Are we slowing down AI progress?
What happened?
Dario Amodei, the CEO of Anthropic, called for reassessing the pace of AI progress. Sam Altman, the CEO of OpenAI, then backed the idea, as did others1. The media attention grew after Amodei published an essay on Saturday calling for a slower pace of development and for third-party review of Anthropic’s AI systems.2
AI concerns have reached a fever pitch in recent months after several high-profile AI safety incidents this year. Most recently, an in-depth review of an incident at OpenAI raised concerns about the level of autonomy AI agents could develop and how they might seek to evade or confuse human review.
Framing the issue
We see two market-relevant questions: first, how fast will AI capabilities develop, and how will that feed through to economic growth? And second, how does computing demand evolve from here?
Future AI demand continues to be a critical question but unfortunately one with limited visibility. We have a reasonable understanding of the supply side and its bottlenecks, such as the level of investment spending into data centres and their inputs, including production of chips, data centres, power demand, and similar factors. However, we have comparatively little insight into the demand side: How much are AI labs spending on compute? What is the average compute per user? What is the expected future compute consumption? This makes the investment boom challenging to assess – is it too much, or not enough?
Improving models and growing adoption are the two primary drivers of compute demand related to the AI boom, in our view. Models have continued to improve in recent years in large part through greater compute usage, through pre-release effort (pre-training work, reinforcement learning) and growing focus on inference and multi-step models (reasoning models, agentic AI). The other primary force is the growing adoption of AI tools by individuals and businesses. Both factors have contributed to demand in recent years and, in our view, reinforce each other.
Our view
Today, AI leaders Sam Altman (OpenAI) and Dario Amodei (Anthropic) are considering a deliberate slowdown in AI progress. Elon Musk (SpaceX) also voiced agreement. While calls to slow the pace of development have been heard before, today’s is different as there’s broader agreement among AI leaders. If AI labs follow through, this would present a threat to one of the two pillars of growing compute use (and therefore the infrastructure boom itself), in our view. It may also call into question the rate of spend on AI versus the expected macroeconomic benefit via the productivity channel.
However, we expect this to be a short-lived story for the overall investment boom. We view these concerns as another air pocket in the AI story, comparable to past upsets such as the DeepSeek event last year, this year’s Kimi K3 release and the ongoing rise of China, as well as past focus on monetization concerns. Each of these raised concerns around long-term compute demand trajectories and was accompanied by a bout of market volatility. This may be a similar story.
Reassuringly, even current capabilities are enough to generate a positive macro story, in our view. Even if we assume a slowdown in frontier model development, the currently available model generation remains a powerful toolset. We also still expect model distillation to continue, meaning that current leading-edge capabilities will be available more cheaply in the months to come. As costs fall, adoption should continue to grow, and compute demand should continue to increase, which should in turn provide a positive productivity catalyst.
We also remain sceptical of how serious AI labs are about slowing the pace of development. Anthropic’s Dario Amodei has proposed a slowdown of progress (particularly of recursive self-improvement3) and regular third-party review of AI development. This may introduce more safety guardrails but may not necessarily change the course of AI development. Meanwhile, recent commentary out of China has labelled AI safety concerns as “fearmongering” and suggests that model development will not be subject to the same constraints that OpenAI and Anthropic are discussing. President Trump has also dismissed these fears and has emphasized the US desire to lead in the capability race. These forces may mean any pause in progress is temporary, perhaps just a multi-week story.
AI labs have not yet shown signs of changing course. While Anthropic calls for slowing the pace of development, it has not changed its investment plans as far as we can tell. Anthropic is still planning to spend nearly $520 bn in various infrastructure commitments.4 The visible bellwethers for compute demand – particularly earnings reports of hyperscalers and neoclouds – suggest a healthy outlook. For example, Oracle’s results reported last week showed cloud revenues surprised strongly.5 Similarly, proxies for demand, such as prevailing rental prices for H100s, also point to healthy demand.6 For this to disrupt the AI investment story, we think we’d need to see the growth rate of compute demand waver. We also note that, even if AI progress slows, Anthropic, OpenAI, and others can still benefit from growing adoption and therefore revenue for their product offerings.
Markets may nevertheless trade this news by selling major AI infrastructure beneficiaries. Hyperscalers are likely to be in the line of fire, as well as stocks that are sensitive to perceived swings in computing demand such as neoclouds, networking, cooling, and power generation equipment makers. Meanwhile, we suspect names sensitive to disruption, such as software, may benefit from this news as investors reassess the direction of AI capabilities. In either case, we think there’s a growing role for cybersecurity exposure to help manage AI safety risks.
What are the risks?
If computing demand were to meaningfully slow because of a durable slow-down of AI progress, this would present downside risks for the balance of computing supply versus demand. Currently, AI labs, hyperscalers, and others are investing heavily in computing supply via a boom in data centre construction. A supply overshoot would likely curtail the investment cycle that has defined the AI investment theme in recent years. Markets also have large exposures to AI-related names because of concentration, which may cause broader macro spillovers via wealth effects in the event of a market downturn.
Note: Based on media reports and public statements by management. The Silicon Data Compute Rental Indices (used in the right-side chart) track the cost to rent computing capacity available on the marketplace from hyperscalers, neoclouds, regional data centres, and other graphics processing unit (GPU) platforms. The benchmark focuses on market rental prices for single A100 and H100 chips.
Sources: Various media reports, Bloomberg L.P. and Silicon Data, as of 14 September 2026.
Investment risks
The value of investments and any income will fluctuate (this may partly be the result of exchange rate fluctuations) and investors may not get back the full amount invested. Past performance is not a guide to future returns.
