When AI Leaders Call for a Slowdown: Caution, Fear or Protecting Their Business?

Image generated with Artlist under licence.

French AI start-up Mistral has just completed the largest venture capital funding round ever raised by a European technology company: €3 billion, at a valuation exceeding €21 billion.

Almost simultaneously, the leaders of the biggest players in AI have begun calling for a slowdown in the artificial intelligence race. Are we dealing with a technology that frightens even its creators, or an industry looking to raise barriers, buy time and protect extraordinary valuations?

Artificial intelligence is facing a contradiction that's difficult to ignore.

On the one hand, money continues to pour into the industry at an extraordinary pace. Mistral AI has closed a €3 billion Series D funding round, reaching a post-money valuation of more than €21 billion.

Image sourced from multiples.vc

According to the company itself, this is the largest funding round ever raised by a European technology business. It was led by Samsung Electronics, alongside Scaleup Europe Fund and PSG Equity, with additional investors including ASML, Nvidia, Salesforce Ventures and BlackRock.

On the other hand, the leaders of major AI companies — Anthropic, OpenAI, Google and even Elon Musk — have all suggested that the development of the most advanced models should be slowed down to allow safety measures to catch up.

Money to accelerate. Public statements calling for restraint. It's tempting to conclude that someone isn't telling the whole truth.

The reality is probably more uncomfortable: they're telling part of the truth while simultaneously protecting their own interests.

Mistral Is Not All Hype, but Its Valuation Reflects More Than Its Current Business

The funding round radically changes Mistral's capabilities. It allows the company to purchase computing capacity, attract researchers, expand its infrastructure, grow internationally and sustain its position for longer in a race where training and operating AI models requires ever-increasing amounts of capital.

The company says it already operates in 20 countries and is involved in the transformation of more than 125 major corporations, including Airbus, ASML and HSBC. Its ambition is no longer simply to sell AI models: it wants to control an ecosystem of open models, enterprise products, infrastructure and computing capacity.

Mistral expects to reach €900 million in annualised recurring revenue by the end of 2026. Excluding the capital just raised, its valuation represents approximately 21 times that ARR. That's a very high multiple, although not extraordinary in the private AI sector (a market that is clearly experiencing inflated valuations).

Nevertheless, three factors should be considered:

• First, ARR represents an estimated annual revenue run rate, not necessarily the revenue actually generated during the year.

• Second, Mistral is a private company, and we don't have precise information about its margins, the cost of delivering its contracts or its infrastructure commitments.

• Third, its technical position alone doesn't justify the valuation. According to the Artificial Analysis Intelligence Index, Mistral's best model scores 15 points, compared with 53 and 54 for the leading US models and 45 for the best Chinese model.

Mistral isn't an empty bubble, but it isn't worth €21 billion because it currently has the best AI model either. Its valuation includes something else: the strategic value of having a sovereign European AI provider. In other words, a geopolitical premium.

For governments, banks, manufacturers and regulated businesses, Mistral offers data control, private deployment, customisation, open-weight models and reduced dependence on US or Chinese providers. Samsung and ASML aren't simply investing in an AI company: they're buying strategic protection against an increasingly politicised technology supply chain.

Today, Mistral is part software company, part infrastructure provider and part geopolitical insurance policy.

The bubble risk emerges if that sovereignty premium fails to translate into lasting economic advantages. Having a European passport isn't enough. Mistral will need to demonstrate that its customers renew their contracts, that those contracts generate margins after computing costs and that its technology can remain sufficiently close to the cutting edge to avoid becoming a sovereign but clearly inferior alternative.

Alongside this enormous funding round by European standards, we're seeing a very different message from the major US players: it's time to pull the handbrake.

AI Leaders: Do They Genuinely Believe in the Risks, or Are They Simply Making Strategic Moves?

When AI leaders call for restraint, we tend to frame the issue as a simple choice: either they genuinely believe in the risks or they're using fear to protect their businesses.

The reality is that both things can be true.

An executive can be sincerely concerned about cybersecurity, agent autonomy or losing control of advanced systems while also recognising that complex regulation favours companies that already have billions in funding, legal teams, extensive evaluation capabilities and privileged access to chips.

There are, however, reasons to take these concerns seriously. OpenAI has stated that it temporarily slowed its development after identifying that some of its models could exceed established safety thresholds. Anthropic, meanwhile, has proposed coordinated legal mechanisms to moderate development. Google was already calling for AI regulation back in 2020.

At the same time, there are reasons to be sceptical. Regulating frontier AI models involves fixed costs that large companies can absorb but start-ups cannot. Licences, audits, safety testing, reporting obligations and early access to evaluators can become a regulatory wall. The OECD has warned that chips and cloud infrastructure already exhibit strong economies of scale and high levels of market concentration.

To determine whether these companies are genuinely sincere, we need to look beyond the tone of their statements and examine the actual measures they're proposing:

• If the rules apply only to genuinely dangerous capabilities, with objective thresholds, independent evaluators and proportionate obligations, we're moving closer to effective safety regulation.

• If they require expensive licences for much less capable models, allow major players to define the standards and exempt their own deployments, we're moving closer to regulatory capture.

• If companies disclose incidents, accept external audits and delay launches even while competitors continue moving forward, their credibility increases.

• If they collectively call for a slowdown while accelerating their own development, the brakes are largely rhetorical.

  
  

Are They Trying to Keep New Competitors Out?

It's certainly possible, and under some versions of this supposedly humanitarian approach, almost inevitable.

The race to develop foundation models already requires scarce resources: capital, chips, energy, data and talent.

Adding a poorly designed regulatory layer raises entry costs even further. Established players could even turn their own internal protocols into standards that everyone else must follow.

That doesn't prove safety is merely an excuse. It demonstrates that good intentions can produce anticompetitive advantages.

This is why regulation should focus on high-risk capabilities and uses rather than protecting particular brands or architectures. It should fund shared evaluation systems, include independent scientific and societal representation, and avoid imposing identical compliance costs on Mistral, a small open-source project and an AI company valued at hundreds of billions.

Is There Also an Intention to Slow China Down?

A slowdown limited to Western countries would effectively give China more time. That's why major players talk about international coordination and rules applicable to all models, regardless of where they originate. However, safety concerns can also be used to justify export controls, restrictions on chip access and limitations on open models that disproportionately affect Chinese competitors.

The safety argument can therefore serve two purposes: restricting dangerous capabilities and preserving a national competitive advantage. Separating the two will be difficult. The relevant question isn't whether a particular rule disadvantages China, but whether the same criteria would apply equally to a US company with identical capabilities and risks as to a company from another country.

Universal regulation is highly unlikely, but critical infrastructure is concentrated among a small number of chip manufacturers, cloud providers and major data centres. Agreements between the jurisdictions controlling these bottlenecks could be incomplete and still have a meaningful impact. The challenge isn't purely technical: it's persuading major powers to accept mutual restrictions when they believe leadership in AI will determine their economic and military strength.

Are They Hiding the Limitations of Their Models?

It's possible that calls for caution could provide an opportunity to delay a disappointing product launch or buy time to fix technical problems. However, there is no clear evidence that these calls to slow AI development are intended to conceal shortcomings.

In fact, acknowledging that a model is unsafe doesn't necessarily mean admitting that it's commercially inadequate. Sometimes the opposite is true: the more capable a system becomes in areas such as programming, persuasion or research, the more challenging it can be to deploy safely. Safety can be a genuine limitation while also providing a very convenient business explanation.

OpenAI's Delayed IPO

This is a particularly ambiguous issue. The company confidentially submitted preliminary IPO documentation in June, but had not publicly announced a definitive listing date. A few days ago, Sam Altman stated that OpenAI would not go public in 2026, arguing that it would be an unsuitable time and that the company needed to focus on safety.

For now, it would be more accurate to describe this as the postponement of a potential public listing rather than the cancellation of a planned IPO.

Remaining private offers clear advantages: less quarterly pressure, greater freedom to absorb enormous costs, tighter control over governance and less public scrutiny of its financial accounts. It also avoids immediately exposing an extraordinary valuation and a business model that still consumes vast amounts of capital to the scrutiny of public markets.

But is safety really a lifeline? OpenAI has continued raising private capital on an enormous scale. Nevertheless, safety can provide both a narrative and a financial shield. It offers a difficult-to-challenge justification for buying time, postponing the scrutiny of public markets and waiting for revenues, margins and reliability to come closer to supporting the company's valuation.

Ultimately, it may be a technically prudent decision, morally defensible and, above all, financially convenient.

The AI Bubble: A Bet on Extreme Outcomes

The AI bubble isn't about the technology having no value. It does have value. It's already generating revenue, productivity gains and products used by millions of people. The speculative element lies elsewhere: in assuming that a handful of companies will capture an enormous share of all future value before we know their margins, the sustainability of their competitive advantages or the energy and financial costs required to maintain them.

In Mistral's case, investors aren't simply buying future revenue. They're buying the possibility that Europe will need its own AI champion and that technological sovereignty will become a purchasing criterion as important as the raw quality of the models.

With OpenAI and its competitors, investors are buying the possibility that whichever company reaches certain capabilities first will capture an enormous market. And now those same companies are calling for mechanisms to prevent that race from ending in disaster.

The result isn't a perfectly coordinated conspiracy, nor is it an act of pure altruism. It's something much more human and familiar: companies that recognise genuine risks, compete fiercely and try to influence the rules that will determine who survives.

In short, the decisive question isn't whether we should believe the AI gurus. It's whether we're willing to build institutions capable of independently verifying their claims without handing those same companies the power to write their own rules.