Some of the most influential figures in artificial intelligence are calling for the pace of frontier AI development to slow, warning that capabilities are advancing faster than the industry’s ability to control them. The safety concerns deserve to be taken seriously. But with trillions being committed to AI infrastructure and questions growing around investment returns, slowing the frontier could also give leading providers something commercially valuable: time.

Something unusual happened in artificial intelligence this month. After years in which competition between the leading AI companies was defined by speed, model capability and the race to reach the next frontier first, several of the industry’s most prominent figures began arguing that development should slow down.

Anthropic CEO Dario Amodei made the clearest case in an essay titled We Must Pace the Frontier, arguing that rapidly improving capabilities, particularly AI systems increasingly contributing to AI development itself, require stronger testing, independent evaluation and greater coordination between leading developers. OpenAI’s Sam Altman, xAI’s Elon Musk and Google DeepMind’s Demis Hassabis subsequently expressed support for some form of slower or more controlled development.

The agreement is far from universal. Meta CEO Mark Zuckerberg has rejected the case for a coordinated slowdown, arguing that competition, liability and individual responsibility already give laboratories incentives to develop models safely. European AI companies including Mistral have been more critical, warning that restrictions imposed around the technological frontier could entrench the position of the American companies that reached it first.

There are therefore good reasons to examine what has changed. The obvious explanation is safety, and recent developments make that difficult to dismiss. But another change has occurred at almost exactly the same time. The financial cost of the AI race has become enormous, while the economic returns required to justify it are becoming increasingly demanding.

That does not mean safety is being used as an excuse for disappointing ROI. There is no evidence to support such a conclusion. It does mean that safety and economics may now be creating incentives that point in the same direction.

The safety argument has become harder to ignore

Concerns about advanced AI are hardly new. Researchers and technology executives have spent years debating alignment, autonomous systems, cybersecurity, biological risks and the possibility that increasingly capable models could behave in ways their developers cannot reliably control.

What has changed is the capability of the systems themselves. AI agents can now perform longer sequences of actions, interact with software environments and increasingly contribute to technical research and software development. Recent demonstrations of AI-assisted cyber operations have made the security implications more immediate, while frontier laboratories are devoting growing resources to determining which capabilities should trigger additional safeguards.

Amodei’s argument is essentially that development and assurance are moving at different speeds. If capability continues accelerating while evaluation, security and governance lag behind, the industry risks deploying systems before it has sufficient mechanisms to understand or contain their behaviour. His proposed response includes stronger safety testing, permanent independent evaluators and cooperation between leading developers.

Those concerns should be considered on their merits. The fact that slowing development might also produce commercial advantages does not make the underlying safety argument invalid.

It does, however, make the economics worth examining.

The AI race has become extraordinarily expensive

The Bank for International Settlements describes the current AI build-out as one of the largest technology-driven investment booms in US history. It estimates that the five largest technology companies will spend more than $1 trillion on AI-related capital expenditure across 2025 and 2026, while industry expectations put global AI investment somewhere between $3 trillion and $4 trillion by 2030.

S&P Global provides an even broader indication of the scale. Its analysis estimates that Amazon, Alphabet, Microsoft, Meta and Oracle have collectively spent around $1.1 trillion in capital expenditure during the past five years, with market expectations pointing towards another $5.3 trillion by 2030. S&P’s concern is not simply the amount being spent, but what happens if the investment cycle begins shifting from cash-funded expansion towards greater dependence on debt before the returns have been established.

The competitive structure of the AI market compounds the problem. A company cannot easily decide that its existing model is sufficiently capable if a rival is preparing something substantially better. Each new generation can require another expensive training programme, more accelerators, additional data-centre capacity and greater energy supply. Falling behind at the frontier could affect enterprise customers, developer ecosystems, talent and ultimately market position.

This creates an investment race in which the rational decision for each individual company can produce an irrational outcome for the industry as a whole.

BIS researchers have attempted to quantify that effect. Their model suggests competitive pressure in a winner-take-most market could push AI investment to around 50 per cent above the socially efficient level under a conservative scenario. The concern is not that artificial intelligence lacks economic value, but that companies competing for a limited number of dominant positions may commit more capital than the eventual returns can support.

Slowing down changes the economics

A slower frontier would not stop AI investment. Data centres would still be built, enterprises would continue deploying models and providers would continue competing for customers. What it could change is the frequency with which the industry has to finance another major leap in model capability.

That creates commercial breathing room.

Providers could concentrate more resources on improving inference efficiency, integrating existing models into enterprise workflows, developing agents and vertical applications, reducing operating costs and converting widespread experimentation into recurring revenue. Customers would have more time to redesign processes around capabilities that already exist rather than continually evaluating the next generation. Infrastructure providers would have longer to generate returns from facilities and hardware already being deployed.

In other words, slowing capability development could extend the commercial life of today’s frontier.

There is no evidence that this is the primary reason Amodei, Altman or others support pacing development. But it would be an economically useful consequence at a time when the amount of capital committed to AI is increasing faster than evidence of equivalent financial returns.

This is particularly relevant because AI infrastructure depreciates economically as well as physically. A data centre remains a data centre, but the accelerators, networking and architectures inside it exist in a market where performance expectations change quickly. A development cycle that continually demands more capable infrastructure can shorten the period available to recover investment from the previous generation.

A slower frontier potentially gives the industry more time to turn capability into revenue before financing the next capability jump.

The incumbent problem

The argument becomes more complicated when viewed from outside the leading US laboratories.

Mistral and other European technology companies have challenged calls for coordinated pacing precisely because the leading developers already possess the models, computing infrastructure, capital and customer relationships that challengers are trying to build. A slowdown that freezes or widens the gap between frontier companies and those attempting to catch them could therefore have competitive consequences irrespective of its safety objectives.

This is why proposals for coordination between leading laboratories raise questions extending beyond AI safety. Amodei has proposed a narrow antitrust exemption that would allow major developers to cooperate on safety standards. Supporters see coordination as necessary because unilateral restraint becomes difficult if competitors continue racing ahead. Critics see a danger that the companies already occupying the frontier could acquire influence over the rules governing who is allowed to approach it.

Europe has an obvious interest in that distinction. The region is already attempting to reduce dependence on American AI infrastructure and develop its own models, computing capacity and industrial ecosystem. A global regime designed primarily around the capabilities and risk thresholds of today’s leading laboratories could inadvertently make technological sovereignty more difficult for those still trying to close the gap.

Safety standards may be necessary. The question is who defines them, how independently they are assessed and whether they protect society without protecting incumbents from competition.

Safety and economics can point in the same direction

The AI slowdown debate does not require a choice between believing the safety argument and believing the economic one. Both can operate simultaneously.

AI capabilities may genuinely be advancing faster than the mechanisms available to control them. Independent evaluation, stronger security and better governance may therefore be necessary. At the same time, the companies developing those systems are participating in an extraordinarily expensive investment race whose eventual returns remain uncertain.

A slower frontier could help with both problems. It could create more time to evaluate and secure advanced systems while also giving providers more time to monetise existing capabilities and infrastructure.

That alignment of incentives deserves scrutiny, but not conspiracy.

The more useful question is not whether AI executives are secretly worried about ROI rather than safety. It is whether the industry has reached a point where continuing to accelerate development creates risks that are simultaneously technical, financial and competitive.

The response also matters beyond Silicon Valley. If slowing AI becomes part of the emerging governance model, decisions about pace will influence investors financing the infrastructure, enterprises deploying the technology, countries trying to build sovereign capability and challengers attempting to compete with today’s frontier laboratories.

For several years, the assumption underpinning the AI race was straightforward: faster progress was better progress. The current debate suggests that assumption is beginning to fracture.

Whether that is primarily because the technology has become more dangerous, the investment race has become less sustainable, or because both pressures have arrived together remains an open question. What is increasingly clear is that the economics of AI development can no longer be separated from the debate about how quickly the technology should advance.

Further Reading: Is AI Worth the Investment?