What triggered the Monday’s AI pullback

The AI debate took a more consequential turn last weekend, as leading AI researchers and executives focused not just on how quickly AI is improving, but on the possibility that AI could increasingly help develop the next generation of AI itself. The discussion around recursive self-improvement has revived questions about whether the pace of development could become difficult to control, prompting calls from some in the industry for greater safeguards and potentially a slower pace of frontier development.

The market reaction was immediate, with Philadelphia Semiconductor sector index losing -5.6% on Monday, and many other AI-exposed hardware vendors outside the semiconductor sector suffered similar declines. We think that reaction was too broad: while the debate could eventually affect the pace of frontier model development, there is little evidence so far that it is changing the much more immediate investment cycle around AI deployment and datacentre infrastructure.

Why we think investors are conflating two separate questions

The debate over the weekend was about frontier development: should the pace of building more capable AI slow down, given the risk that humans lose meaningful control over increasingly autonomous systems? That is a real and unresolved question.

But the market priced it as if it were a different question: will companies keep buying the computing capacity to run the AI that already exists? Those are not the same question, and it is the AI capex investment outlook, not the frontier-safety debate, that drives stock prices today. The first has genuinely become more uncertain this week. The second has not moved at all.

Recursive self-improvement is not yet the autonomous intelligence explosion the phrase implies. AI today assists human researchers, writing and debugging code, generating training data, designing experiments, evaluating models, rather than independently redesigning its successor end to end, implementing the redesign, and iterating without a human in the loop. The distinction matters because it defines where the actual risk sits. Today, humans still set the research agenda and decide whether results are good enough to deploy, even as AI does more of the underlying work. Recursive self-improvement, in its fuller form, would erode that human bottleneck: each generation of AI designs a better research process, that process produces a more capable model, and the more capable model designs a still-better process again. If the humans supervising that loop eventually stop being able to fully explain why the new system is more capable, oversight, not compute, becomes the binding constraint. That is a legitimate reason for caution at the frontier.

It is also a more complicated argument than "AI is dangerous, slow down." AI development is a geopolitical and commercial competition as much as a technical one. If one lab or one country voluntarily slows while rivals keep pushing, the party exercising restraint simply cedes technological ground. A unilateral pause is not commercially or strategically sustainable in the way a coordinated, multi-party slowdown might be. Coordination of that kind requires agreement not just between companies but ultimately between governments, a materially higher bar. That is why we'd expect any deceleration in frontier development to be gradual and contested, at best.

Slower frontier progress does not mean slower infrastructure spending

The read-through to infrastructure spend is where we push back on Monday's price action. The market's implicit logic looks like this: if frontier development slows, fewer and models get trained; if fewer models are trained, fewer GPUs are needed; if less compute is needed, the datacentre investment cycle slows. We think that chain is too linear, because it treats the AI infrastructure build as a bet on the next frontier model when it is increasingly a bet on deployment.

The first phase of this investment cycle was dominated by training ever-larger frontier models. But the next (current) phase is broader: companies need capacity not just to train models but to run them, inference, networking, storage, power and cooling, on top of raw datacentre capacity. That build has its own physical constraints that are largely decoupled from the frontier debate. A datacentre takes years to plan, permit, construct and connect to the grid; semiconductor supply agreements and power procurement are similarly locked in years ahead. The capex decisions being made today are therefore bets on the pace of broader AI adoption across search, cloud, enterprise software, coding and customer service, not bets on how capable the next model will be.

The corporate evidence supports that framing. Hyperscaler capex guidance has moved up through the year, not down, with several management teams raising outlooks rather than trimming them. Nvidia’s own guidance points the same way, with commentary suggesting continued demand from hyperscalers and AI labs rather than an imminent slowdown. Directionally, nothing in recent guidance from the companies that are spending the money signals an air pocket. On the contrary, the current guidance implies that the combined AI capex of 6 top hyperscalers will reach $1.5tn in 2027, a 70% increase YoY. Moreover, the magnitude of the pre-committed spend of $2.7tn between these 6 companies suggests that their capex plans will extend well beyond next year.