of computing power in AI development, stating: "Significant computing resources are essential for AI creation. We aim to offer SMEs and startups privileged access to our network of European supercomputers. Our commitment is to foster innovation in AI and create a vibrant AI ecosystem in Europe."
Internal Market Commissioner Thierry Breton also announced the introduction of "AI Factories," which will serve as centralized hubs providing the 'raw materials' for AI—such as computing power, data, algorithms, and talent. "These factories will act as one-stop shops for Europe’s AI startups, allowing them to develop advanced AI models and applications. We are determined to make Europe the premier location for trustworthy AI."
Previously, in November, the Commission rolled out the Large AI Grand Challenge to support AI startups financially and provide supercomputing access. This new package puts that commitment into action, broadening support mechanisms for AI startups and innovation, including a proposal for privileged access to supercomputers for these startups and the wider innovation community.
The Commission plans to establish “AI Factories” around the EU's supercomputing centers through a combination of policy initiatives that will facilitate AI-focused hardware upgrades, enhance access for startups, and offer a support “one-stop shop” for startups to utilize supercomputer-friendly programming facilities and other AI-enabling services. Additional support measures will promote the development of AI applications based on General Purpose AI models.
"We have numerous promising startups in generative AI and related fields, but they require access to computing power and other essential resources,” a Commission official stated ahead of today’s announcements. “The essential components start with data, followed by the computing power needed to train models on that data, and of course, effective algorithms, with talent and skills being the final integral part.”
Addressing compute infrastructure, the EU has proposed amending current legislation regarding the joint procurement of high-performance supercomputers to allow greater flexibility for upgrades. Commission officials suggest this amendment aims to significantly increase the capacity of supercomputers more adaptively than current regulations permit.
"It’s far more cost-effective to double the capacity of existing supercomputers by adding AI capabilities now rather than waiting two to three years for a new AI-specific supercomputer," a second Commission official elaborated during a press briefing.
This support package is designed to complement the EU’s developing regulatory framework for AI, known as the AI Act, which encourages AI application adoption by fostering user trust. However, the current package focuses on more immediate support for developers of AI applications.
While the AI Act has faced some resistance from the industry, there is ongoing pressure on the Commission to provide incentives for general-purpose AI, as seen with lobbying from firms like France’s Mistral. Mistral's CEO, Arthur Mensch, praised the strategy, affirming it is effective.
Recently, a political agreement was reached on the AI Act, which remains in draft form and requires a vote in the Council to confirm the agreed-upon text. This means complications could still arise for the flagship AI legislation even at this stage. Publicly, the Commission expresses confidence that the AI Act will pass, while officials privately convey they are “hopeful” it will secure approval.
In anticipation of the AI Act’s finalization, the EU is taking steps to create a new public body, the AI Office, tasked with enforcing the Act for general-purpose AIs. This body will also serve as a central coordination point for EU-level AI policy, collaborating with various stakeholders.
The AI Office will promote the EU’s approach to AI governance and support international AI initiatives, while also improving knowledge and understanding of AI to encourage innovation across the region.
The establishment of the AI Office is now formalized, with its operations starting in the coming months. However, the proposed revisions to supercomputing procurement rules require endorsement from the European Parliament and Council to proceed.
Demand Crunch
Despite having a network of supercomputers for generative AI model training, these resources are significantly overbooked, with demand currently exceeding capacity by at least twofold. Additionally, the supercomputers weren’t initially designed for AI model training, necessitating urgent upgrades to enhance the bloc’s competitive standing in the AI race.
Recognizing that AI startups may require additional guidance to maximize their use of these computational resources, the support package will also provide necessary assistance. Research access to the supercomputing infrastructure remains free, with AI startups benefiting from waived fees during model pre-training or training phases. Thus, investments aimed at upgrading the EU’s supercomputers and configuring them for generative AI training could bridge the competitive gap currently faced by startups reliant on partnerships with hyperscalers, which currently dominate the market.
Nonetheless, many AI startups seeking access to free EU compute infrastructure will inevitably increase demand for the limited supercomputer resources. The success of the EU’s supercomputing strategy will depend on effectively delivering critical infrastructure upgrades—swiftly.
The Commission plans to alleviate the demand bottleneck by amending rules to enhance the flexibility of supercomputer upgrades, tailored to the requirements of generative AI developers. This entails substantial hardware improvements, including adding numerous racks of accelerators (GPUs) and optimizing components for efficient data transfer.
“With this newfound flexibility for upgrades, we can nearly double the capacity within a year and provide users the necessary access to train large models, keeping pace in an ever-evolving global marketplace,” a Commission official indicated.
Though the EU can adjust legislation to optimize supercomputer capabilities, securing the necessary GPUs remains a challenge, as global demand for these components is extremely high. During a recent press briefing, officials acknowledged this significant concern, understanding that acquisition delays are possible due to escalating demand. The bloc has been actively seeking commitments from leading GPU manufacturer Nvidia regarding supercomputer procurement as a proactive measure.
In the long term, the EU is also optimistic that initiatives like the Chips Act will help strengthen chip availability, creating a European supply of accelerators to alleviate demand pressures. However, officials admit that their short-term ambitions to make high-performance compute infrastructure suitable for the generative AI boom could be hindered by chip supply limitations. “While we cannot guarantee we won’t face similar shortages or delays as other Nvidia clients, we are implementing several mitigation strategies, both immediate and long-term,” officials noted.
Aside from addressing compute constraints, the "AI Innovation Package" aims to facilitate the growth of European generative AI startups through various initiatives. Member States intending to host dedicated AI-optimized supercomputers will need to ensure a direct connection to data centers for "quick and easy access” to resources. Additionally, these states must provide comprehensive support services, enabling startups to fully leverage this infrastructure. "All these resources should be accessible as a one-stop shop for users throughout the Union,” the official reiterated.
Further financial resources will be made available for AI startups through existing programs, such as Horizon Europe and Digital Europe, focusing specifically on generative AI. The Commission anticipates this will lead to an overall public and private investment increase of around €4 billion by 2027.
Moreover, the EU will strive to generate more public and private investments in AI startups and scale-ups through avenues like venture capital or equity support, utilizing new initiatives from the EIC Accelerator Program and InvestEU.
To tackle the data dimension of the AI ecosystem, the support package aims to expedite the development and deployment of Common European Data Spaces, outlined in the Commission’s February 2020 data strategy, making them accessible to the AI community.
Additionally, the Commission has introduced the "GenAI4EU" initiative, aimed at fostering novel use cases and emerging AI applications across Europe’s 14 industrial ecosystems as well as the public sector. This initiative will cover areas such as robotics, health, biotechnology, manufacturing, mobility, climate, and virtual worlds.
In another significant move, two European Digital Infrastructure Consortiums (Edics) are being established. One, the Alliance for Language Technologies, will focus on creating a common European infrastructure for language technologies to address the shortage of training data in European languages. The objective is to enhance the development of large language models (LLMs) while promoting Europe’s linguistic diversity and cultural richness.
The second Edic, CitiVerse, will concentrate on developing AI tools for city planning and management—focusing on "Local Digital Twins for Smart Communities"—utilizing generative AI to help cities model and optimize functions like traffic and waste management.
Responsible AI?
Last fall, EU President Ursula von der Leyen hinted at plans to allow AI startups to utilize EU supercomputers for model training, asserting that access would be granted to "responsible" AI innovators who adhere to the EU's risk-based governance framework for AI applications.
The OECD confirms that today’s announcement similarly indicates the package of measures will support European startups and SMEs in developing trustworthy AI in alignment with EU values and regulations. The term "trustworthy" is echoed several times throughout the announcement.
However, with the AI Act still pending, and even if adopted soon, the law won't be fully enforced for several years. This raises questions about how to determine the trustworthiness of AI startups in the interim to allocate supercomputer access.
The EU has been collaborating with industry to establish a non-legally binding AI Pact as a provisional code of conduct until the AI Act is operational. AI startups aspiring to access free supercomputing training time may demonstrate their "trustworthiness" by signing onto this Pact.
Despite this, reports reveal that some AI startups, including Mistral from France, have already received access to supercomputers ahead of the AI Pact’s implementation. Therefore, a disconnect appears to exist between the EU’s assurances of access for "responsible" AI startups and the current operational reality of this program.
The urgency for the EU to catch up in the generative AI domain seems to be the primary catalyst behind these initiatives, while rigorous AI governance policies are still under review. If the AI Act fails to materialize, the AI Pact may ultimately serve merely as a symbolic gesture, undermining the EU’s commitment to directing taxpayer funding towards "trustworthy" AI solutions.
“To fully activate the AI Pact, we first needed the AI Act's text finalized,” a Commission official explained when discussing this issue. “With the recent political agreement reached in December, we anticipate that now we can start to operationalize the AI Pact further—seeking increased commitments from companies and startups.”
“As dedicated AI supercomputers and factories come online, the AI Pact can then be leveraged as a validation tool to ensure compliance with EU standards and values,” they added.
EU to Empower Trusted AI Startups with Supercomputing Access
EU Enhances Support for AI Startups to Leverage Supercomputers for Model Development
AI, EU AI Innovation Package, EU Generative AI, EU Supercomputers for AI, Europe, Government & Policy, Startups
Ranking
