Questions remain over the business case for the €20 billion plan, amid wider doubts over the profitability of large language models

Photo credits: Growtika / Unsplash
The EU’s €20 billion plan to build so-called gigafactories to train world-leading AI models has come in for criticism from industry groups, who worry there isn’t a plan for what they will be used for or how they will be financially sustainable.
In February, as the EU feared it was falling behind the US and China in AI, the European Commission said it would invest €20 billion to help create gigafactories containing around 100,000 chips to create “the most complex, very large, AI models.”
The EU and its governments will fund up to 35% of the capital expenditure to build the gigafactories, with consortia of private firms footing the rest of the bill, as well as covering operating costs. In June, the Commission announced it had received 76 expressions of interest to run gigafactories, and plans to release a formal call later this year.
However, many in Europe’s digital industries have questions about exactly what the gigafactories are for, and if they will be financially sustainable.
“No one has really asked the question, do we really need them?” said Agata Hidalgo, European affairs lead for the start-up association France Digitale. “Where is the demand for it going to come from? Who's going to maintain this?”
Given the huge amount of public funding committed, “maybe it would have been better to think first on what is the need that we really want to answer,” she said.
The gigafactories are touted as the massive hardware boost Europe needs to train its own large language models (LLMs), which underpin chatbots like ChatGPT.
While European firms such as Paris-based Mistral have trained their own LLMs, the field is dominated by US and Chinese players. Leading edge LLMs are enormously expensive to train. Costs are opaque, but are estimated to be in the hundreds of millions to a billion dollars.
But it’s unclear whether any European firms currently have this kind of money to train an LLM using a gigafactory, said Jörg Bienert, president of the German AI Association. “The question is, who will come,” he said. “I don't see any single corporation that will be able to spend the money in order to train a large LLM on its own.”
While Bienert said Europe certainly needs more AI hardware, “I'm missing the overall picture and the overall strategy,” he said of the gigafactories initiative.
No business case?
In addition, there are doubts that building and running gigafactories will be profitable for operating consortia.
“We hear from most of the companies that they don't see a business case in gigafactories,” said Kai Zenner, head of office to Axel Voss, a German MEP with the European People's Party group and a leading voice on EU digital and AI policy.
Companies have told the Commission of these concerns, but the Commission “had an order from upstairs to do it, and now they need to do it,” he said.
Firms are forming consortia to get involved in the gigafactories initiative, said Zenner, but were largely doing this for “political reasons,” to satisfy national politicians, rather than a strong business case.
No discussion
Alexander Rabe, managing director of Germany’s Association of the Internet Industry (Eco), also complained of a lack of consultation. “We have not been in discussions before,” he said. The gigafactories announcement was a “typical political situation” where the EU felt it “had to do something,” he said.
Such a huge concentration of 100,000 chips was not needed by European AI companies “in the first step,” said Rabe.
“There are companies who want to train their own models,” he said. “The question is, do they need gigafactories for that?”
Eco’s members would rather see a better framework for private investment in AI hardware, like making energy cheaper and planning permission easier, than public subsidies, said Rabe. “You don't need to spend so much public money, but to enable private companies,” he said.
Strong demand
However, the Commission says interest in running gigafactories is strong. In June, it said it had received many more expressions of interest than expected.
Consortia involving data centre operators, telecoms companies, power suppliers, tech firms and investors have proposed establishing gigafactories at 60 different sites across 16 different EU states, the announcement said.
“It is a clear testament to the very high momentum and interest in AI across Europe, and especially in AI gigafactories,” the Commission said in its announcement. A spokesperson later confirmed that it had received proposals worth €230 billion from governments and the private sector.
Still, these are not yet full bids, but rather expressions of interest. Bienert is sceptical they will all turn into concrete plans. “I really doubt that there will be 70 companies or consortiums, really being able and willing to go into this endeavour,” he said.
Not enough customers
In May, Bienert’s German AI Association released a report questioning the business case for the gigafactories. The private consortia running them risk “being left with a huge investment and ongoing costs without finding enough paying customers,” it warns.
Instead of funding four or five gigafactories, the Commission should start with just two, but with a higher proportion of public funding, said Bienert.
Others in the European AI scene are more optimistic about the business case. “I don’t see them [the operating consortia] spending billions of euros without a clear business case,” said Björn Ommer, a computer vision and learning professor at the Ludwig Maximilian University of Munich. “I’ve spoken with firms that see a clear value in committing to AI within this framework, and others that don’t.”
Related articles
- EU launches AI strategies to boost competitiveness and science
- Europe’s independence hinges on innovation, von der Leyen says
- AI in science: is it useful?
What’s more, there appears to be some flexibility in how the gigafactories might be used. Although they are primarily touted as a means of training leading LLMs, Commission documents show they can be put to a range of other uses, such as fine-tuning models, or simply running models that have already been trained.
“Gigafactories are a strategic investment in Europe’s technological sovereignty, focused not on short-term profit but on securing long-term capability.” a Commission spokesperson said.
Capacity ahead of demand
Some industry bodies are supportive of gigafactories. Connect Europe, a lobbying group for telecoms providers, backs the plan.
“In technology, often you need to invest in capacity ahead of demand,” said a spokesman. “It’s good news that Europe is creating capacity with gigafactories, so that we don’t leave this only to the Americans and the Chinese.”
However, “having a strong business case is key” he said, meaning that “demand-side measures” and “buy European” initiatives were needed to make sure the gigafactories get enough business.
Wider bubble?
Questions over the gigafactories plan come amid rising fears of an AI bubble in the US, with OpenAI chief executive Sam Altman admitting in August that investors were “overexcited” about the technology.
So far, no LLM has proved profitable, with consumer and business subscription income dwarfed by the vast cost of training and running the models, and the often-astronomical wages commanded by AI engineers. Even on its own projections, OpenAI does not expect to be profitable before 2029, although other models are bankrolled by US tech giants with other sources of revenue.
A disappointing launch of the company’s latest model, ChatGPT 5, has also added to worries that LLMs will not get much more capable than they are now, and will continue to fabricate answers.
Disquieting news out of OpenAI over the summer has increased Zenner’s scepticism about the EU’s own initiative. “It did not convince me that the gigafactories are something useful,” he said.
“Many researchers I speak to think LLM performance will plateau,” said Philip Piatkiewicz, secretary general of the European AI-Data-Robotics Association.
Open LLMs
Bienert agreed that “we are achieving some kind of plateau,” but this is not a reason to stop trying to create leading LLMs. Instead, it gave hope that Europe can catch up with the US. “If we are hitting a plateau, it might be even easier,” he said.
However, European LLMs should be openly available to build upon. “We definitely need open source, alternative, sovereign, LLM solutions coming from Europe,” Bienert said.
There are also cultural arguments for Europe having its own models, said Piatkiewicz, for example, ones that are tailored to European languages. “Current US models aren’t good enough for non-English,” he said.
Europe does have its own private AI champions. Perhaps the best known is Mistral, which has developed LLMs with a comparable performance to its US rivals.
Mistral, however, is in large part owned by US investors, Bienert pointed out. The company continues to have to raise money from venture capital firms, and has been linked to a buyout by Apple.
“The question is, how sustainable is this European sovereignty, if you just bet on Mistral,” said Bienert.