Trio of reports say the EU is betting too heavily on generative AI

Three new reports have criticised the EU’s artificial intelligence strategy, including its plan to spend €20 billion on enormous supercomputers known as gigafactories to train and run large language models (LLMs), the form of AI that powers chatbots like ChatGPT. 

The European Commission and EU member states will cover up to 35% of the gigafactories’ capital costs, with groups of private firms paying for the rest. Each cluster will contain around 100,000 graphics processing unit (GPUs), the chips best optimised to train and run LLMs. 

But three think tanks, including the European Parliament’s Research Service, question whether the Commission is placing too big a bet on LLMs, which remain prone to hallucination and are proving harder to improve with more computing power, leading to fears of an AI investment bubble. 

“The current generation of models exhibit well-publicised weaknesses that might not simply disappear by using more data and processing power or smarter training,” according to the research service briefing, released in late November. 

The “reliability issues” inherent in generative AI, which includes text-creating LLMs, image generation tools such as Midjourney, or the song-creator Suno, “call into question” EU hopes that AI can drive economic growth and be trustworthy, the briefing, What if generative AI is reaching its limits?, says. 

Yet despite these problems, the EU has continued to focus largely on generative AI, the briefing says, rather than more exploratory research into other forms of AI. These include models that learn from physical environments, or so-called neuro-symbolic approaches, which might prove better at logic and abstraction. 

For example, the EU has committed €700 million to be spent on a GenAI4EU package, “rather than on research and development of alternative and complementary AI approaches.” 

On December 4, Germany’s Sprind innovation agency launched a new challenge to develop fresh forms of AI, deriding LLMs as "combustion-engine AI" that is ripe for displacement by far more energy efficient and transparent "electric-drive AI.” 

A Commission spokesman for research, Thomas Regnier, said in a statement that the EU was “not only investing in generative AI.” 

For example, the next EU Framework Programme for research and innovation, which starts in 2028, could include a “moonshot” mission on “next generation AI” that is modelled “on the laws of nature and grounded in physics and biology.” The European Innovation Council is also launching a new challenge to develop physically-aware AI, worth €6 million in the first stage. 

Over-hyped

Two other think thanks have also recently criticised the EU’s generative AI focus, including its gigafactories plan.  

It its report EU Plans for AI (Giga) Factories: Sanctuaries of Innovation, or Cathedrals in the Desert? the Brussels-based think tank CEPS says that the EU is “adhering to the dominant, possibly over-hyped wave of generative AI” and should instead support “more trustworthy approaches.”

Nicoleta Kyosovska, one of the report’s authors, said she was “sceptical” that the returns to generative AI would justify the current level of investment. Instead, generative AI might need to be combined with other, new forms of AI to become more useful. 

“It's really hard to forecast where this will go,” she said, but “we need something completely new.”

The problem for gigafactories is that they risk being “inflexible” as processing units and memory chips evolve, said Kyosovska.  

Never realistic

A third report, from the Berlin-based technology think tank Interface, also casts doubt on the viability of gigafactories, saying that there are no big European AI companies ready to come in and train leading models on them. 

When the gigafactories idea was launched in February this year, Commission President Ursula von der Leyen said that “they would develop the most advanced very large models.”

However, “the original goal of building five AI gigafactories to train and deploy frontier AI models is no longer really on the table, at least in the short-term,” said Julia Hess, one of the authors of Built for Purpose? Demand-Led Scenarios for Europe's AI Gigafactories. 

“This goal was never very realistic given that Europe currently has only one frontier AI lab, Mistral,” she said, referring to the Paris-based AI firm. 

CEPS’s Kyosovska agreed that there’s a risk of a lack of demand. “I don't think demand will necessarily come if you build these,” she said. “It's just very unclear to me exactly who will be doing this training.” 

The Interface report says that, rather than training leading edge models from one anchor AI firm, gigafactories will have to cobble together enough demand from small to medium sized AI training projects, for example, training specialised industrial AI models, or conducting academic research. 

Should start small

However, this will put the gigafactories in competition with so-called neoclouds, existing firms such as Coreweave that build AI-focused data centres and lease capacity on demand.

Neoclouds “already offer significant flexibility: hourly GPU rental, surrounding services and the ability to adjust quickly to changing workloads. It is unclear whether the gigafactories will be able to provide comparable flexibility,” said Hess. 

Related articles

• The EU’s AI gigafactories need a rethink, says industry
• Explained: the EIC advanced innovation challenges pilot

What’s more, some neocloud firms, such as Nebius in the Netherlands, are already run by EU-based companies, undermining the argument that gigafactories provide a unique advantage in terms of technological sovereignty, the report says. 

Overall, Hess also thinks that the Commission is betting too heavily on generative AI. Instead, the EU should “start small” and develop more “modest capacities” in AI computing initially, and then adjust plans as the technology demands it. 

“This would avoid idle capacity and allow the gigafactories to remain flexible in the face of future AI innovations, rather than locking in a single technology path too early,” she said. 

Build in the cold

Both CEPS and Interface also think the gigafactories should be concentrated in northern states, such as Sweden and Finland, where there is cheap and green energy, urban areas with a decent amount of AI expertise, and cold weather, which will reduce data centre cooling costs. 

But CEPS worries that the EU distributes the gigafactory sites across member states as a form of political compromise.

“No decision on the location of AI gigafactories has been made yet,” said Regnier in a statement. 

Powerful signal

Asked whether there will be enough demand for the gigafactories, the Commission said there had been an “overwhelming response” to a call for interest in helping to set them up. In June, it announced 76 expressions of interest, although the details remain private, and are not binding. 

“These submissions come from industry leaders, investors and member states, with 60 sites proposed across 16 member states. This is a powerful signal,” said Regnier. 

But for Kyosovska, these interested firms may just be “following the hype,” and still assuming that “more compute means better models,” even though scaling up LLMs had led to diminishing performance returns. 

With the business case in her view unclear, Kyosovska said she thought the Commission had announced the gigafactories project “to be taken seriously” by the rest of the world. “It’s not a bad thing to be taken seriously,” she said. “But AI is becoming more of a goal [in itself] than a tool for something.”