Will the Logic of CapEx and Tech Stocks Change After AI Slows Down?

Original Title: The bottleneck isn't the model. It never was.
Original Author: Saket Mehrotra, Beta to Alpha

Editor’s Note: Recently, Jacob Coxon, who previously worked on pre-training at OpenAI and Anthropic, announced his departure from Anthropic and publicly called for leading AI labs to coordinate in limiting the continued enhancement of model capabilities. Evan Hubinger, the head of alignment science at Anthropic, subsequently responded that he agrees with Coxon’s main assessment of the risks and estimates the probability of AI leading to human extinction in the next decade to be over 10%.

This debate is shifting the discussion of "AI slowdown" from an ethical issue to the capital markets: If frontier labs really slow down model development, will the AI capital expenditures that have been continuously revised upwards over the past few years also reach a turning point? Will tech stocks that have gained valuation premiums based on AI narratives need to be repriced?

Saket Mehrotra, author at Beta to Alpha, argues in his article "The bottleneck isn't the model. It never was." that the concerns of labs regarding AI risks are not the same as a cooling in infrastructure investment. As long as companies remain worried about falling behind in the race, safety anxieties may continue to drive investments in computing power; meanwhile, the industry bottleneck has gradually shifted from GPUs, HBM, and advanced packaging to electricity and the power grid.

This does not mean that the AI slowdown has no impact on the market. More accurately, the impact depends on which layer the slowdown occurs: If it only extends safety testing and slows model releases, CapEx may not necessarily decline; if the government directly restricts training clusters, chip supplies, or data center electricity usage, the entire investment framework may undergo fundamental changes.

Below is the original text compilation:

Jacob Coxon posted his resignation statement on X, calling for frontier AI labs to coordinate in slowing down the capability race.

Jacob Coxon’s departure has once again brought the risks of frontier AI to the forefront.

Coxon believes that OpenAI and Anthropic are competing to develop self-improving superintelligence without taking sufficient responsible safety measures. This is not an isolated voice from external critics. Evan Hubinger, head of alignment science at Anthropic, publicly stated that he agrees with Coxon’s main assessment and estimates the probability of AI leading to human extinction in the next decade to be over 10%. This statement has also been reported by media outlets such as WIRED.

However, for investors, another question may be more direct: As frontier labs begin to discuss slowing down, will AI capital expenditures at tech companies also decline?

Mehrotra believes that, at least at this stage, one cannot draw a direct equivalence between the two. The AI safety controversy occurs at the level of models and lab governance, while the constraints of the current CapEx cycle have gradually sunk to physical infrastructure such as chips, packaging, electricity, and data centers.

The More Concerned About Losing the Race, the Harder It Is for Labs to Slow Down

According to Coxon’s description, frontier labs are not unaware of the risks. The real problem is that no lab is willing to be the first to slow down.

The logic of the game is: If one lab chooses to slow down, other competitors with weaker safety awareness may develop more powerful models first. Rather than allowing less cautious competitors to break through first, it is better to continue investing resources to ensure that one remains a leader and strives to control technology in a safer manner.

In the author’s view, this mindset is difficult to translate into a contraction of capital expenditures; rather, it may reinforce the arms race. The more labs believe they are participating in a decisive technological competition, the harder it is to cut GPU purchases, training clusters, and data center investments.

Therefore, safety anxieties and CapEx growth can coexist. It may even form a self-reinforcing cycle: labs worry about the rapid enhancement of AI capabilities, yet invest more resources due to fears of falling behind competitors, further accelerating the capability race.

This is the author’s explanation of the industry’s game structure, not a result that has been confirmed by company capital expenditure plans. To determine whether this logic holds, it is still necessary to observe whether major cloud providers and AI labs begin to adjust their actual procurement and construction plans.

Slowing Down Models Does Not Mean the End of the Infrastructure Cycle

Mehrotra’s second judgment is that the core bottleneck of the AI industry has migrated multiple times, and each migration has increased capital intensity.

Initially, the limitation on model training was GPU supply and chip quotas; subsequently, the bottleneck shifted to HBM high-bandwidth memory and advanced packaging; now, electricity supply, grid access, and data center construction are becoming constraints that are more difficult to resolve quickly.

Chip and packaging capabilities can be alleviated by expanding production lines, but electricity issues usually require the construction or restart of power generation facilities, increasing gas turbine and transformer capacity, and upgrading transmission and grid systems. These types of projects require larger investment scales, longer construction cycles, and cannot be resolved through a single software upgrade.

Therefore, even if the frequency of releasing frontier models decreases, already signed chip orders, electricity procurement agreements, and data center construction projects may not immediately stop. There is a time lag between model development and infrastructure construction, and CapEx typically does not shift in sync with public opinion or the pace of research and development.

Meanwhile, a slowdown does not necessarily mean that the demand for computing power naturally declines. Longer safety tests, more complex reasoning processes, and the deployment of existing models in enterprise and consumer scenarios may still consume a significant amount of computing resources.

The author believes that investors should not only focus on the model layer that is most hotly debated on social media but should also look for new bottlenecks forming in the next 18 months. According to this framework, the currently more noteworthy scarce resources have gradually shifted from "model intelligence" to electricity and its supporting infrastructure.

AI Tech Stocks May Shift from Broad Gains to Differentiation

If slowing down models does not mean stopping CapEx, its impact on tech stocks will not be a simple overall negative but is more likely to reflect a repricing of the industry chain.

First, companies that rely on continuous leaps in frontier models and rapid monetization of products may be more easily affected. Once model iterations slow down, the market’s assumptions about revenue growth, commercialization pace, and valuation multiples may need to be adjusted.

Second, companies that control GPU, networking equipment, power distribution, cooling, and data center resources may not see their order logic deteriorate simultaneously. As long as cloud providers continue to expand infrastructure or the computing power bottleneck continues to shift towards electricity, related investments may maintain strong inertia.

Third, the market needs to distinguish between training and inference. Even if labs coordinate to limit ultra-large-scale frontier training, the inference demand of existing models, enterprise deployments, and the popularization of AI applications may still drive computing power consumption. At that time, CapEx may undergo structural changes rather than a cliff-like contraction.

This means that the first thing that may change due to the AI slowdown is not the growth direction of the entire tech industry, but rather the situation where different assets share the same set of AI valuation logic. In the past, model companies, cloud providers, semiconductor companies, and power infrastructure suppliers could all gain premiums from the AI narrative; if the competition for model capabilities slows down, the market may begin to more strictly differentiate between technological leadership, commercialization capabilities, and order fulfillment.

The Variable That Truly Changes CapEx Logic Is Regulation

Mehrotra believes that what could truly interrupt this round of AI capital expenditure cycle is not another supply chain shortage, but rather the government imposing hard constraints from outside the system.

Shortages of GPUs, HBM, electricity, and transformers are essentially bottlenecks that can still be alleviated by increasing investment. They may delay the launch of data centers but will also guide capital towards new scarce segments.

Regulation is different. If the government directly restricts the procurement of high-end chips, the scale of training clusters, the computing power required for model training, or the electricity usage of data centers, companies will find it difficult to break through constraints merely by increasing investment. In this scenario, cloud providers’ CapEx expectations, infrastructure orders, and valuations of AI-related tech stocks may simultaneously face downward revisions.

However, the original text views this situation as a low-probability tail risk rather than a baseline scenario. This judgment also needs continuous verification and cannot be understood as regulatory risks being eliminated. As the concerns within AI labs become public, external policy pressures may still rise.

Moving forward, the market truly needs to observe not the vague slogan of "AI slowdown," but three more specific signals: whether frontier labs will write slowing down into formal R&D plans, whether major cloud providers will cut data center orders, and whether regulation will begin to directly touch on chips, computing power, and energy supply.

Coxon’s departure proves that the safety debate in frontier AI has moved from external criticism into the labs themselves. However, there remains a gap between researchers expressing concerns and tech companies actually cutting capital expenditures, influenced by competitive pressures, construction inertia, and infrastructure shortages.

AI slowdown may not immediately end the CapEx cycle, but it may conclude the phase where all AI assets share the same rising logic.

[Original link]