European Union lawmakers are pushing for the bloc to be a strong player in the generative AI field and are introducing a package of support measures to help local AI startups and scale ups grow.
Artificial intelligence technologies — especially generative AI models that are trained on huge data-sets and can do things like understand natural language and create text, images or audio on demand — are seen as a key area for the bloc’s future success. But Commission officials admit lawmakers were not ready for the compute infrastructure needed to train such AIs.
They say they were caught off guard by the sudden popularity of generative AI tools like OpenAI’s ChatGPT last year. The EU has a great network of high performance supercomputers — like the new MareNostrum 5 — but this strategic infrastructure was mainly for scientific users and uses, not for training the new generation of disruptive generative AI models. This is a big gap in the EU’s AI strategy that lawmakers are trying to fill.
The “AI innovation package”, as the Commission calls the policy bundle that it adopted today, aims to support the growth of Europe’s AI ecosystem with actions in different areas — like high performance computing infrastructure for training models and access to the skills and talent needed for generative AI.
In a statement, Margrethe Vestager, the EU’s digital chief, said: “You need a lot of computing power to make AI. So we want to give SMEs and start-ups special access to the network of European supercomputers. We are committed to AI innovation and innovation with AI. And we will do our best to create a thriving AI ecosystem in Europe.”
“Today, we announce the start of AI Factories, which bring together the ‘raw materials’ for AI: computing power, data, algorithms and talent,” said internal market commissioner, Thierry Breton, in another statement. “They will be a one-stop shop for Europe’s AI start-ups, helping them make the most advanced AI models and industrial applications. We are making Europe the best place in the world for trustworthy AI.”
The Commission had already started the Large AI Grand Challenge in November, a prize to help AI startups with money and supercomputing access. The Commission said that “Today’s package makes this commitment real with many measures to support AI startups and innovation, like a proposal to give special access to supercomputers to AI startups and the wider innovation community.”
The EU’s executive is thinking of creating “AI Factories” near the bloc’s supercomputers — with a mix of policy actions that will bring in AI-focused hardware improvements; actions to increase access to the resource, for startups too; a support “one-stop-shop” for startups and innovators to use “supercomputer-friendly programming facilities and other AI services”; and more support actions to help make AI applications based on General Purpose AI models.
A Commission official said to the press before today’s announcements that “We have many good startups in generative AI and in related areas but they need this computing power and they need other important things for AI.” He said that “It starts with data, then it’s the computing power to train the models with that data and you need the algorithm of course but the fourth thing, I would say, is talent and skills.”
On compute infrastructure, the EU’s executive has suggested changing existing EU rules about the joint buying of high performance supercomputers (also called, the Regulation establishing the European High Performance Computing Joint Undertaking) to make upgrades more flexible.
Commission officials say the suggested change to the rules is meant to let the bloc increase the capacity of its supercomputers more easily than with the current rules.
A second Commission official, who also spoke to the press on background, said that “It is much cheaper to double the capacity of the existing supercomputers by adding AI capabilities now, instead of waiting for another two or three years until we have bought a new supercomputer dedicated to AI.”
The support package is meant to go along with the bloc’s new regulatory framework for AI — also known as the AI Act — which EU lawmakers say is also a supportive measure. But while that law wants to increase the use of AI apps by building user trust, today’s package is more about direct support for AI app makers.
The AI Act has also met some industry resistance. Last year, there was a lot of lobbying for regulatory exceptions for general purpose AI from companies like France’s Mistral1. So the Commission might be feeling the pressure from the industry to offer some attractive incentives too. (Mistral’s CEO and co-founder, Arthur Mensch, liked the EU’s AI support package, saying: “It works well.”)
On top of that, the AI Act is still a draft law, waiting for a vote in the Council to confirm the agreed text from last month — so there is still a risk of something going wrong with the main AI law, even at this late stage.
The Commission is showing confidence in public that the AI Act will pass these final steps and be adopted. In private, Commission officials are slightly more careful and say they are “hopeful” the regulation will be adopted.
Still, today the EU’s executive is moving forward publicly in the hope of the main AI law being approved by taking steps to create a new public body, called the AI Office, which will be in charge of enforcing the Act on general purpose AIs.
The AI Office, which will be part of the Commission, will also have other roles. The Commission sees it as a “central coordination body” for AI policy at EU level — working with other departments, EU bodies, Member States and the stakeholder community.
The Commission says in a press release that “It will have an international role and support the EU approach to AI governance and take part in the EU’s international activities on AI.” It also says that “The AI Office should learn more about AI and help AI adoption and innovation.”
The decision to set up the AI Office starts today — and it should start working in the next few months. But the suggested changes to supercomputing buying rules will need the support of the European Parliament and Council to go ahead.
Demand crunch
The bloc’s supercomputers might seem like a useful advantage in the global AI race, but they are already too busy, according to Commission officials. They also were not made for this kind of AI model training. So they need upgrades — both in hardware and capacity — to improve the bloc’s chances.
The EU also knows that AI startups will need a lot of support to use the resource well. That’s why support services are another part of today’s package.
The bloc’s supercomputing infrastructure is free for research and AI startups doing model pre-training or training. So the planned spending to improve the EU’s supercomputers and make the infrastructure ready for training generative AIs could, in theory, help make the competition fairer — instead of the uneven situation we have now, where startups with a close relationship with a hyperscaler and their cloud computing infrastructure can do better than others.
But many AI startups want to use the free EU compute infrastructure, which will increase the demand for slots on the supercomputers. So the EU’s supercomputing-for-AI plan will depend on getting key infrastructure upgrades — and quickly.
The Commission’s idea to reduce the demand problem is to change current rules to make it easier to upgrade the existing supercomputers. These upgrades will focus on changing the infrastructure to fit the needs of generative AI model makers. So upgrades here mean adding a lot of racks full of accelerators (like GPUs). And some changes to components to make sure data can move well.
One Commission official said that “With this flexibility on upgrading the existing infrastructure, in a year, we can almost double the capacity and give the AI users the access they need to train big models and stay competitive in a very fast changing market in the world.”
The EU can change rules to make upgrades to its supercomputers more flexible, but it can’t be sure it will get all the GPUs it needs — because there is a lot of global demand for them.
Commission officials said this is a big concern when they talked to the press. They said there might be delays, because of the high demand for the chips. But they also said the bloc has been asking for promises from Nvidia, the main GPU maker, about supercomputer buying.
In the long term, the EU also hopes other actions it’s taking to get more chips, like the Chips Act, will help get European accelerators to reduce the demand problem. But the official said its short term goals to make its high performance compute infrastructure ready for the generative AI boom might fail because of chip supply issues. They said “There’s no guarantee we will not have the same shortage or delay as other Nvidia clients. But we’re working on some ways to deal with it, short term or long term.”
The “AI innovation” package also wants to help European generative AI startups grow in other ways.
For example, EU Member States that want to have a special AI-optimized supercomputer will have to promise to give a direct connection to data centers to make sure there is “quick and easy access” to the infrastructure, the Commission said. They will also have to promise to give support services to help startups use the infrastructure. The official said “All this should be available, as a one stop shop, for any user wherever he lives in the Union.”
The EU will also give more money to AI startups — through the existing Horizon Europe and the Digital Europe programs — with a focus on generative AI. The Commission expects this to bring in more public and private investment of about €4BN by 2027.
The EU also wants to get more public and private investments in AI start-ups and scale-ups, with venture capital or equity support (like new initiatives of the EIC accelerator Programme and InvestEU).
To deal with the data part of the AI challenge, the support package wants to speed up the making and using of the Common European Data Spaces, which the Commission’s February 2020 data strategy talked about, to share them with the “AI community”.
The Commission is also talking about a “GenAI4EU” initiative to help make “new use cases and emerging” AI applications in Europe’s 14 industrial ecosystems, and the public sector. It said “Application areas include robotics, health, biotech, manufacturing, mobility, climate and virtual worlds.”
As another step, two European Digital Infrastructure Consortiums (or Edics) are being set up — one, called the Alliance for Language Technologies, will work on making a common European infrastructure in language technologies to deal with a lack of European languages data for training AIs. The aim is to help make European large language models (LLMs), and keep Europe’s “linguistic diversity and cultural richness”, as the Commission says.
The other Edic, called CitiVerse, will work on making AI tools for city planning and management — or “Local Digital Twins for Smart Communities”, as the PR says — to use generative AI to help cities model and improve processes like traffic management and waste management, or see how a planned development might affect a city area.
Responsible AI?
The bloc’s president, Ursula von der Leyen, announced the plan to let AI startups use the EU’s supercomputers for model training last fall. She said the access would be given to “responsible” AI startups — which agree to follow the EU’s risk-based governance model for AI applications.
The Commission’s PR today says something similar — with the announcement that the package of measures will help European startups and SMEs “in the making of trustworthy Artificial Intelligence (AI) that respects EU values and rules”. And the word “trustworthy” is repeated several times in the announcement.
But the EU has not passed the AI Act yet — and, even if it’s passed soon, the law will not be fully in force for several years — so how will AI startup trustworthiness be checked, in order to decide who gets supercomputer access, in the meantime?
The bloc has been working with industry on a stand-in (not legally binding) AI Pact — which is meant to act as a temporary code until the AI Act is in force. And one of the ways AI startups that want to get free training time on the supercomputers will be able to show they’re “trustworthy” enough is by signing up to this Pact, according to the Commission.
But, as we reported last month, the EU is already giving some AI startups slots on the supercomputers — like France’s Mistral1 — even though the AI Pact is not officially working yet… So there seems to be a gap between the bloc’s publicity, about access for “responsible/trustworthy” AI startups, and how the early stages of this program are working.
The EU’s need to catch up in generative AI seems to be the main reason behind its early moves here, while hard AI governance rules are still being discussed. But if the AI Act does not get passed, the AI Pact will be just a ‘pinkie promise’, and the EU’s promise to give taxpayer-funded support only to “trustworthy” AIs will sound empty.
“A Commission official told TechCrunch that “For the AI Pact to really work we first needed to have the full text of the AI Act.” They said that “Now that we have made this big step of the political agreement [on the Act] reached in December — and hoping for adoption soon — this is really a time where we would start making the AI Pact work better — looking for more commitments from companies and from startups.”
They also said that “When these AI-only supercomputers and these AI factories start working then, indeed, the AI Pact can be used as one of the ways to show that the proposal follows EU rules and EU values.”