The AI Trade Just Hit an Uncomfortable Question: What If Progress Slows?

The artificial intelligence trade has spent the past few years running on a simple premise: AI capabilities will continue to improve rapidly, demand for computing power will keep rising, and the companies supplying the infrastructure behind that growth will continue to benefit.

That assumption was challenged over the weekend.

Anthropic CEO Dario Amodei called for the AI industry to “pace the frontier”, arguing that the rate at which AI model capabilities are improving needs to slow so that safety work can keep up. OpenAI CEO Sam Altman and xAI CEO Elon Musk backed the call, while Google DeepMind CEO Demis Hassabis said the proposal was

directionally right, although the details still needed to be worked through.

The reaction from markets was swift. AI linked stocks fell sharply on Monday, with semiconductor stocks taking much of the damage. The Philadelphia Semiconductor Index fell 5.9%, while Nvidia, AMD, Broadcom and Micron all posted significant declines. Semiconductor equipment companies were also hit.

The question now facing the market is not whether AI matters. It is whether the rate of improvement investors have been extrapolating is a fixed feature of the technology, or a choice that the industry has just started debating out loud.

What AI's Leaders Are Actually Saying

It is important to understand what Dario Amodei is, and is not, proposing. In his September 2026 essay, We Must Pace the Frontier, the Anthropic CEO argues that the industry should slow the rate at which it improves the capabilities of frontier AI models, giving safety and alignment work more time to keep pace. He is explicit that pacing does not mean halting model training or technical progress. Instead, he wants companies to take more time to test and safeguard increasingly capable systems.

Part of his concern is that AI is increasingly helping to develop better AI itself, a process he calls recursive self-improvement. Amodei argues that if this accelerates too quickly, it could eventually move faster than the industry's ability to understand and control these systems. His proposal includes independent evaluators working inside frontier AI companies, greater coordination among companies in democratic countries and, eventually, broader international cooperation.

What makes the argument notable is who has supported it. OpenAI CEO Sam Altman and xAI CEO Elon Musk said they agreed with Amodei, while Google DeepMind CEO Demis Hassabis said the proposal pointed in the right direction. These are executives and founders at companies competing to build increasingly capable AI systems. Their support does not mean there is agreement on how much development should slow, but it does show that the debate is now happening inside the industry itself.

Amodei is also not arguing that the US should give up its commercial or technological lead. He explicitly says pacing should give companies more time to improve safety without sacrificing the United States' position in AI. That makes this less a case for stepping away from the technology and more a debate over how quickly the frontier should move.

Can the US Really Afford to Slow Down?

This is where the issue becomes more complicated. The US currently leads the frontier AI race, while China is close enough behind for any deliberate slowdown to have strategic consequences. Helen Toner, executive director of Georgetown's Center for Security and Emerging Technology and a former OpenAI board member, said the best estimates put China's very best models around six to nine months behind the very best US models, although the size of that gap varies by model and capability and is difficult to measure precisely.

Beijing has made clear that it does not support the idea of the US slowing its own development while tightening restrictions on Chinese AI. Foreign Ministry spokesman Guo Jiakun criticised what he described as fearmongering and said confrontation and vicious competition would disrupt global AI governance. Chinese state media has also portrayed Amodei's proposal as an attempt to contain China's technological development.

China is not ignoring AI safety either. Senior officials have warned about the risks that increasingly capable AI could pose to political security, critical infrastructure and information systems, while regulators continue to develop rules around AI safety and the use of autonomous systems. The difference is that Beijing appears unwilling to accept a slowdown that could leave China further behind in a strategically important technology.

That creates a difficult equation for Washington. If the US slows while China continues to push forward, Chinese developers have more time to close the gap. US agencies have also accused Chinese AI companies of using techniques such as model distillation to extract capabilities from leading US models, allegations that Beijing rejects. President Trump has meanwhile argued that the US needs to maintain its lead in AI. Against that backdrop, the Trump administration's planned discussions with President Xi around 24 September could take on added significance. Any agreement on AI would have to balance safety concerns against the risk of giving up technological ground.

Why Did AI Stocks Sell Off?

The market reaction makes more sense when viewed through the AI infrastructure cycle. For several years, the story has been relatively straightforward: better AI models drive more adoption, which drives greater demand for compute, data centres and semiconductors, ultimately supporting earnings across the AI supply chain.

The concern now is whether a slower pace of frontier model development could weaken some of those links, particularly the need to keep expanding training capacity at an accelerating rate. The selling was concentrated in companies closest to that infrastructure buildout, particularly semiconductors and equipment.

Still, it would be too simplistic to attribute the entire market move to Amodei's essay. Wall Street was already under pressure from several other factors. The S&P 500 fell 0.48%, the Nasdaq 0.56% and the Dow 0.29%, while the US 10 year Treasury yield briefly moved above 5%. Oil prices were also rising amid renewed disruption to Middle East energy supplies, adding to inflation and interest rate concerns.

The AI story and the broader market story were therefore hitting at the same time. Higher oil prices and bond yields put pressure on growth valuations, while the debate over pacing raised a separate question about how durable the enormous AI infrastructure cycle will be. Monday's selling looks less like a verdict that AI demand is disappearing and more like a market questioning how quickly the buildout can continue and what returns will ultimately come from it.

What Does This Mean for Nvidia and the Chipmakers?

Nvidia is the obvious focal point because it sits at the centre of the AI compute cycle. But the relevant question is not whether the world will continue to need substantial AI compute. It is whether that demand can continue growing at the extraordinary rate currently embedded in expectations.

There are two separate sources of demand to consider. The first is training, where frontier labs use vast amounts of computing power to develop increasingly capable models. The second is inference, where those models are actually run across search, software, coding, customer service, cybersecurity, robotics and other applications.

A slower frontier could therefore affect parts of the semiconductor cycle without necessarily undermining AI demand as a whole. If training growth moderates but inference continues to expand rapidly, overall compute demand could remain strong even as the composition of demand changes.

That distinction is becoming increasingly important. The question is no longer simply whether AI needs more chips, but whether it needs them at the same rate the market has come to expect.

Could Slower AI Progress Change the Capex Cycle?

The scale of the investment cycle is enormous. Morgan Stanley forecast earlier this year that AI spending would surpass USD 1.3 trillion by 2027, according to Reuters. That helps explain why a debate over slowing frontier development can unsettle the market, but it also shows why a sudden collapse in spending is difficult to assume.

There is also a physical constraint to the buildout. Data centres require power, land and grid connections, with capacity often planned years in advance. Goodman Group chief executive Greg Goodman has described a market still constrained by power and land, with undersupply expected into 2027 and 2028. That makes the link between a slower model cycle and near term infrastructure demand less direct.

AI adoption could also continue even if frontier models improve more slowly. Inference, enterprise deployment and AI agents may create additional demand for computing capacity, while US-China competition could encourage American technology companies to keep investing. A slower pace of model development does not automatically mean a smaller infrastructure market.

For Australia, the exposure sits further down the chain through data centres, digital infrastructure, power and property rather than frontier model development. Any impact on local companies is therefore more likely to come through hyperscaler spending decisions and the pace at which new capacity is commissioned, making upcoming infrastructure commitments more important to watch than the timing of individual model releases.

The AI Boom Isn’t Over. Here’s What Matters Next

Slower frontier progress is not the same as slower AI adoption. Existing models are already being used across software, cybersecurity, healthcare, robotics and enterprise applications, while inference demand could continue to grow even if the pace of new model development moderates.

The next phase of the AI trade may therefore shift from capability to monetisation. Monday's market reaction offered an early example of that rotation. While semiconductor stocks were heavily sold, several software names including ServiceNow, Adobe and Workday rallied, alongside strength in parts of cybersecurity and cloud software. The market was already beginning to distinguish between companies selling the infrastructure required to build more powerful models and those using AI to improve existing products and services.

The signals to watch are becoming clearer. Hyperscaler capex guidance and semiconductor orders will show whether infrastructure spending is actually changing. The pace of frontier model releases will show whether calls for pacing are translating into behaviour, while inference demand and enterprise adoption will reveal whether usage remains strong. The US-China race will matter too, particularly if geopolitical pressure makes it harder for US companies to voluntarily slow development.



For now, there is no evidence that the AI boom is over. What has changed is the assumption that capability, compute demand and capital spending can all accelerate indefinitely. The next phase of the AI trade may depend less on how quickly the technology improves and more on how effectively companies turn what already exists into sustainable economic value.

The AI trade has not necessarily broken, but its assumptions are being tested.