The AI race hits the brakes: what changes for investors?

Key takeaways

• Slower frontier development could shift AI spending from training bigger models towards deploying, securing and using existing ones.

• Hardware faces the biggest question because valuations assume computing demand keeps expanding quickly.

• Tougher safety requirements could unintentionally strengthen today’s largest AI laboratories by raising the cost of competing.

For the past few years, the artificial intelligence race has followed a simple rule: bigger models, more computing power, faster development. Now some of the people leading that race are asking whether everyone should slow down.

Anthropic chief executive Dario Amodei has called for companies to pace improvements in frontier AI capabilities so safety systems have time to catch up. OpenAI chief scientist Jakub Pachocki says no laboratory has solved safety well enough to keep scaling at maximum speed indefinitely.

The obvious question is whether this threatens the AI boom. The more interesting answer is that it may simply change where the money goes.

Slower models, not slower AI

The key distinction is between building the next frontier model and using the models we already have.

Amodei is not proposing that AI development stops. His idea of “pacing” means allowing more time for testing, monitoring and safeguards before capabilities move another large step forward. OpenAI has already temporarily slowed scaling when one upcoming model approached critical cybersecurity capabilities.

That matters because AI spending extends far beyond training.

Companies need computing power to run models, a process known as inference. Businesses need software to integrate them into workflows. They need cybersecurity to control them, data systems to feed them information and applications that turn intelligence into something customers will actually pay for.

In simple terms, the AI value chain may move from:

training → inference → applications → security → deployment

A slower first step does not necessarily shrink everything that follows. In fact, businesses have barely begun extracting value from the capabilities already available. The investment question could therefore shift from “how quickly can we build a smarter model?” towards “how much economic value can we create from the models we already have?” For investors, that may be the healthier question.

The picks-and-shovels trade gets tested

The first phase of the AI boom rewarded the companies supplying scarce computing power. Graphics processing units, or GPUs, advanced memory, networking equipment, data centres and electricity all benefited from one powerful assumption: frontier AI would keep requiring substantially more computing power.

A slower frontier challenges the simplest version of that story. If the next generation of models arrives later, some training clusters could be delayed. That matters most for businesses whose growth expectations depend on capital spending continuing to accelerate almost automatically.

But training demand and AI demand are not the same thing. Existing models still require enormous amounts of computing power every time people use them. AI agents, enterprise adoption and scientific applications could increase inference demand even if new frontier models arrive less frequently.

This is why Monday’s market reaction may be more useful as a warning than as a verdict. Hardware companies face a new question about the pace of demand, while software and cybersecurity could gain more attention if the industry spends longer commercialising, monitoring and protecting existing AI systems.

The AI trade may be becoming less about one bottleneck and more about the entire stack.

Safety can become a moat

There is another, less obvious consequence. Slowing development sounds like a constraint on OpenAI and Anthropic. It could also strengthen their competitive position.

Imagine that developing frontier AI increasingly requires independent safety evaluators, sophisticated monitoring, heavily secured infrastructure and extensive testing before a model can be released.

Those measures cost money. OpenAI, Anthropic, Google and other well-funded laboratories can probably afford them. A smaller challenger may struggle. That creates an uncomfortable tension. Rules designed to make AI safer could also raise barriers to entry and make today’s leaders harder to challenge.

Then there is China. US laboratories can voluntarily pace development, but they cannot guarantee competitors elsewhere will do the same. Amodei acknowledges this directly, arguing that any slowdown must consider the technological gap between the United States and China.

That makes a permanent global pause difficult. Competitive pressure has not disappeared just because the safety debate has become louder.

The risks are on both sides

Investors should avoid extrapolating Monday’s sell-off too far. Frontier laboratories are discussing slowing the pace of capability advances, not abandoning AI. OpenAI still aims to build automated AI researchers, while Anthropic continues to argue that advanced AI could generate enormous economic and scientific benefits.

The opposite risk also matters. If voluntary restraint proves impossible because competitors keep pushing ahead, the current debate may lead mainly to more testing and regulation rather than materially slower computing demand.

The early signs to watch are therefore practical: delays to major training programmes, lower capital-spending plans from hyperscalers, slower orders for advanced chips, or evidence that spending is rotating towards inference, software and security.

Investor playbook

• Separate training from usage. Slower frontier development does not automatically mean lower demand for every part of the AI ecosystem.
• Watch capital spending. Hyperscaler budgets provide a better demand signal than headlines about whether one model launch is delayed.
• Follow where revenue appears. Growing inference and software revenues would suggest AI is shifting from infrastructure build-out towards commercial use.
• Mind concentration. A portfolio heavily exposed to one part of the AI stack becomes more sensitive when the industry’s spending priorities change.

From bigger to better

For several years, the AI race has largely been measured by one question: who can build the most powerful model next? Investors may increasingly need to ask another: who can make the most money from the intelligence we already have?

A slower frontier would challenge the simplest version of the AI trade, where ever-larger models automatically require ever-more chips, power and data centres. But it would not necessarily weaken AI itself. It could move the industry’s centre of gravity from creating intelligence towards deploying, securing and monetising it.

The first AI race was about building the biggest brain. The next may be about turning that brain into a sustainable business.