Editor's note: Recently, Jacob Coxon, who previously worked on pre-training at OpenAI and Anthropic, announced his departure from Anthropic and publicly called on leading AI labs to coordinate restrictions on further improving model capabilities. Evan Hubinger, Anthropic's alignment science lead, subsequently responded that he agrees with Coxon's main assessment of the risks and estimates the probability of AI causing human extinction over the next decade at more than 10%.
This debate is pushing "AI slowdown" from an ethical issue into the capital markets: if frontier labs really slow down model development, will the AI capital expenditures that have been repeatedly revised upward over the past few years also face a turning point? And do tech stocks that have gained valuation premiums from the AI narrative need to be repriced?
In the article "The bottleneck isn't the model. It never was.," Beta to Alpha author Saket Mehrotra argues that labs' concerns about AI risk are not the same thing as a cooling of infrastructure investment. As long as companies still fear falling behind in the race, safety anxiety may continue to drive compute investment; at the same time, the industry bottleneck has gradually shifted from GPUs, HBM, and advanced packaging to power and the electrical grid.
This does not mean that an AI slowdown has no impact on the market at all. More precisely, the impact depends on which layer the slowdown occurs at: if it is only extended safety testing and a slower pace of model releases, CapEx may not necessarily decline as a result; only if governments directly restrict training clusters, chip supply, or data center electricity use could the entire investment framework undergo a fundamental change.
The following is a translation of the original article:
Jacob Coxon's departure has once again brought the risks of frontier AI to the forefront.
Coxon believes that OpenAI and Anthropic are racing to develop superintelligence capable of self-improvement without taking sufficiently responsible safety measures. This is not an isolated voice from outside critics. Evan Hubinger, Anthropic's alignment science lead, subsequently said publicly that he agrees with Coxon's main assessment and estimates the probability of AI causing human extinction over the next decade at more than 10%. Related remarks were also reported by media outlets such as WIRED.
But for investors, another question may be more direct: as frontier labs begin to discuss slowing down, will tech companies' AI capital expenditures also decline accordingly?
Mehrotra argues that, at least at this stage, the two cannot be directly equated. AI safety controversies occur at the model and lab governance level, while the constraints of this CapEx cycle have gradually trickled down to physical infrastructure such as chips, packaging, power, and data centers.
The more labs fear losing the race, the harder it is for them to slow down voluntarily
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 game theory logic is this: if one lab chooses to slow down, other competitors with weaker safety awareness may be the first to develop more powerful models. Rather than let less cautious rivals break through first, it is better to continue investing resources to ensure one becomes the leader and strives to control the technology in a safer way.
In the author's view, this line of thinking is unlikely to translate into a contraction in capital expenditures; instead, it may reinforce the arms race. The more labs believe they are engaged in a decisive technological competition, the harder it is to cut GPU procurement, training cluster, and data center investments.
Therefore, safety anxiety and CapEx growth can coexist. It may even form a self-reinforcing cycle: labs worry about the rapid advancement of AI capabilities, yet because they fear falling behind rivals, they invest more resources, further accelerating the capability race.
This is the author's explanation of the industry's game structure, not a result already confirmed by companies' capital expenditure plans. To judge 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.
Model slowdown does not mean the infrastructure cycle is over
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, what limited model training was GPU supply and chip quotas; subsequently, the bottleneck shifted to HBM high-bandwidth memory and advanced packaging; now, power supply, grid access, and data center construction are becoming constraints that are harder to resolve quickly.
Chip and packaging capacity can be alleviated by expanding production lines, but power issues typically require building or restarting power generation facilities, increasing gas turbine and transformer capacity, and upgrading transmission and grid connection systems. Such projects involve larger investment scales and longer construction cycles, and cannot be solved through a single software upgrade.
Therefore, even if the release frequency of frontier models declines, chip orders already placed, power purchase agreements, and data center construction projects may not necessarily halt immediately. There is a time lag between model R&D and infrastructure buildout, and CapEx typically does not pivot in sync with public opinion or the pace of research and development.
At the same time, a slowdown does not necessarily mean that compute demand will naturally decline. Longer safety testing, more complex inference processes, and the deployment of existing models into enterprise and consumer scenarios may still consume substantial computing resources.
The author thus argues that investors should not focus solely on the model layer where social media debates are most intense, but should instead look for new bottlenecks forming over the next 18 months. Under this framework, the scarcer resource currently worth watching has gradually shifted from “model intelligence” to power and its supporting infrastructure.
AI tech stocks may shift from broad gains to divergence
If a model slowdown does not equal a halt in CapEx, its impact on tech stocks will not simply be broadly bearish, but more likely manifest as a repricing across the industry chain.
First, companies that rely on continuous leaps in frontier models and rapid product monetization may be more vulnerable. Once model iteration slows, market assumptions about revenue growth, commercialization pace, and valuation multiples may need adjustment.
Second, companies that control GPUs, networking equipment, power supply and distribution, cooling, and data center resources may not see their order logic deteriorate in tandem. As long as cloud providers continue expanding infrastructure, or the compute bottleneck continues migrating toward the power segment, related investment may maintain strong inertia.
Third, the market needs to distinguish between training and inference. Even if labs coordinate to restrict ultra-large-scale frontier training, inference demand for existing models, enterprise deployment, and the proliferation of AI applications may still drive compute consumption. At that point, CapEx may undergo structural changes rather than a cliff-like contraction.
This means that what an AI slowdown may first change is not the growth direction of the entire tech industry, but rather the situation in which different assets share the same AI valuation logic. In the past, model companies, cloud providers, semiconductor firms, and power infrastructure suppliers could all command a premium from the AI narrative; if the model capability race slows, the market may begin to more rigorously differentiate between technological leadership, commercialization capability, and order fulfillment.
The variable that truly changes the 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 governments imposing hard limits from outside the system.
Insufficient GPU, HBM, power, and transformer capacity are essentially bottlenecks that can be alleviated by increasing investment. They may delay data center launches, but they will also direct capital toward new scarce links.
Regulation is different. If the government directly restricts high-end chip procurement, training cluster scale, computing power required for model training, or data center electricity use, companies will find it difficult to break through the constraints simply by increasing investment. In this scenario, cloud providers' CapEx expectations, infrastructure orders, and AI-related tech stock valuations could all face downward revisions at the same time.
However, the original article treats this situation as a low-probability tail risk, not a baseline scenario. This judgment also needs continuous verification and should not be understood as meaning that regulatory risk has already been ruled out. As concerns within AI labs become public, external policy pressure may still rise.
Next, what the market really needs to watch is not the broad slogan of "AI slowdown," but three more specific signals: whether frontier labs write slowdown into formal R&D plans, whether major cloud providers cut data center orders, and whether regulation begins to directly touch chips, computing power, and energy supply.
Coxon's departure proves that safety disputes over frontier AI have moved from external criticism into the internal laboratory. But from researchers expressing concerns to tech companies actually cutting capital expenditure, there are still multiple links in between, including competitive pressure, construction inertia, and infrastructure shortages.
An AI slowdown may not immediately end the CapEx cycle, but it may end the phase in which all AI assets share the same upward logic.
Welcome to join the official BlockBeats community:
Telegram Subscription Group: https://t.me/theblockbeats
Telegram Discussion Group: https://t.me/BlockBeats_App
Official Twitter Account: https://twitter.com/BlockBeatsAsia
