Jensen Huang’s fall further exposes EU leaders on AI

BRUSSELS — It takes an 80-year-old German diplomat to put the Jensen Huang spy scandal in proper historical context.

Asked about allegations that the Nvidia CEO was lured into an intimate relationship with a Chinese spy, Wolfgang Ischinger, Germany’s former ambassador to Washington, pointed to the past. “In World War I, you had female spies who used the honey trap to talk to a general, and maybe sleep with him and extract some secrets,” he told POLITICO.

As Europe’s 27 leaders meet for their first conversation on AI the day after Huang’s resignation, his fall from grace is a reminder that it’s American tech bros — not European military men — who now hold the secrets that can decide the fate of the world.

Friday’s roundtable of European leaders in Dublin must reckon with the reality that the pursuit of AI superiority remains a two-horse race between the U.S. and China — and that Nvidia’s leadership drama doesn’t change the fact that a single American company’s chips will continue to drive the artificial intelligence boom.

With the meeting meant to help forge a new direction for Europe on AI, leaders will mostly be in listening mode, with presentations by Arthur Mensch, CEO of French AI company Mistral, and computer scientist Geoffrey Hinton, considered one of the “godfathers of AI” and an advocate for more stringent safety measures.

“We’re not going to pretend that we are experts,” an EU official said.

The session comes as Europe struggles to find its place in global AI. The rise of advanced U.S.-built proprietary models and open-weight artificial intelligence progress in China has created a new range of national security issues — and while the EU has proven to be a leader in regulating the technology, its own would-be frontier AI labs continue to play catch-up.

Although this week’s events are a reminder to European politicians that they are spectators in the AI race, an emerging split between top industry leaders in Europe is also dampening momentum on a bloc-wide effort to take on the competition.

“Switching dependency from California to France doesn’t solve Europe’s sovereignty problem, right?” Uljan Sharka, CEO of leading Italian company Domyn, said in an interview.

Sovereignty

Sovereignty and security are the main arguments for Europe not to abandon efforts to develop its own frontier AI models. In the past year, EU officials and companies have struggled to access the latest models from OpenAI and Anthropic, which drive much of the current risks around the technology.

“How can Europe become a leading and sovereign player in strategic areas of the AI value chain?” European Council President António Costa asked EU leaders in his invitation letter.

Yet in classic European fashion, officials disagreed ahead of the meeting on how to catch up across many layers of the AI supply chain, from chips and cloud services to cutting-edge models.

“We cannot abandon the frontier ambition. It’s an issue of our sovereignty, it’s an issue of our competitiveness,” Malgorzata Nikowska, head of unit at the European Commission’s Artificial Intelligence Office, said at an event last week.

It was a moment of clarity in a frantic debate about whether Europe should still chase the frontier in AI, which would require ploughing billions of euros into compute power — or focus on deriving value by applying the technology in key industries.

Mistral’s Mensch has been a key voice in that debate, warning that giving up the frontier would be “a fundamental mistake” in an essay titled “Build or Surrender.”

Two European PMs — Finland’s Petteri Orpo and Estonian Kristen Michal — came out swinging ahead of Friday’s meeting, saying “Europe can compete at the frontier” and that the bloc’s ambition “must go beyond adoption of AI,” in a letter exclusively seen by POLITICO.

These statements, and Mensch’s presence at the table Friday, suggest EU leaders will commit to continuing to build competitive frontier models. But how to get there is a far more vexed question — and European industry executives are offering different answers.

Mistral has taken the route of building large open models. Its latest Large 4 model crossed the 1-trillion-parameter mark, allowing it to rival some of the world’s best open models such as from China’s Moonshot AI.

To get there, the French company relied on President Emmanuel Macron as its best-placed salesman, and multi-billion euro funding rounds valuing the company at over €21 billion.

But Mistral’s European rivals don’t necessarily agree with Mistral’s approach of training large models through vast amounts of compute power.

Sharka — whose company was chosen to lead an EU-backed consortium to develop an open-source AI model that can cover the 24 EU languages by next year — argues Europe should not follow the example of OpenAI and Anthropic to use ever-more compute power for ever-larger models.

Instead, Sharka argued, it should aim to optimize the training runs of the models. “To build the European frontier, we need to be original,” he said. “I think we need to shift away from the concept of copy-pasting like unfortunately some players are pushing.”

He argued that the typical AI scaling laws, which claim that a model performs better if it is larger and has more compute power, have “plateaued.”

“There’s only one way, and it’s not the Mistral way to be clear,” he said, adding that there’s a “lot of space for optimization.”

Mistral did not respond to a request for comment.

The Nvidia problem

Sharka was joined in his assessment by Aldo Podestà, CEO of Swiss AI model provider Giotto.ai., who argued there are ways to be more efficient about “pre-training,” typically the most compute-intensive part of building an AI model.

“We have the opportunity to be at the frontier without having that massive pre-training cost,” Podestà said. “My bet would be to have one very strong common open-source pre-training exercise” — after which different companies can tweak the outcome to their needs.

Under any scenario, Europe will still rely on the U.S. for the hardware components that power its AI models — especially Nvidia’s AI chips — since the continent has little design or production capacity of its own.

Podestà said the bloc shouldn’t worry merely about building its own models.

“We need to open up a serious discussion about the hardware dependency we have,” he said. “Today it’s a monopoly for a single company.”